arXiv CSopen access

To Build or Not to Build? Factors that Lead to Non-Development or Abandonment of AI Systems

To Build or Not to Build? Factors that Lead to Non-Development or Abandonment of AI Systems SHREYA CHAPPIDI, University of Cambridge, United Kingdom and National Cancer Institute, National Institutes of Health, United States

JATINDER SINGH, Research Centre Trust, UA Ruhr, University Duisburg-Essen, Germany and University of

arXiv:2604.28053v1 [cs.CY] 30 Apr 2026

Cambridge, United Kingdom Responsible AI research typically focuses on examining the use and impacts of deployed AI systems. Yet, there is currently limited visibility into the pre-deployment decisions to pursue building such systems in the first place. Decisions taken in the earlier stages of development shape which systems are ultimately released, and therefore represent potential, but underexplored, points for intervention. As such, this paper investigates factors influencing AI non-development and abandonment throughout the development lifecycle. Specifically, we first perform a scoping review of academic literature, civil society resources, and grey literature including journalism and industry reports. Through thematic analysis of these sources, we develop a taxonomy of six categories of factors contributing to AI abandonment: ethical concerns, stakeholder feedback, development lifecycle challenges, organizational dynamics, resource constraints, and legal/regulatory concerns. Then, we collect data on real-world cases of AI system abandonment via an AI incident database and a practitioner survey to evidence and compare factors that drive abandonment both prior to and following system deployment. While academic responsible AI communities often emphasize ethical risks as reasons to not develop AI, our empirical analysis of these cases demonstrates the diverse, and often non-ethics-related, levers that motivate organizations to abandon AI development. Synthesizing evidence from our taxonomy and related case study analyses, we identify gaps and opportunities in current responsible AI research to (1) engage with the diverse range of levers that influence organizations to abandon AI development, and (2) better support appropriate (dis)engagement with AI system development. CCS Concepts: • Computing methodologies → Artificial intelligence; • Social and professional topics → Computing / technology policy; • Software and its engineering → Software creation and management. Additional Key Words and Phrases: Responsible AI, AI governance, sociotechnical systems, AI lifecycle, decision-making technology abandonment, organisational dynamics, stakeholder engagement, risk management, system development ACM Reference Format: Shreya Chappidi and Jatinder Singh. 2026. To Build or Not to Build? Factors that Lead to Non-Development or Abandonment of AI Systems. In The 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’26), June 25–28, 2026, Montreal, QC, Canada. ACM, New York, NY, USA, 25 pages. https://doi.org/10.1145/3805689.3806736

1

Introduction

There are growing calls and concerns over responsible AI (RAI) development and deployment, across academic, industry, legal/policy, and public spheres. Many academic works interrogate and propose artifacts to improve development practices, including auditing datasets and their creation [56], toolkits for fairness evaluations [94], and approaches to improve model transparency [107]. Emerging regulation seeks to govern algorithms through protections against automated decision-making and scrutiny including impact assessments for high-risk AI systems [68, 120]. There are also growing civil advocacy concerns over increasing AI development, including ACM acknowledges that this contribution was authored or co-authored by an employee, contractor, or affiliate of the United States government. As such, the United States government retains a nonexclusive, royalty-free right to publish or reproduce this article, or to allow others to do so, for government purposes only. Request permissions from owner/author(s). FAccT ’26, Montreal, QC, Canada © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2596-8/2026/06 https://doi.org/10.1145/3805689.3806736

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

journalism and associated databases documenting harms from deployed systems [4, 130] and general skepticism regarding AI hype and overuse [147]. These calls and tools for supporting RAI development often advocate for critical evaluation practices and the appropriate governance mechanisms. At the same time, many critical decisions are made earlier in the AI development lifecycle—including explicit or implicit decisions to pursue AI development in the first place—that are less frequently explored in RAI works [59, 88]. Existing analyses of AI disuse or abandonment largely focus on systems that were deployed and subsequently monitored (either formally or informally) [78]. As such, works addressing AI development challenges often focus on mitigating risks that are surfaced, or could surface, postdeployment (even if the resulting tool or method itself affects pre-deployment stages, such as dataset bias for example) [59, 88, 169]. In contrast, there is limited empirical research investigating challenges in AI development that occur before deployment, particularly in scenarios where AI development is abandoned or not pursued at all, creating multiple gaps in RAI research. First, it is critical to interrogate all blockers to AI development, as they can reveal earlystage failure points in development processes that later stage interventions may not appropriately or effectively address (e.g., fairness evaluations may not always mitigate or reveal issues that stem from decisions made during early-stage problem formulation) [30, 121, 160]. Second, investigating AI abandonment can reveal real-world incentives that shape whether and how AI systems are built. For example, examining what kinds of AI systems are abandoned and for what reasons can highlight what criteria are considered and prioritized during organizational decision-making. A deeper understanding of these broad and diverse incentives can enable the RAI community to better orient and target processes to support practitioners during AI development. Lastly, non-development is also generally an underexplored decision that relates to AI governance. As such, investigating abandoned AI development cases can make non-development more visible and defensible as an RAI practice, and reveal levers that influence decision-making to (not) pursue development. By focusing on cases where AI development is abandoned or not pursued, we seek to identify current gaps in tooling, processes, and incentives that shape AI (non-)development in practice. Definitions. In this paper, we define non-development as organizational decisions to not build a proposed system, and abandonment as organizational decisions to stop building or using an AI system. AI abandonment can occur at any time, including prior to deployment, such as during early-stage data collection or late-stage testing and validation during the AI development lifecycle, or following deployment. We also note that AI abandonment is viewed through a context-dependent lens, and not inherently negative or positive. In some cases, abandonment can be beneficial, preventing development of nonfunctional, unethical, or underresourced AI systems. Conversely, abandonment may be undesirable for well-designed systems that could offer tangible benefits, but face practical challenges leading to or after deployment.

1.1

Contributions

To characterize, analyze, & increase visibility of factors leading to AI abandonment/non-development, this paper: (1) performs a scoping review of academic and grey literature to surface and establishes a taxonomy of factors contributing to AI non-development and abandonment, organizing the factors into six categories: ethical concerns, resource constraints, development lifecycle challenges, legal/regulatory concerns, stakeholder feedback, and organizational dynamics; (2) empirically analyzes real-world cases of AI abandonment collected via a public AI incident database and an online practitioner survey, providing evidence on how both ethical concerns and a diverse range of other factors drive AI abandonment and non-development in real-world contexts; (3) highlights gaps and opportunities in current responsible AI research, tooling, and practice to investigate and engage levers influencing AI non-development and abandonment.

Factors that Lead to Non-Development or Abandonment of AI Systems

2

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Related Work

AI systems research often focuses on model/system performance (including benchmarks, evaluations, and fairness metrics) [43, 49], or the adoption or impacts of using systems that have already been built (and often deployed). However, given broad critiques of automation and other risks posed by AI systems, it is critical to explore conditions that lead to disengagement with technologies, from development to deployment. AI (non-)adoption. Many human-computer interaction (HCI) studies [29, 89, 137] draw upon the technology acceptance model (TAM) [37, 158] to explore end user adoption (including uptake and subsequent patterns of use) of AI systems. For example, Russo [137] examines TAM, diffusion of innovation theory [134], and social cognitive theory to understand why employees adopt generative AI, arguing that adoption is driven by compatibility of tools with existing workflows rather than conventional TAM factors like usefulness. Meanwhile, other industrialorganizational psychology works find that performance expectations, organizational resources [157], cultural values [118] and social dynamics of transferring knowledge [166] drive successful AI adoption. These works offer complex sociotechnical models mediating AI adoption and use, but mainly concern systems that have already been built and often focus on user-specific factors like trust. While adoption studies provide critical insights into whether deployed AI systems will be adopted and used, they can overlook broader factors mediating the development process itself, including initial decisions to even pursue AI development. Thus, to address this gap, we now discuss indicative literature on not pursuing AI development. AI non-functionality or non-use. Other works have proposed non-use of algorithms under specific conditions. Raji et al. [130] create a taxonomy of AI ‘functionality’ issues, arguing for critical reflection over what AI systems can actually do (and not do) and recommend legal, policy, and organizational interventions to improve functional safety of AI systems. Their primary focus on failures relating to the AI system itself can importantly support identification of (in)appropriate AI use cases, but they do not specifically discuss non-development or focus on the “infrastructural or environmental ‘meta’ failures” [130] that may influence decisions to abandon AI development. Focusing on AI nonuse, Bhatt and Sargeant [18] propose algorithmic resignation as a form of governance where deployed algorithms could be intentionally decommissioned under certain conditions. AI non-development. While prior work has examined AI non-adoption at the deployment stage and AI non-use in specific scenarios, there is relatively less work focusing on the conditions that encourage or discourage AI development in the first place. Some academic works [14] and civil tech initiatives [80] urge practitioners to refrain from building unnecessary or objectionable systems. Meanwhile, Mun et al. [112] explicitly examine lay perceptions over the hypothetical impacts of not developing AI tools, finding preferences to avoid technosolutionism but also fears over lost potential benefits from not developing certain use cases. Yet, there has been limited focus on the principle of ‘not building’ in the growing context of real-world AI development. Two key studies offer empirical insights into AI non-development. First, Kawakami et al. [86] use semistructured interviews and design activities to explore how U.S. public sector agencies decide whether to create or adopt new AI tools. They highlight pressure from external organizations as critical influences on public sector decision-making, indicating the need to explore these factors across broader contexts and application domains. Next, Johnson et al. [78] investigate 40 public cases of AI abandonment, outlining a common six-stage process of “campaigns to abandon an algorithm.” The work provides important foundations for investigating decomissioned AI systems; however, it conceptualizes abandonment as a reactive process concerned over the ability of an algorithm to cause harm. As such, the work primarily analyzes cases of post-deployment AI abandonment due to real or potential harms, and does not explore other reasons why AI development may be abandoned (e.g., resource constraints, etc.). While Johnson et al. [78] offer a process to specifically describe how organizations abandon AI systems, there remains a gap in establishing the factors why organizations abandon AI development, particularly where systems are not deployed or lack clear potential for harm.

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

Summary. Further research is needed to investigate why and how why individuals, teams, or organizations choose to pursue AI development, and also how development processes might be challenged pre-deployment. Both cases of choosing to not develop or stop developing can result in systems that never get deployed, rendering these cases underexplored in RAI literature. Our work aims towards this gap, investigating factors mediating decisions to abandon AI development across diverse domains and industries.

3

A taxonomy of factors contributing to AI abandonment

We establish a taxonomy of factors contributing to AI abandonment to improve understanding of both 1) decisionmaking processes to pursue AI development and 2) diverse failure points, challenges, and trade-offs that can lead to abandonment. We developed our taxonomy through a scoping multivocal literature review [53] that includes both academic and grey literature sources [101], including civil society organizations and resources, and technology journalism. The high-level goal of this research was to identify What factors contribute to AI system non-development or abandonment? Methodology. To build the taxonomy, we perform a multivocal scoping literature review, including both academic and grey literature, to explore influences on why AI development may be abandoned or not pursued. We incorporate grey literature, including civil society resources, technology journalism, and industry reports and white papers, to capture other timely, pragmatic factors that could drive organizational decision-making and may be overlooked in traditional academic databases [53, 54, 85]. Academic Literature Databases ● Google Scholar ● ACM Digital Library ● Semantic Scholar [“Not” OR “stop” OR “abandon” OR“halt”] AND [“develop” OR “build”] AND [“AI” OR “ML”]

Civil Society Resources ● AI Harm/Risk/Incident Repositories ● Advocacy groups (e.g., Don’t Build It)

Grey Literature ● Web search (Google), including ○ Industry surveys ○ White papers ○ Tech journalism [“Not” OR “stop” OR “abandon” OR“halt”] AND [“develop” OR “build”] AND [“AI” OR “ML”]

1. Initial Literature Probe (48 sources selected) 2. Snowballing (32 sources added) 3. Inductive Thematic Analysis (6 categories identified) Ethical Concerns

Organizational

Dynamics

Development Lifecycle Challenges

Stakeholder Feedback

Resource Constraints

Legal/ Regulatory Concerns

4. Category-Specific Literature Search (80 sources added) 5. Establish Taxonomy of Factors Contributing to AI System Abandonment 6. Apply Taxonomy to Analyze Real-World Cases of AI System Abandonment

Fig. 1. Overall methodology, including scoping literature workflow, used to establish and analyze categories of factors contributing to AI abandonment.

Using search terms and sources detailed in Figure 1, we performed preliminary queries across academic literature databases and grey literature sources (web search) for relevant publications discussing factors that could influence not developing or abandoning/stopping the development of AI systems. Then, we reviewed the titles and abstracts of the first 50 search results from each search source, and catalogued for further review those discussing formal mechanisms (e.g., exit plans, injunctions or suspensions) or more informal challenges (e.g.,

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

public backlash, data collection issues, etc.) that could influence pausing or stopping AI development. Through this process, we identified 48 initial sources. Then, using a snowballing approach [168] we explored relevant citations included within these 48 initial sources, gathering 32 additional relevant sources. Since the goal of this scoping review was to identify broad, diverse factors contributing to AI abandonment, we did not systematically record information on works screened out as not relevant. From the collected works that referenced formal and informal factors influencing AI development decisions, inductive thematic analysis [22] was performed by the lead researcher to identify factors that could contribute to AI development challenges and abandonment. These factors were grouped into higher-order categories based on shared themes, which were iterated and discussed with the broader research team until consensus was reached. After establishing the six broad categories of factors contributing to abandonment, we performed categoryspecific literature searches to address any gaps from the initial literature review and ensure relevant coverage. For example, after we established development lifecycle challenges (see §3.3) as a contributing category, we performed additional searches related to both ‘AI/ML development lifecycle’ and its individual stages (e.g., ‘problem formulation,’). These category-specific searches increased the reviewed corpus by an additional 80 sources. The six categories identified via inductive thematic analysis and indicative literature review works underpinning each relevant factor are detailed in Table 1, with further contextual discussions on the individual factors now presented in §3.1–§3.6. Ethical Concerns (adapted from [145])

Stakeholder Feedback

Development Lifecycle Challenges

Discrimination Could not measure target variable [62, 90, 152] Stakeholder Toxicity engagement [84] OR Challenging to collect desired data [12, 103] Privacy & security Resistance [40] from: Dataset too small [16, 35, 156] Misinformation • Employees [1, 115] Lacked ground truth [61, 62] Too challenging to pre-process/curate data Malicious actors & • Users [40, 81] [122, 139, 163] misuse • Data subjects [5] Too challenging to label data [17, 57, 150, 172] Concerns over • Civil society Inappropriate model selection [99, 130] overreliance [36, 147] Concerns over • Domain experts [67] Model training too technically challenging • Activists [5] [32, 99] unsafe use Model training too resource intensive [41, 170] Loss of human Undefined or inappropriate success/evaluation agency & criteria [19, 99] autonomy Model performance not sufficient [3, 130] Labor Difficult to integrate into pipelines [13, 137] displacement Inability to conduct evaluation/pilot [130] Environmental Low adoption in pilot deployment [45] concerns [69]

Organizational Dynamics Changing incentives/priorities [74] Not indicated/scoped well [20, 97, 130] Not aligned with organizational strategy [131] Insufficient leadership sponsorship [96, 106] Clients/customers did not want it [55, 147] Concerns over keeping company assets private [32]

Resource Constraints

Legal/Regulatory Concerns

Lacked technical AI regulations expertise to build [39] compliance [8, 46, 69] Lacked technical Data protection expertise to maintain regulations [67] [28, 117, 125] Domain-specific Too expensive to regulations [114, 132] build [55] Too expensive to Deemed too high risk maintain [155] [46] Compute availability Lack of regulatory guidance [127] concerns [10, 23] Compute cost Liability concerns concerns [23, 167] [138, 143, 154] Development timeline Intellectual property too long [67, 91] concerns [7, 102, 173] Cheaper to outsource [32, 144, 164] Easier to outsource [13, 144, 164]

Table 1. Taxonomy of factors contributing to AI non-development or abandonment, with illustrative references from the reviewed literature corpus indicated.

3.1

Non-Development Category 1: Ethical Concerns

Our analysis of factors mediating AI abandonment identified growing ethical concerns over the individual and/or societal risks of developing AI systems that could be grouped into three main categories.1 First, there are concerns that AI deployments can bring negative societal impacts, including if they lead to job loss, or provide inconclusive evidence due to its probabilistic nature [109]. Second, there are concerns over AI being used for ends themselves deemed unethical across various principles (e.g., weapons deployment) [1]. Finally, there are critiques that the process of constructing and implementing AI systems itself can perpetuate individual or societal harms, such as 1 Since our analysis is guided by the overarching question of factors motivating AI abandonment, we discuss ethical concerns in this paper

from a perspective of that likely used in AI organizational practice, rather than through a more formal moral or philosophical lens.

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

by extracting resources or labor from already marginalized communities [58, 108] or enabling discrimination, often through the guise of ‘objectivity’ via automation [109]. Many RAI frameworks have analyzed societal risks and related ethical issues within AI systems [72, 133, 151, 162, 169], including AI risk repositories. Slattery et al. [145] perform a comprehensive meta-review of 777 AI risks identified across 43 separate taxonomies, synthesizing the reviewed taxonomies into 7 main risk domains. To reflect common themes identified across AI risk taxonomies, these risk domains were adapted from Slattery et al. [145] in §3 to include factors involving Discrimination, Toxicity, issues with Privacy & security (also discussed in §3.4), Misinformation, Malicious actors & misuse, Concerns over overreliance, Concerns over unsafe use, Loss of human agency & autonomy, Labor displacement, and Environmental concerns (including the hardware and energy resources required to sustain development [41, 69, 170]).

3.2

Non-Development Category 2: Stakeholder Feedback

Stakeholder feedback may affect decisions to abandon AI development when explicitly solicited through §3.2.1 stakeholder engagement performed by organizations, or when externally imposed through §3.2.2 resistance and/or collective actions. Note that feedback may often communicate information reflected in other categories of factors driving abandonment (e.g., concerns over ethics, low model performance, cost of development). As such, this section specifically considers the influence of the stakeholder feedback itself (i.e., that the relevant concern is amplified or only taken seriously when levied through a stakeholder feedback mechanism). In these cases, abandonment may be driven by key considerations including but not limited to reputational risks [77], fear of repercussions, or general lack of public support [78]. 3.2.1 Stakeholder engagement. The RAI community has advocated for stakeholder involvement throughout the development lifecycle to surface potential risks and harms, and shape AI system design to address actual stakeholder needs [15, 33, 66, 82–84, 113]. Stakeholder engagement frameworks have been proposed across application domains, including education [52], medicine [161], and public sector systems [87], and may also be scoped by [66] or support [15] regulatory guidance. Businesses also often rely on stakeholder engagement during project scoping and requirements gathering to ensure downstream success with users and clients, financial viability, and possibly avoid legal fees [84]. For example, agile techniques advocate for frequent collaboration with end-users and customers to shape development processes [81, 82]. Through these processes, organizations may receive feedback from stakeholders to not build certain features or systems, or surface potential challenges to development that may lead to AI abandonment. Many stakeholder engagement processes may only involve those directly tied to commercial interests [66, 84]; yet, other relevant stakeholders include those impacted by the system (e.g., communities subject to surveillance), employees building and maintaining systems, and domain-specific actors such as caregivers and administrators in the case of medicine [67]. Nedzhvetskaya and Tan [115] highlight the critical role of workers as stakeholders in AI development, arguing that while system “designers” are often given more prominence and power, “trainers” (including those generating, cleaning, and processing data) are often more likely to be subject to harm from such algorithms. Ajmani et al. [5] also advocate improved engagement with secondary stakeholders, including data contributors and activists. At the same time, it can be challenging to identify cases where stakeholder involvement influences AI abandonment, as evidence is less likely to be published on systems that don’t materalize. 3.2.2 Resistance and/or collective action. Stakeholder groups may also be mobilized outside of formal stakeholder engagement process, ‘disseminating’ feedback via offline or online mechanisms [78] or engaging in explicit resistance. Employees tasked with developing AI systems can advocate abandonment of AI projects, with Abdalla [1] analyzing technology worker resistance actions, frequently employed to halt AI development contracts over

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

purposes found objectionable by employees (e.g., military uses). Collective action by employees through organized labor [51] or by affected communities [124] may also be used to shape and resist AI development practices [21, 115]. Resistance can also be employed by other parties, including data subjects (those whose data is used to train such algorithms [24]) or those whose jobs would be replaced or altered by AI development [140]. For example, users may push back against development of AI systems through prompt injections [2], jailbreaks [63, 142, 171], opting-out, and other guides on avoiding AI [140, 147]. DeVrio et al. [40] taxonomize how people respond “from below” towards algorithmic harms, including refusing legitimate engagement with systems. Collective labor movements have also fought for protections against AI development by those potentially subject to labor replacement, including actors and writer’s guilds [9] and specialized professions [26, 140]. While these forms of actual or proposed user resistance may appear to reflect low AI adoption, these collective actions have implications for AI development, as they can prohibit, ban, or require oversight on upstream AI development [9, 28].

3.3

Non-Development Category 3: Development Lifecycle Challenges

Common challenges during AI development can be pinpointed to different stages of the AI development lifecycle. These issues may drive abandonment if the challenges cannot be overcome or ultimately create a nonfunctional AI system. As outlined in §3, we performed broad searches into literature specifically on challenges appearing throughout the AI development lifecycle [12, 30, 95, 119, 128, 149], including problem formulation; data collection, pre-processing, and curation; model building, training, and fine-tuning; and model performance evaluation and monitoring. Then, we performed targeted literature searches on the broad stages described in §3.3.1-§3.3.4 (adapted from [30, 119]) to identify additional factors that may prevent completion of the development lifecycle. 3.3.1 Problem formulation. Problem formulation involves translating a real-world problem into a well-defined task, including selecting appropriate target variables, metrics, and evaluation criteria. More specifically, problem formulation has been linked as an early AI development process shaping later, or downstream, stages of fairness assessment [75, 76, 121], ground truth construction [92, 111, 152], and performance evaluation [123, 160]. Yet, Mao et al. [105] report that ML practitioners often struggle to understand alternative ways to construct questions and hypotheses. Kernahan et al. [90] investigate ML target variable construct invalidity, finding that many algorithms are unable to measure their target variable, thus contributing to implementation failure. Undefined or inappropriate success/evaluation criteria is a challenge surfaced in later stages often involving problem formulation choices [19, 99]. 3.3.2 Data curation. Data curation involves collecting, cleaning, labelling, and preprocessing data for use in training and evaluation of ML models. It may be challenging to collect desired data [19, 135], leading to missing and imbalanced datasets that result in performance issues and bias in high-stakes settings, including clinical decision-making [6, 44, 47, 98, 153]. Datasets that are too small can result in overfitting [16, 35, 156], potentially leading organizations to abandon certain use cases in favor of other uses cases enabled by data they can access. Other studies of ML practitioners and general ML data practices have identified challenges in pre-processing and curating data [159], including opaque data processing habits and insufficient handling of data from minority protected attributes [122, 139, 163]. It can also be challenging to label data in various domains, particularly as labor-intensive, subjective labeling processes may create label bias and low inter-rater reliability [17, 150]. Lack of ground truth target labels is also a frequent challenge [27, 61, 62], and can introduce measurement bias to systems when proxies for an actual yet unmeasurable outcome are used [38, 50, 61, 152, 172]. 3.3.3 Model building. Inappropriate model selection can influence AI functionality and abandonment of systems [130]. For example, models may be too simple, too complex, or involve incorrect assumptions for the available data [99]. Increasingly complex model architectures and/or algorithmic supply chains can also make model training too technically challenging. For example, deep learning architectures often require

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

sophisticated interpretability techniques to investigate non-linear interactions [99], and foundation models can require additional database/software dependencies that complicate AI system development [32, 70, 99, 165]. Model training and inference can also be resource intensive lifecycle stages [170], influencing decision-making where resources such as GPUs or data center storage are limited. 3.3.4 Evaluating & monitoring real-world performance. Many AI systems may be abandoned at the evaluation stage if model performance is not sufficient, including incorrect predictions [3], hallucinations [11, 25], or unexpected system behavior [130]. Depending on prior system design decisions, evaluations may suffer from undefined or inappropriate success/evaluation criteria [19, 99]. For example, inappropriate evaluation metrics may prevent rigorous statistical evaluations [19] or provide misleading results (e.g., F1 score better captures performance over accuracy for highly imbalanced datasets) [99]. Pilot deployments may also influence abandonment [45], where low adoption indicates that the developed system will not be used. Lastly, integration into existing workflows is a key factor mediating success of AI system deployments [13, 60, 106, 137].

3.4

Non-Development Category 4: Legal/Regulatory Concerns

Law and regulation, including those principle-based [34, 48, 125, 148] or domain-specific, can directly impact whether system development proceeds. This section considers the challenges of complying with law and regulation and the consequences of non-compliance which may impact AI system non-development and abandonment. Concerns over complying with (emerging) AI regulation may stop or slow down AI development [46], including instances of ‘regulatory flight’ [8] where organizations choose to not develop technologies in certain regions due to more stringent regulatory restrictions. Non-development may occur where systems are deemed too high risk, either explicitly prohibited if deemed to be of ‘unacceptable’ risk under regulation such as the EU AI Act [69] or bringing additional ‘high-risk’ regulatory compliance requirements that organizations may not wish to undertake [46]. Lack of regulatory guidance, or uncertainty over how to interpret emerging regulations, may also slow or hamper development [48, 71, 127]. Law and regulation concerning data rights, protection, and ownership can also shape decision-making surrounding AI development. Data protection mechanisms such as the EU’s GDPR can impose restrictions on appropriate data collection, storage, and use [125] that shape organizational decisions to pursue AI development [117]. For example, Chappidi et al. [28] find evidence that record-keeping mechanisms and requirements, including those mandated by GDPR, can influence and even dissuade practitioners from building systems that record data subject to such protections. Similarly, domain-specific data privacy and protection regulations can also shape decision-making to pursue AI development. For example, systems using individually identifiable health information (PHI) may be subject to additional regulatory requirements [114], bringing additional compliance costs, considerations, and potential regulatory bodies/actions. Generative AI integrations have also driven scrutiny around data ownership rights, copyright law [125, 173], and intellectual property [102]. Compliance with privacy and copyright protection laws may include opt-out provisions [7], machine unlearning [7, 173], or system removal altogether. Abandonment over intellectual property concerns may also occur given ongoing lawsuits seeking injunctions on AI system sales over copyright infringements [102]. Lastly, legal concerns can also emerge outside specific internal and external regulatory frameworks, including through fear of liability risks and lawsuits [104]. Organizations may seek to avoid risks of AI development involving product liability [130], or stakeholders may hold domain-specific concerns over individual legal responsibility when AI systems are developed and used, including in medicine [64, 67, 126, 136, 154].

3.5

Non-Development Category 5: Organizational Dynamics

Organizational dynamics, including tensions between individuals, teams, and organizations and their own incentives, workflows, and mandates, can shape AI development practices and related decisions to abandon AI

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

systems. Organizational incentives to pursue technology development [146] often involve decisions on revenuegenerating directions for the organization [106]. Leader sponsorship is also a critical factor shaping outcomes of AI system development, with Marwaha et al. [106] arguing the necessity of both an “internal champion” and “executive sponsor.” AI development at the organization-level has been driven by hopes of profitability, [73], “AI-first” policies [110], and other incentives. At the same time, changing priorities or incentives can lead to abandonment of AI development strategies, such as retail company Klarna backtracking on its AI-first policy [74]. Increasing algorithmic supply chain complexity, including through growing AI integrations with foundation models, may bring concerns over keeping company, user, or client information or assets private due to limited visibility into other parts of the supply chain [32]. Misalignment or lack of alignment between organizational strategies and individual/team operations can affect the success of AI development. Rakova et al. [131] study organizational RAI implementation, identifying “misalignment between individual and team incentives and org-level mission statements.” Kawakami et al. [86] also identify disconnects between frontline worker concerns and employees with higher institutional power across public sector agencies developing AI. Organizations may also struggle with faithful translation between goals, data, and computational problems [83, 121], resulting in projects not well scoped/indicated [20, 97, 130].

3.6

Non-Development Category 6: Resource Constraints

Resource constraints may also drive organizational decision-making to not develop or abandon development. AI systems may be too expensive to build, bringing costs associated with the system itself (e.g., employee compensation, hardware, energy, etc. [65]) and those related to mitigating risks (e.g., stakeholder engagement, legal/regulatory compliance). Comparing development costs to value generated by such systems, IBM Institute for Business Value [73] find that only 25% of CEOs currently report returns on investment in AI. Systems also carry additional costs to maintain, potentially evidenced by reports that only 16% of AI initiatives have scaled across enterprises [155]. Compute cost and compute availability concerns can also mediate decision-making, with reported shortages in GPUs critical for model training [10], increasing demand for data centers [36], and energy bottlenecks in sustaining data centers [23]. Costs to maintain generative AI frontier models significantly outpace revenue currently generated by subscriptions, creating further concerns of costs increasingly being passed to users, including through price hikes [167]. Another constrained resource may include technical expertise to build and maintain AI systems. AI systems are growing in complexity due to increasing data needs, complex ML architectures, and dependencies introduced by other AI system components [32, 70, 116]. Organizations may lack necessary resources to maintain AI systems [67]. For example, some tech executives report that AI development is scaling down (due to complexity) [141], while 54% of surveyed CEOs report hiring for AI-related roles that did not exist a year ago [73]. AI abandonment may also be influenced by development timelines that are too long [67, 91]. Resource constraints relating to costs, time, and expertise may also drive abandonment of in-house development in favor of outsourcing development [13, 91, 164] or contracting, purchasing, or procuring off-the-shelf products [13, 71, 79]. Summary. Table 1 establishes and summarizes the six categories of factors contributing to AI abandonment identified from inductive thematic analysis applied to our scoping literature review. While existing work on AI nonuse or abandonment tends to focus on factors mediating technology adoption and related user-focused models, our taxonomy presents an expanded lens to analyze AI abandonment throughout the development lifecycle. As such, our taxonomy emphasizes both decisions made prior to system deployment and factors beyond user-level interactions, including broader organizational and related stakeholder dynamics. While categories in §3 are presented separately for clarity, these factors are often linked/related across categories, and decisions to abandon AI development are likely shaped by multiple factors overlapping, interacting, and/or layering on top of each other. For example, stakeholder feedback may raise concerns over ethical risks, likely

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

development lifecycle challenges, or reveal conflicting organizational dynamics. As such, our taxonomy enables the identification of multiple categories that layer and interact to eventually trigger abandonment of AI development.

4

Analyzing Real-World Cases of AI Abandonment

Next, we sought to gather evidence and analyze how the factors identified in §3 influenced decisions to abandon AI system development in real-world practice. Overall Methods. We gathered and analyzed case studies on abandoned AI systems from both AI incident repositories and an empirical AI practitioner survey. Data from the AI incident repository in §4.1 enabled analysis of factors leading to abandonment in publicly documented systems (mostly reflecting those reaching deployment stages), while the practitioner survey in §4.2 augmented our investigations with further data on systems that are abandoned at various stages, including pre-deployment. Details on case filtering, data collection, and analyses are described in §4.1.1 and §4.2.1-§4.2.2. We present initial case details and summaries in §4.3, and discuss broader trends and findings on real-world AI system abandonment in §5.

4.1

Analyzing Abandoned AI Case Studies via AI Incident Repositories

The AI, Algorithmic and Automation Incidents and Controversies (AIAAIC) Repository crowdsources documentation of algorithmic and/or AI-related risks, harms, and incidents [4]. Since the AIAAIC repository relies on publicly available details and reports, many entries originate from deployed systems and frequently discuss ethical concerns. We analyzed AIAAIC case studies to investigate factors contributing to why AI systems are abandoned. 4.1.1 Methods. The AIAAIC Repository web database (accessed 18/10/2025) was reviewed for curated data categories, including system details like the purpose, technology type, and developers/deployers, and impacts of the harm or controversy. We filtered for cases using the ‘Response Taxonomy’ column, which describes actions taken by the individuals or organisations who developed/deployed the AI system, including employee termination, public apology, policy review, and system suspension. We chose to filter on the ‘Response Taxonomy’ field as it would most closely capture the perspectives and decision-making by the relevant party to abandon AI development. We selected all AIAAIC cases mentioning key terms of [‘termination,’ ‘suspension,’ ‘pause,’ ‘closure,’ ‘scrapped,’ ‘deletion,’ ‘cancellation,’ ’retraction,’ ‘loss,’ ’disabled,’ ‘withdrawal,’ ‘halted,’ or ‘removal’] in the ‘Response’ column to collect abandoned AI system cases. Then, the lead researcher reviewed relevant case details summarized directly within the repository entry and links to external coverage of the incident. Cases were excluded from further analysis if they did not involve decisions to not develop or stop developing AI systems (e.g., merely updating the system, eventually reinstating the system, only removing problematic data from training datasets). Finally, we applied the taxonomy identified in §3 to categorize and analyze factors leading to AI abandonment. Details from these sources were used to identify and code 1) the stage at which the AI system was abandoned (e.g., project scoping, data curation) and 2) relevant factors contributing to AI abandonment under the taxonomy outlined in the §3.

4.2

Analyzing Abandoned AI Case Studies via Practitioner Survey

Building on work conducted in §4.1, we sought to expand our analysis of abandoned AI systems across diverse domains and industries, past those specifically scoped as ‘incidents’ and likely to be mature enough in development to meet news reporting thresholds. As such, we collected further abandoned AI case studies through an empirical online practitioner survey to increase our reach, diversity, and spread of analyzed abandoned AI cases. Practitioners were asked to report, detail, and analyze cases where they abandoned AI development. 4.2.1 Study population and recruitment. The survey targeted individuals involved in commissioning, designing, developing, implementing, and/or monitoring AI systems, including software engineers, product/project managers,

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

data scientists, and stakeholders in compliance, legal, and ethics roles. Participants needed to self-report at least one relevant role; otherwise, the survey was terminated early and their responses were excluded. Recruitment occurred via open calls through professional networks (including LinkedIn, Slack, emailing lists, and relevant workshops), consistent with methods in similar studies [82]. Participation was voluntary and anonymous, with human research subjects approval obtained through the department ethics committee. 4.2.2 Survey design. We designed a practitioner survey to collect case studies on instances where AI development projects were abandoned or continued, and to explore factors and processes influencing these decisions. The survey collected cases where an individual, team, and/or organization chose to not develop or stop developing an AI system for any reason. The survey also asked practitioners to report cases where their organization continued AI development, despite their assessment that the system should have been abandoned. Both of these survey sections gathered details about the AI system, including its purpose, relevant training data, user population, and the application domain. Participants reported the stage at which development was (or should have been) abandoned and indicated relevant factors towards abandonment using category-specific select-all-the-apply questions based on §3. The order of these category-specific questions was randomized to minimize ordering effects. Lastly, the survey collected details on the decision-making process contributing to (or preventing) abandonment, including who was involved, tools or methods used, and under what circumstances the AI system might be reconsidered for future development. Finally, participant demographics were collected, including role, years of experience, industry, and country (reported in the Appendix).

4.3

Results

This section presents descriptive analyses of our collected case studies to characterize the diverse categories of factors contributing to AI non-development and abandonment, with further qualitative insights presented in §5. 4.3.1 AI Incident cases. Out of 2109 total entries (as of 10/2025), 143 AIAAIC cases included a suspension key word (§4.1.1) in the ‘Response’ category. 52 AIAAIC cases were excluded from further analysis for 1) not involving AI systems specifically (e.g., incidents concerning dataset content, scams incorrectly claiming AI involvement) or 2) not involving an AI system reported as abandoned (e.g., system was updated and/or re-introduced). We ultimately identified 91 AIAAIC cases involving AI abandonment for further analysis, ranging from 2016 to 2025. 85% (n=77) of analyzed incidents were abandoned post-deployment, with 5% abandoned at the project proposal stage and 5% abandoned at the pilot deployment stage. Ethical concerns were reported as contributing to abandonment for 81 cases, with 51% (n=41) raising concerns over bias/discrimination and 51% (n=41) involving privacy and security concerns such as potential for surveillance. Stakeholder feedback was relevant towards abandonment for 48 cases, frequently involving public backlash on social media from affected users or data subjects; expert criticism from civil rights, privacy, and other advocacy groups; and investigative reporting from news outlets and think tanks. Legal/regulatory concerns reportedly contributed to abandonment in 38 AIAAIC cases, with data protection regulations and domain-specific regulations (e.g., anti-discrimination laws in housing, hiring, etc.) shaping concerns over AI systems. Development lifecycle challenges were identified in 26 cases, with 23 specifically reporting issues or challenges with model performance during pilot or formal deployment. Resource constraints and organizational dynamics also influenced abandonment in 4 and 3 AIAAIC cases, respectively. 4.3.2 Survey cases. A total of 36 complete survey responses were collected, with 8 responses removed from participants who did not report involvement with design, development, or deployment of AI systems. Of the remaining 28 practitioner responses, we collected 12 cases of abandoned AI system development, 5 cases of AI systems that participants reported should have been abandoned in their view (but development was pushed forward regardless), and 23 cases of AI system development reported as justifiable or suitable. Participants reported a wide range of factors contributing to AI abandonment, including most frequently development lifecycle

AI System Description ID Abandoned?

Chatbot for public assistance program applications P1 AI-mediated e-commerce integration w/ web search P2 Clustering to identify fraudulent crypto users P3 Clustering to identify high-risk labor operations P4 Inference to improve data quality gaps P5 In-house LLM development (to compete w/industry) P6 LLM-based infrastructure support for SWEs P7 Concept extraction from larger NLP model P8 LLM-based notification summaries for SWEs P9 RAG system for sustainability business contacts P10 Decision trees for COVID-19 travel warnings P11 LLMs for functional form identification P12 LLM summarization of congressional/policy press P13 Predict system bugs using troubleshooting logs P14 LLM support to improve vulnerability management P15 Extract lease details from unstructured documents P16 Public health outbreak detection P17

Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No No No No No

Stage Relevant to Abandonment

Deployment Model retraining Model retraining Problem operationalization Data collection Model evaluation Problem formulation Problem operationalization Problem operationalization Data collection Model evaluation Model evaluation During development During development During development Before development During development

✓

✓ ✓

✓

✓

✓ ✓ ✓ ✓ ✓ ✓

✓

✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

CoRes ns ou tra rce in ts Le ga l/R Co eg nc ul er ato ns r y

Chappidi & Singh

CoEth nc ica er l ns Or ga Dy niz na atio m n ics al De ve l L o Ch ifec pm all ycl ent en e ge s Sta Fe keh ed ol ba de ck r

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

✓ ✓ ✓

✓ ✓ ✓ ✓ ✓

✓ ✓

✓ ✓ ✓ ✓

✓

✓ ✓ ✓

Table 2. Surveyed case studies indicate a wide array of factors contributing to abandonment, with many reporting abandonment prior to deployment. P13-P17 reflect cases where participants reported that the system should have been abandoned due to the indicated factors, but the organization continued with development. Factor-level details contributing to cases indicating abandonment are reported in A1. SWEs = software engineers

challenges, followed by resource constraints, organizational dynamics, and stakeholder feedback. Factors contributing to abandonment are detailed by category in Table 2 and individually in Figure A1. Summary. A majority of AIAAIC cases analyzed indicated ethical concerns and related stakeholder feedback as key factors contributing to AI abandonment (Table 3). This finding is expected since these cases were gathered via a crowdsourced AI incident database cataloging public sources more likely to describe mature, deployed AI systems. Less than 5% of all AIAAIC repository entries involved AI abandonment, indicating that non-development is not a frequently documented system outcome. We also identified other non-ethics-related factors driving abandonment across these cases, including legal/regulatory concerns and development lifecycle challenges. In contrast, cases collected via survey tended to report non-ethics concerns as influencing abandonment (see Tables 2 and 3). Surveyed cases also usually reported multiple categories of factors as contributing to abandonment, most frequently including organizational dynamics, development lifecycle challenges, and resource constraints. While 85% of AIAAIC cases were abandoned post-deployment, many of the surveyed abandoned cases reported issues prior to deployment, including during problem operationalization and model adjustment. As such, our initial analysis of abandoned AI cases via survey indicate a greater influence of internal factors in deciding to abandon AI development, particularly where systems do not make it to the deployment stage.

5

Discussion

We now discuss and synthesize findings from our analyses of abandoned AI case studies gathered via AI incident repository and practitioner surveys. We also highlight gaps and opportunities for the RAI community to consider, address, and engage with AI abandonment and non-development as practices influencing RAI development and governance. Practitioner survey IDs [P#] assigned in Table 2 and AIAAIC-generated source ID codes [AIAAIC#] providing evidence are indicated in brackets where relevant.

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Factors Contributing to Abandonment

Ethical Concerns

Organizational Dynamics

Development Lifecycle Challenges

Stakeholder Feedback

Resource Constraints

Legal/ Regulatory Concerns

# of AIAAIC cases (n=91) # of survey cases (n=17)

81

10

23

45

4

37

3

9

14

8

9

1

Table 3. Our analysis of AI incident repository cases naturally indicates that ethical concerns, stakeholder feedback, and legal/regulatory concerns are primary influences on AI abandonment. Meanwhile, surveyed cases indicating abandonment frequently report factors beyond those 3 categories. Darker shaded cells reflect a higher proportion of cases in the category.

Ethical concerns are not the only factors that contribute to abandoned AI development. Many abandoned systems analyzed from the AI incident repository identified concerns over potential or realized harm, including discrimination, toxicity, etc. This finding is not surprising given that these repositories seek to categorize instances of harm created by algorithms, and reflects the utility of these resources in establishing real-world risks and harms created by AI system development. At the same time, cases collected via practitioner survey frequently reported non-ethics reasons across resource constraints, organizational dynamics, and challenges during the AI development lifecycle as key factors in deciding to abandon AI development. While it remains critical to address ethical concerns when deciding whether to pursue AI development, we identify and evidence diverse, frequently overlapping drivers or levers influencing decision-making to abandon AI development. These levers can indicate different avenues to motivate change within organizations, and the relative influence and relationship between these levers should be further explored and investigated in RAI work. Few surveyed participants reported using formal processes to determine whether to even proceed with building AI. While other works discuss public responses and influences on decision-making to abandon harmful algorithms [40, 78], our surveyed cases often involved abandoned AI systems that did not explicitly flag ethical concerns or receive negative stakeholder feedback. As such, these scenarios often required internal reflections on whether to abandon AI development. Most participants reported no specific individual, team, or organizational decisionmaking processes, including P7 specifically flagging the need for “a centralized AI strategy for the organization.” Some RAI works advocate for formal processes to rule-in AI use, including problem formulation [130] and exploration of alternative methods [99]. Yet, very few RAI artifacts explicitly advocate or address AI abandonment as an available, acceptable, or appropriate decision throughout the development process. This is echoed by Kawakami et al. [87] finding that “most existing toolkits assume that the decision to develop a particular AI system has already been made” and Wong et al. [169] discussing ‘solutionism’ present in RAI toolkits that do not suggest “fundamental changes to the corporate values systems or business models that may lead to harms from AI system.” Resources that reference not commissioning or de-commissioning systems [151] often do not highlight or suggest conditions where abandonment may be appropriate. As such, our findings emphasize the need for processes that guide organizations in reflecting on conditions indicating or requiring AI abandonment. Resource constraints can underpin many additional factors driving abandonment, bringing nuance to decision-making behind non-development. Many surveyed case studies reported resource constraints as a key factor in deciding to abandon AI development. These survey responses often indicated that procurement of existing tools was cheaper and easier, and that they lacked sufficient compute resources or dataset sizes [P6] to effectively maintain and justify their AI system. Resource constraints may often drive factors indicating abandonment in other categories of the taxonomy. For example, insufficient dataset sizes within development lifecycle challenges may be driven by insufficient funds or technical expertise to collect, curate, and appropriately pre-process data. In the case of La Buona Scuola teachers’ mobility algorithm in Italy [AIAAIC0705], a lack of technical expertise to build the system

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

(translated as “a bombastic, redundant and non-maintenance-oriented system” by Lopera [100] from an Italian audit report) resulted in an inability to conduct an evaluation of the system, ultimately driving abandonment. Costs may be an additional (and potentially more salient) lever that encourages non-development or abandonment for organizations, even for AI systems with identified ethical concerns. For example, in the case of Alpaca [AIAAIC1029], a language model criticized for ethical concerns over misinformation, a Stanford spokesperson only discussed in their public statement that costs to maintain factored into the abandoned AI development [93, 129]. Other categories of abandonment factors may also carry their own associated costs that influence overall decision-making to pursue AI development. For example, emphasizing the costs of lawsuits or regulatory probes into data protection compliance issues could make abandonment a more salient option for AI systems with ethical privacy & security risks. Similarly, emphasizing the costs associated with systems abandoned post-deployment due to stakeholder resistance can make the upfront costs of stakeholder engagement [84, 88] more reasonable. Early-stage challenges in AI development lifecycle often influence abandonment. We identified two cases where organizations could not measure their target variable indicating the critical influence of early-stage decisionmaking. P4 abandoned their AI system at the problem operationalization stage, noting that “although full ground truth is challenging to obtain, it is possible to collect offending vessels from news reports, or use IUU (illegal, unreported and unregulated) fishing list as a proxy.” In contrast, the criminality prediction algorithm reported in AIAAIC0467 was abandoned after development, indicating potential challenges in identifying that the AI problem design was not possible [130]. Moreover, two surveyed case studies reported challenges in labeling data [P3, P5] and four reported challenges in obtaining ground truth [P3, P4, P5, P12]; yet, both P3 and P12 reported that development was abandoned at model training stages. This indicates that they may have proceeded past the data curation stage despite challenges, or that they only identified early-stage issues at the evaluation stage (e.g., P12 reporting “we realized that the method our colleague was set on [...] was not scientifically rigorous or of potential interest to our broader research community.”) Thus, our work indicates the potentially beneficial role of deliberating on early-stage development challenges to identify abandoned AI projects sooner. Resurrection of abandoned AI systems is a likely outcome. Many reviewed AIAAIC cases were excluded from final inclusion because the system was not actually abandoned, or in some cases “reincarnated” (as termed by Johnson et al. [78]). For example, reviewed AI systems were often suspended in one locality (e.g., Uber suspended their vehicles in Tempe, but then continued operations in Pittsburgh [AIAAIC0187]), for a certain period of time (e.g., Malta stopped their ‘Safe City’ video surveillance program, but then appointed a new board to oversee the initiative years later [AIAAIC1051]), or re-introduced after a feature was tweaked (e.g., HireVue removed facial analysis from its screening algorithm [AIAAIC0579]). These findings reflect how projects abandoned in the present can still have impacts (or ‘imprints’ [42]) in the future. Many surveyed cases also reported that their organization was likely to revisit development of abandoned systems, including under conditions involving “better employee bandwidth” [P1], “more time” [P2], and “if we have close partnership with front line organisations” [P4]. Surveyed cases also indicated that their organization realized it was cheaper or easier to outsource development [P6, P8, P9]. As such, our analyses echo the growing influence and relevance of algorithmic supply chains [32], AI-as-a-Service contracts [31] (e.g., P16 reporting that they warned of issues over hallucinations but their client “favored an AI-integrated solution first”), and procurement in RAI development [71, 79].

5.1

Summary & Recommendations

Our taxonomy developed in §3 and analysis of real-world abandoned AI development cases in §4 indicate a diverse set of factors or levers influencing AI non-development. We call for further attention and reflection on AI abandonment and non-development as a multi-faceted practice, with the following recommendations for responsible AI communities:

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Recommendations 1. Responsible AI communities should investigate, interrogate, and reflect on cases of AI abandonment occuring prior to deployment to better understand current practices, challenges, and gaps in AI system development. By documenting and analyzing cases of AI abandonment at all stages of development, the community can gain a more complete picture of drivers and blockers towards responsible AI development, including those non-ethics-related. 2. Artifacts and tooling should explicitly increase visibility and outline discussions on non-development or abandonment, particularly where organizations identify that they lack appropriate resources, expertise, or capacity to develop their desired system. 3. Responsible AI development toolkits, frameworks, and artifacts can make more explicit the influence of resources, including costs, technical expertise, and development timelines, on the potential success of other development stages. For example, encouraging organizations to explicitly outline the costs of collecting sufficiently large datasets or maintaining deployed systems can use resource constraints as a lever to discourage development of AI systems that also carry ethical risks or would face development lifecycle challenges. 4. Responsible AI tooling should explicitly encourage organizations to reflect on, outline, and update conditions required to appropriately re-visit or resurrect development of abandoned AI systems.

6

Limitations & Future Work

The taxonomy presented in §3 is meant to illustrate the diverse levers that can facilitate non-development rather than be exhaustive, and additional factors may emerge in future work. While we aimed to review all available details, analysis of abandoned AIAAIC cases may not capture every single factor that ultimately contributed to AI abandonment, as these databases largely concern deployed systems and external reports may lack visibility into organizational dynamics or resource constraints. Relatedly, data on non-development and abandoned projects is hard to collect as these systems receive fewer resources, are often less formally documented, and can be less salient to practitioners. Our real-world data analyses affirmed this, as very few systems in the AIAAIC repository were reported as abandoned and surveyed practitioners were more inclined to report on continued AI developments rather than abandoned ones. Thus, our work reflects a first step towards this gap by engaging and gathering insights from organizations that abandoned AI development across diverse domains and purposes, particularly in cases prior to system deployment. Future work can continue this approach, encouraging researchers, organizations, and advocacy groups to 1) continue collection, analysis, and sharing of abandoned AI cases across all lifecycle stages, 2) conduct further empirical analysis of incentives and drivers preventing abandonment, and 3) examine the broader impacts of AI system removal.

7

Concluding Remarks

Our work seeks to increase visibility into AI non-development, departing from and complementing other works by conceptualizing abandonment as a practice enabled by diverse levers present throughout the AI development lifecycle. While academic and civil society resources frequently emphasize ethical risks as reasons to not develop AI systems, we find that decisions to abandon development often involve factors beyond ethical concerns, including development lifecycle challenges, organizational dynamics, and resource constraints. Through this work, we indicate real-world incentives and levers that influence organizations to pursue or abandon AI development, and in turn, highlight actionable opportunities to better align RAI research and interventions with the factors that actually influence organizational decision-making. We urge the RAI community to further investigate AI abandonment, particularly including those abandoned in early-stages or prior to deployment, and advocate for updates to ongoing research and tooling to support appropriate (dis)engagement with AI development.

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

Generative AI Usage Statement Asta (allen.ai) was used to identify links to potentially relevant academic sources, with any resulting academic sources reviewed directly in the externally linked/original source location (e.g., full PDF upload). Limited editorial support was provided from LLMs to adjust formatting commands for LaTeX tables (Copilot integration in VSCode, GPT-5 mini) and support ideation on rephrasing select statements (GPT-4o & 5 mini).

Acknowledgments We would like to thank the practitioner survey participants for their responses, as well as Anna Neumann and Anna Ida Hudig for their feedback on study materials. SC is a PhD student in the NIH Oxford-Cambridge Scholars Program. This research was supported [in part] by the Intramural Research Program of the National Institutes of Health (NIH). The contributions of the NIH author(s) are considered Works of the United States Government. The findings and conclusions presented in this paper are those of the author(s) and do not necessarily reflect the views of the NIH or the U.S. Department of Health and Human Services.

References [1] Mohamed Abdalla. 2025. $100,000 or the Robot Gets It! Tech Workers’ Resistance Guide: Tech Worker Actions, History, Risks, Impacts, and the Case for a Radical Flank. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society 8, 1 (Oct. 2025), 2–14. doi:10.1609/aies.v8i1.36526 [2] William Agnew, Harry H. Jiang, Cella Sum, Maarten Sap, and Sauvik Das. 2024. Data Defenses Against Large Language Models. (2024). doi:10.48550/ARXIV.2410.13138 Publisher: arXiv Version Number: 1. [3] Zo Ahmed. 2024. AI coding assistants do not boost productivity or prevent burnout, study finds. https://www.techspot.com/news/ 104945-ai-coding-assistants-do-not-boost-productivity-or.html [4] AIAAIC. 2025. AI, Algorithmic and Automation Incidents and Controversies (AIAAIC). https://www.aiaaic.org/ [5] Leah Hope Ajmani, Nuredin Ali Abdelkadir, and Stevie Chancellor. 2025. Secondary Stakeholders in AI: Fighting for, Brokering, and Navigating Agency. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 1095–1107. doi:10.1145/3715275.3732071 [6] Saeed Amal, Lida Safarnejad, Jesutofunmi A. Omiye, Ilies Ghanzouri, John Hanson Cabot, and Elsie Gyang Ross. 2022. Use of MultiModal Data and Machine Learning to Improve Cardiovascular Disease Care. Frontiers in Cardiovascular Medicine 9 (April 2022), 840262. doi:10.3389/fcvm.2022.840262 [7] Archer Amon, Zhipeng Yin, Zichong Wang, Avash Palikhe, Tongjia Yu, and Wenbin Zhang. 2026. Uncertain Boundaries: Multidisciplinary Approaches to Copyright Issues in Generative AI. SIGKDD Explor. Newsl. 27, 2 (Dec. 2026), 1–12. doi:10.1145/3787470.3787472 [8] Markus Anderljung, Joslyn Barnhart, Anton Korinek, Jade Leung, Cullen O’Keefe, Jess Whittlestone, Shahar Avin, Miles Brundage, Justin Bullock, Duncan Cass-Beggs, Ben Chang, Tantum Collins, Tim Fist, Gillian Hadfield, Alan Hayes, Lewis Ho, Sara Hooker, Eric Horvitz, Noam Kolt, Jonas Schuett, Yonadav Shavit, Divya Siddarth, Robert Trager, and Kevin Wolf. 2023. Frontier AI Regulation: Managing Emerging Risks to Public Safety. doi:10.48550/arXiv.2307.03718 arXiv:2307.03718 [cs]. [9] Dani Anguiano and Lois Beckett. 2023. How Hollywood writers triumphed over AI – and why it matters. The Guardian (Oct. 2023). https://www.theguardian.com/culture/2023/oct/01/hollywood-writers-strike-artificial-intelligence [10] Guido Appenzeller, Matt Bornstein, and Martin Casado. 2023. Navigating the High Cost of AI Compute. https://a16z.com/navigatingthe-high-cost-of-ai-compute/ [11] Zahra Ashktorab, Michael Desmond, Qian Pan, James Johnson, Michelle Brachman, Casey Dugan, Marina Danilevsky, and Werner Geyer. 2025. Emerging Reliance Behaviors in Human-AI Content Grounded Data Generation: The Role of Cognitive Forcing Functions and Hallucinations. Proceedings of the 4th Annual Symposium on Human-Computer Interaction for Work (June 2025), 1–17. doi:10.1145/ 3729176.3729179 Conference Name: CHIWORK ’25: Proceedings of the 4th Annual Symposium on Human-Computer Interaction for Work ISBN: 9798400713842 Place: Amsterdam Netherlands Publisher: ACM. [12] Rob Ashmore, Radu Calinescu, and Colin Paterson. 2022. Assuring the Machine Learning Lifecycle: Desiderata, Methods, and Challenges. Comput. Surveys 54, 5 (June 2022), 1–39. doi:10.1145/3453444 [13] Deepika Badampudi, Claes Wohlin, and Kai Petersen. 2016. Software component decision-making: In-house, OSS, COTS or outsourcing - A systematic literature review. Journal of Systems and Software 121 (Nov. 2016), 105–124. doi:10.1016/j.jss.2016.07.027 [14] Eric P.S. Baumer and M. Six Silberman. 2011. When the implication is not to design (technology). In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ’11). Association for Computing Machinery, New York, NY, USA, 2271–2274. doi:10.1145/1978942.1979275

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

[15] Andrew Bell, Oded Nov, and Julia Stoyanovich. 2023. Think about the stakeholders first! Toward an algorithmic transparency playbook for regulatory compliance. Data & Policy 5 (Jan. 2023), e12. doi:10.1017/dap.2023.8 [16] Visar Berisha, Chelsea Krantsevich, P. Richard Hahn, Shira Hahn, Gautam Dasarathy, Pavan Turaga, and Julie Liss. 2021. Digital medicine and the curse of dimensionality. npj Digital Medicine 4, 1 (Oct. 2021), 153. doi:10.1038/s41746-021-00521-5 [17] Mélanie Bernhardt, Daniel C. Castro, Ryutaro Tanno, Anton Schwaighofer, Kerem C. Tezcan, Miguel Monteiro, Shruthi Bannur, Matthew P. Lungren, Aditya Nori, Ben Glocker, Javier Alvarez-Valle, and Ozan Oktay. 2022. Active label cleaning for improved dataset quality under resource constraints. Nature Communications 13, 1 (March 2022), 1161. doi:10.1038/s41467-022-28818-3 Publisher: Nature Publishing Group. [18] Umang Bhatt and Holli Sargeant. 2024. When Should Algorithms Resign? A Proposal for AI Governance. Computer 57, 10 (Oct. 2024), 99–103. doi:10.1109/MC.2024.3431328 [19] Stella Biderman and Walter J. Scheirer. 2021. Pitfalls in Machine Learning Research: Reexamining the Development Cycle. doi:10. 48550/arXiv.2011.02832 arXiv:2011.02832 [cs]. [20] Emily Black, Rakshit Naidu, Rayid Ghani, Kit Rodolfa, Daniel Ho, and Hoda Heidari. 2023. Toward Operationalizing Pipeline-aware ML Fairness: A Research Agenda for Developing Practical Guidelines and Tools. In Equity and Access in Algorithms, Mechanisms, and Optimization. ACM, Boston MA USA, 1–11. doi:10.1145/3617694.3623259 [21] William Boag, Harini Suresh, Bianca Lepe, and Catherine D’Ignazio. 2022. Tech Worker Organizing for Power and Accountability. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22). Association for Computing Machinery, New York, NY, USA, 452–463. doi:10.1145/3531146.3533111 [22] Virginia Braun and Victoria Clarke. 2006. Using thematic analysis in psychology. Qualitative Research in Psychology 3, 2 (Jan. 2006), 77–101. doi:10.1191/1478088706qp063oa Publisher: Routledge _eprint: https://doi.org/10.1191/1478088706qp063oa. [23] Claire Burch. 2025. Deep Dive: Economics of the AI Build-Out. Technical Report. Contrary Research. https://research.contrary.com/ report/the-economics-of-ai-build-out [24] Matt Burgess. 2024. How to Stop Your Data From Being Used to Train AI. Wired (Oct. 2024). https://www.wired.com/story/how-tostop-your-data-from-being-used-to-train-ai/ Section: tags. [25] Garance Burke and Hilke Schellmann. 2024. Researchers say an AI-powered transcription tool used in hospitals invents things no one ever said. https://apnews.com/article/ai-artificial-intelligence-health-business-90020cdf5fa16c79ca2e5b6c4c9bbb14 Section: Technology. [26] Center for Labor and a Just Economy. 2025. Bargaining Victory: SEIU Local 688 won strong protections for workers and the public in AI agreement with Pennsylvania Governor Shapiro. https://clje.law.harvard.edu/news/bargaining-victory-seiu-local-688-won-strongprotections-for-workers-and-the-public-in-ai-agreement-with-pennsylvania-governor-shapiro/ [27] Shreya Chappidi, Mason J. Belue, Stephanie A. Harmon, Sarisha Jagasia, Ying Zhuge, Erdal Tasci, Baris Turkbey, Jatinder Singh, Kevin Camphausen, and Andra V. Krauze. 2025. From manual clinical criteria to machine learning algorithms: Comparing outcome endpoints derived from diverse electronic health record data modalities. PLOS Digital Health 4, 5 (May 2025), e0000755. doi:10.1371/journal.pdig. 0000755 [28] Shreya Chappidi, Jennifer Cobbe, Chris Norval, Anjali Mazumder, and Jatinder Singh. 2025. Accountability Capture: How Record-Keeping to Support AI Transparency and Accountability (Re)shapes Algorithmic Oversight. doi:10.48550/arXiv.2510.04609 arXiv:2510.04609 [cs]. [29] Hyesun Choung, Prabu David, and Arun Ross. 2023. Trust in AI and Its Role in the Acceptance of AI Technologies. International Journal of Human–Computer Interaction 39, 9 (May 2023), 1727–1739. doi:10.1080/10447318.2022.2050543 [30] Jennifer Cobbe, Michelle Seng Ah Lee, and Jatinder Singh. 2021. Reviewable Automated Decision-Making: A Framework for Accountable Algorithmic Systems. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. ACM, Virtual Event Canada, 598–609. doi:10.1145/3442188.3445921 [31] Jennifer Cobbe and Jatinder Singh. 2021. Artificial intelligence as a service: Legal responsibilities, liabilities, and policy challenges. Computer Law & Security Review 42 (2021), 105573. doi:10.1016/j.clsr.2021.105573 [32] Jennifer Cobbe, Michael Veale, and Jatinder Singh. 2023. Understanding accountability in algorithmic supply chains. In 2023 ACM Conference on Fairness, Accountability, and Transparency. ACM, Chicago IL USA, 1186–1197. doi:10.1145/3593013.3594073 [33] Eric Corbett, Remi Denton, and Sheena Erete. 2023. Power and Public Participation in AI. In Proceedings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO ’23). Association for Computing Machinery, New York, NY, USA, 1–13. doi:10.1145/3617694.3623228 [34] Nicholas Kluge Corrêa, Camila Galvão, James William Santos, Carolina Del Pino, Edson Pontes Pinto, Camila Barbosa, Diogo Massmann, Rodrigo Mambrini, Luiza Galvão, Edmund Terem, and Nythamar De Oliveira. 2023. Worldwide AI ethics: A review of 200 guidelines and recommendations for AI governance. Patterns 4, 10 (Oct. 2023), 100857. doi:10.1016/j.patter.2023.100857 [35] James L. Cross, Michael A. Choma, and John A. Onofrey. 2024. Bias in medical AI: Implications for clinical decision-making. PLOS Digital Health 3, 11 (Nov. 2024), e0000651. doi:10.1371/journal.pdig.0000651

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

[36] Data Center Watch. 2025. $64 billion of data center projects have been blocked or delayed amid local opposition. https://www. datacenterwatch.org/report [37] Fred D. Davis. 1989. Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology. MIS Quarterly 13, 3 (1989), 319–340. doi:10.2307/249008 Publisher: Management Information Systems Research Center, University of Minnesota. [38] Matthew DeCamp and Charlotta Lindvall. 2020. Latent bias and the implementation of artificial intelligence in medicine. Journal of the American Medical Informatics Association 27, 12 (Dec. 2020), 2020–2023. doi:10.1093/jamia/ocaa094 [39] Innovation & Technology Department for Science and Media & Sport Department for Digital, Culture. 2021. Quantifying the UK Data Skills Gap - Full report. https://www.gov.uk/government/publications/quantifying-the-uk-data-skills-gap/quantifying-the-uk-dataskills-gap-full-report [40] Alicia DeVrio, Motahhare Eslami, and Kenneth Holstein. 2024. Building, Shifting, & Employing Power: A Taxonomy of Responses From Below to Algorithmic Harm. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’24). Association for Computing Machinery, New York, NY, USA, 1093–1106. doi:10.1145/3630106.3658958 [41] Jesse Dodge, Taylor Prewitt, Remi Tachet des Combes, Erika Odmark, Roy Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A. Smith, Nicole DeCario, and Will Buchanan. 2022. Measuring the Carbon Intensity of AI in Cloud Instances. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22). Association for Computing Machinery, New York, NY, USA, 1877–1894. doi:10.1145/3531146.3533234 [42] Upol Ehsan, Ranjit Singh, Jacob Metcalf, and Mark Riedl. 2022. The Algorithmic Imprint. In 2022 ACM Conference on Fairness Accountability and Transparency. ACM, Seoul Republic of Korea, 1305–1317. doi:10.1145/3531146.3533186 [43] Maria Eriksson, Erasmo Purificato, Arman Noroozian, Joao Vinagre, Guillaume Chaslot, Emilia Gomez, and David Fernandez-Llorca. 2025. Can We Trust AI Benchmarks? An Interdisciplinary Review of Current Issues in AI Evaluation. doi:10.48550/arXiv.2502.06559 arXiv:2502.06559 [cs]. [44] Andre Esteva, Brett Kuprel, Roberto A. Novoa, Justin Ko, Susan M. Swetter, Helen M. Blau, and Sebastian Thrun. 2017. Dermatologistlevel classification of skin cancer with deep neural networks. Nature 542, 7639 (Feb. 2017), 115–118. doi:10.1038/nature21056 [45] Sheryl Estrada. 2025. MIT report: 95% of generative AI pilots at companies are failing. https://fortune.com/2025/08/18/mit-report-95percent-generative-ai-pilots-at-companies-failing-cfo/ [46] Aidatul Fitriyah and Daryna Dzemish Abdulovna. 2024. EU’s AI Regulation Approaches and Their Implication for Human Rights. Media Iuris 7, 3 (Oct. 2024), 417–438. doi:10.20473/mi.v7i3.62050 [47] Alyssa M. Flores, Alejandro Schuler, Anne Verena Eberhard, Jeffrey W. Olin, John P. Cooke, Nicholas J. Leeper, Nigam H. Shah, and Elsie G. Ross. 2021. Unsupervised Learning for Automated Detection of Coronary Artery Disease Subgroups. Journal of the American Heart Association 10, 23 (Dec. 2021), e021976. doi:10.1161/JAHA.121.021976 [48] Luciano Floridi and Josh Cowls. 2019. A Unified Framework of Five Principles for AI in Society. Harvard Data Science Review 1, 1 (July 2019). doi:10.1162/99608f92.8cd550d1 Publisher: The MIT Press. [49] James Fodor. 2025. Line Goes Up? Inherent Limitations of Benchmarks for Evaluating Large Language Models. doi:10.48550/arXiv.2502. 14318 arXiv:2502.14318 [cs]. [50] Riccardo Fogliato, Alexandra Chouldechova, and Max G’Sell. 2020. Fairness Evaluation in Presence of Biased Noisy Labels. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics. PMLR, 2325–2336. https://proceedings.mlr.press/ v108/fogliato20a.html ISSN: 2640-3498. [51] Richard B. Freeman and James L. Medoff. 1979. The Two Faces of Unionism. https://papers.ssrn.com/abstract=261218 [52] Caterina Fuligni, Daniel Domínguez Figaredo, and Julia Stoyanovich. 2025. "Would You Want an AI Tutor?"Understanding Stakeholder Perceptions of LLM-based Systems in the Classroom. https://www.semanticscholar.org/paper/ 469a07c98826b0d937ae846b4e1c3ca0f1d0d84f [53] Vahid Garousi, Michael Felderer, and Mika V. Mäntylä. 2019. Guidelines for including grey literature and conducting multivocal literature reviews in software engineering. Information and Software Technology 106 (Feb. 2019), 101–121. doi:10.1016/j.infsof.2018.09.006 [54] Vahid Garousi, Michael Felderer, Mika V. Mäntylä, and Austen Rainer. 2020. Benefitting from the Grey Literature in Software Engineering Research. In Contemporary Empirical Methods in Software Engineering, Michael Felderer and Guilherme Horta Travassos (Eds.). Springer International Publishing, Cham, 385–413. doi:10.1007/978-3-030-32489-6_14 [55] Gartner. 2024. Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025. https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projectswill-be-abandoned-after-proof-of-concept-by-end-of-2025 [56] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. 2021. Datasheets for Datasets. http://arxiv.org/abs/1803.09010 arXiv:1803.09010 [cs]. [57] Apoorva Gondimalla, Varshinee Sreekanth, Govind Joshi, Whitney Nelson, Eunsol Choi, Stephen C. Slota, Sherri R. Greenberg, Kenneth R. Fleischmann, and Min Kyung Lee. 2024. Aligning Data with the Goals of an Organization and Its Workers: Designing Data Labeling for Social Service Case Notes. In Proceedings of the CHI Conference on Human Factors in Computing Systems. ACM, Honolulu HI USA, 1–21. doi:10.1145/3613904.3642014

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

[58] Mary L. Gray and Siddharth Suri. 2019. Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. https://www.semanticscholar.org/paper/Ghost-Work%3A-How-to-Stop-Silicon-Valley-from-a-New-Gray-Suri/ 47f936331872d75df74473883bb65068c14fa7da [59] Daniel Greene, Anna Lauren Hoffmann, and Luke Stark. 2019. Better, Nicer, Clearer, Fairer: A Critical Assessment of the Movement for Ethical Artificial Intelligence and Machine Learning. In Proceedings of the 52nd Hawaii International Conference on System Sciences (Critical and Ethical Studies of Digital and Social Media). doi:10.24251/HICSS.2019.258 [60] Trisha Greenhalgh, Joseph Wherton, Chrysanthi Papoutsi, Jennifer Lynch, Gemma Hughes, Christine A’Court, Susan Hinder, Nick Fahy, Rob Procter, and Sara Shaw. 2017. Beyond Adoption: A New Framework for Theorizing and Evaluating Nonadoption, Abandonment, and Challenges to the Scale-Up, Spread, and Sustainability of Health and Care Technologies. Journal of Medical Internet Research 19, 11 (Nov. 2017), e367. doi:10.2196/jmir.8775 [61] Luke Guerdan, Amanda Coston, Zhiwei Steven Wu, and Kenneth Holstein. 2023. Ground(less) Truth: A Causal Framework for Proxy Labels in Human-Algorithm Decision-Making. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’23). Association for Computing Machinery, New York, NY, USA, 688–704. doi:10.1145/3593013.3594036 [62] Luke Guerdan, Devansh Saxena, Stevie Chancellor, Zhiwei Steven Wu, and Kenneth Holstein. 2025. Measurement as Bricolage: Examining How Data Scientists Construct Target Variables for Predictive Modeling Tasks. (2025). doi:10.48550/ARXIV.2507.02819 Publisher: arXiv Version Number: 3. [63] Hangzhi Guo, Pranav Narayanan Venkit, Eunchae Jang, Mukund Srinath, Wenbo Zhang, Bonam Mingole, Vipul Gupta, Kush R. Varshney, S. Shyam Sundar, and Amulya Yadav. 2024. Hey GPT, Can You be More Racist? Analysis from Crowdsourced Attempts to Elicit Biased Content from Generative AI. doi:10.48550/arXiv.2410.15467 arXiv:2410.15467 [cs]. [64] Zach Harned, Matthew P. Lungren, and Pranav Rajpurkar. 2019. Machine Vision, Medical AI, and Malpractice. https://papers.ssrn. com/abstract=3442249 [65] Will Henshall. 2024. The Billion-Dollar Price Tag of Building AI. https://time.com/6984292/cost-artificial-intelligence-compute-epochreport/ [66] Eleanore Hickman and Martin Petrin. 2021. Trustworthy AI and Corporate Governance: The EU’s Ethics Guidelines for Trustworthy Artificial Intelligence from a Company Law Perspective. European Business Organization Law Review 22, 4 (Dec. 2021), 593–625. doi:10.1007/s40804-021-00224-0 [67] Henry David Jeffry Hogg, Mohaimen Al-Zubaidy, Technology Enhanced Macular Services Study Reference Group, James Talks, Alastair K. Denniston, Christopher J. Kelly, Johann Malawana, Chrysanthi Papoutsi, Marion Dawn Teare, Pearse A. Keane, Fiona R. Beyer, and Gregory Maniatopoulos. 2023. Stakeholder Perspectives of Clinical Artificial Intelligence Implementation: Systematic Review of Qualitative Evidence. Journal of Medical Internet Research 25, 1 (Jan. 2023), e39742. doi:10.2196/39742 Company: Journal of Medical Internet Research Distributor: Journal of Medical Internet Research Institution: Journal of Medical Internet Research Label: Journal of Medical Internet Research Publisher: JMIR Publications Inc., Toronto, Canada. [68] Tomasz Hollanek, Yulu Pi, Cosimo Fiorini, Virginia Vignali, Dorian Peters, and Eleanor Drage. 2025. A Toolkit for Compliance, a Toolkit for Justice: Drawing on Cross-sectoral Expertise to Develop a Pro-justice EU AI Act Toolkit. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 1184–1194. doi:10.1145/3715275.3732078 [69] Tomasz Hollanek, Yulu Pi, Dorian Peters, Selen Yakar, and Eleanor Drage. 2025. The EU AI Act in Development Practice: A Pro-justice Approach. doi:10.48550/arXiv.2504.20075 arXiv:2504.20075 [cs]. [70] Aspen Hopkins, Sarah H. Cen, Isabella Struckman, Andrew Ilyas, Luis Videgaray, and Aleksander Mądry. 2025. AI Supply Chains: An Emerging Ecosystem of AI Actors, Products, and Services. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society 8, 2 (Oct. 2025), 1266–1277. doi:10.1609/aies.v8i2.36628 [71] Anna Ida Hudig, Emma Kallina, and Jatinder Singh. 2026. “It’s Just a Wild, Wild West”: Harnessing Public Procurement as an AI Governance Mechanism. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. ACM, Barcelona, Spain, 22. doi:10.1145/3772318.3791968 [72] Wiebke Hutiri, Orestis Papakyriakopoulos, and Alice Xiang. 2024. Not My Voice! A Taxonomy of Ethical and Safety Harms of Speech Generators. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’24). Association for Computing Machinery, New York, NY, USA, 359–376. doi:10.1145/3630106.3658911 [73] IBM Institute for Business Value. 2025. 2025 CEO Study - 5 mindshifts to supercharge business growth. Technical Report 32nd edition. https://www.ibm.com/downloads/documents/us-en/12f5a711174dc2ac [74] Irina Ivanova. 2025. As Klarna flips from AI-first to hiring people again, a new landmark survey reveals most AI projects fail to deliver. https://fortune.com/2025/05/09/klarna-ai-humans-return-on-investment/ [75] Abigail Z. Jacobs. 2021. Measurement as governance in and for responsible AI. http://arxiv.org/abs/2109.05658 arXiv:2109.05658 [cs]. [76] Abigail Z. Jacobs and Hanna Wallach. 2021. Measurement and Fairness. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21). Association for Computing Machinery, New York, NY, USA, 375–385. doi:10.1145/3442188. 3445901

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

[77] Pierre Le Jeune, Jiaen Liu, Luca Rossi, and Matteo Dora. 2025. RealHarm: A Collection of Real-World Language Model Application Failures. (2025). doi:10.48550/ARXIV.2504.10277 Version Number: 1. [78] Nari Johnson, Sanika Moharana, Christina Harrington, Nazanin Andalibi, Hoda Heidari, and Motahhare Eslami. 2024. The Fall of an Algorithm: Characterizing the Dynamics Toward Abandonment. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’24). Association for Computing Machinery, New York, NY, USA, 337–358. doi:10.1145/3630106.3658910 [79] Nari Johnson, Elise Silva, Harrison Leon, Motahhare Eslami, Beth Schwanke, Ravit Dotan, and Hoda Heidari. 2025. Legacy Procurement Practices Shape How U.S. Cities Govern AI: Understanding Government Employees’ Practices, Challenges, and Needs. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 772–789. doi:10.1145/3715275.3732049 [80] Luke Jordan. 2021. Don’t Build It: A Guide For Practitioners In Civic Tech / Tech For Development. https://mitgovlab.org/resources/dontbuild-it-a-guide-for-practitioners-in-civic-tech/ [81] Gabriela Jurca, Theodore D. Hellmann, and Frank Maurer. 2014. Integrating Agile and User-Centered Design: A Systematic Mapping and Review of Evaluation and Validation Studies of Agile-UX. In Proceedings of the 2014 Agile Conference (AGILE ’14). IEEE Computer Society, USA, 24–32. doi:10.1109/AGILE.2014.17 [82] Emma Kallina, Thomas Bohné, and Jatinder Singh. 2025. Stakeholder Participation for Responsible AI Development: Disconnects Between Guidance and Current Practice. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 1060–1079. doi:10.1145/3715275.3732069 [83] Emma Kallina, Constanze M Leeb, and Jatinder Singh. 2026. The Limits of Stakeholder Participation in Safety-Critical Contexts: Lessons from Air Traffic Control. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, 1–19. doi:10.1145/3772318.3790959 [84] Emma Kallina and Jatinder Singh. 2024. Stakeholder Involvement for Responsible AI Development: A Process Framework. In Proceedings of the 4th ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization. ACM, San Luis Potosi Mexico, 1–14. doi:10.1145/3689904.3694698 [85] Fernando Kamei, Igor Wiese, Crescencio Lima, Ivanilton Polato, Vilmar Nepomuceno, Waldemar Ferreira, Márcio Ribeiro, Carolline Pena, Bruno Cartaxo, Gustavo Pinto, and Sérgio Soares. 2021. Grey Literature in Software Engineering: A critical review. Information and Software Technology 138 (Oct. 2021), 106609. doi:10.1016/j.infsof.2021.106609 [86] Anna Kawakami, Amanda Coston, Hoda Heidari, Kenneth Holstein, and Haiyi Zhu. 2024. Studying Up Public Sector AI: How Networks of Power Relations Shape Agency Decisions Around AI Design and Use. Proc. ACM Hum.-Comput. Interact. 8, CSCW2 (Nov. 2024), 450:1–450:24. doi:10.1145/3686989 [87] Anna Kawakami, Amanda Coston, Haiyi Zhu, Hoda Heidari, and Kenneth Holstein. 2024. The Situate AI Guidebook: Co-Designing a Toolkit to Support Multi-Stakeholder, Early-stage Deliberations Around Public Sector AI Proposals. Proceedings of the CHI Conference on Human Factors in Computing Systems (May 2024), 1–22. doi:10.1145/3613904.3642849 Conference Name: CHI ’24: CHI Conference on Human Factors in Computing Systems ISBN: 9798400703300 Place: Honolulu HI USA Publisher: ACM. [88] Anna Kawakami, Daricia Wilkinson, and Alexandra Chouldechova. 2024. Do Responsible AI Artifacts Advance Stakeholder Goals? Four Key Barriers Perceived by Legal and Civil Stakeholders. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society 7, 1 (Oct. 2024), 670–682. doi:10.1609/aies.v7i1.31669 [89] Sage Kelly, Sherrie-Anne Kaye, and Oscar Oviedo-Trespalacios. 2023. What factors contribute to the acceptance of artificial intelligence? A systematic review. Telematics and Informatics 77 (Feb. 2023), 101925. doi:10.1016/j.tele.2022.101925 [90] Jacqueline Kernahan, Richard Bartels, Mark de Reuver, Daniel Oberski, and Roel Dobbe. 2025. Burying the Lead: Adjusting Goals to Manage Functional Limitations of AI Tools in Healthcare. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society 8, 2 (Oct. 2025), 1401–1412. doi:10.1609/aies.v8i2.36640 [91] Miikka Kuutila, Mika Mäntylä, Umar Farooq, and Maëlick Claes. 2020. Time Pressure in Software Engineering: A Systematic Review. Information and Software Technology 121 (May 2020), 106257. doi:10.1016/j.infsof.2020.106257 arXiv:1901.05771 [cs]. [92] Himabindu Lakkaraju, Jon Kleinberg, Jure Leskovec, Jens Ludwig, and Sendhil Mullainathan. 2017. The Selective Labels Problem: Evaluating Algorithmic Predictions in the Presence of Unobservables. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, Halifax NS Canada, 275–284. doi:10.1145/3097983.3098066 [93] Frank Landymore. 2023. Stanford Pulls Down ChatGPT Clone After Safety Concerns. https://futurism.com/the-byte/stanford-pullsdown-chatgpt-clone [94] Michelle Seng Ah Lee and Jat Singh. 2021. The Landscape and Gaps in Open Source Fairness Toolkits. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM, Yokohama Japan, 1–13. doi:10.1145/3411764.3445261 [95] Michelle Seng Ah Lee and Jatinder Singh. 2021. Risk Identification Questionnaire for Detecting Unintended Bias in the Machine Learning Development Lifecycle. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society (AIES ’21). Association for Computing Machinery, New York, NY, USA, 704–714. doi:10.1145/3461702.3462572 [96] Nick Lichtenberg. 2025. This CEO laid off nearly 80% of his staff because they refused to adopt AI fast enough. 2 years later, he says he’d do it again. https://fortune.com/2025/08/17/ceo-laid-off-80-percent-workforce-ai-sabotage/

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

[97] Hongjin Lin, Anna Kawakami, Catherine D’Ignazio, Kenneth Holstein, and Krzysztof Gajos. 2025. Funding AI for Good: A Call for Meaningful Engagement. (2025). doi:10.48550/ARXIV.2509.12455 Publisher: arXiv Version Number: 2. [98] Jana Lipkova, Richard J. Chen, Bowen Chen, Ming Y. Lu, Matteo Barbieri, Daniel Shao, Anurag J. Vaidya, Chengkuan Chen, Luoting Zhuang, Drew F.K. Williamson, Muhammad Shaban, Tiffany Y. Chen, and Faisal Mahmood. 2022. Artificial intelligence for multimodal data integration in oncology. Cancer Cell 40, 10 (Oct. 2022), 1095–1110. doi:10.1016/j.ccell.2022.09.012 [99] Michael A. Lones. 2024. Avoiding common machine learning pitfalls. Patterns 5, 10 (Oct. 2024), 101046. doi:10.1016/j.patter.2024.101046 [100] Ricardo Zapata Lopera. 2020. The algorithm that decided the destiny of thousands of families. https://rzapatal.medium.com/thealgorithm-that-decided-the-destiny-of-many-families-7b374fa2574b [101] Qinghua Lu, Liming Zhu, Xiwei Xu, Jon Whittle, Didar Zowghi, and Aurelie Jacquet. 2024. Responsible AI Pattern Catalogue: A Collection of Best Practices for AI Governance and Engineering. Comput. Surveys 56, 7 (July 2024), 1–35. doi:10.1145/3626234 [102] Nicola Lucchi. 2024. ChatGPT: A Case Study on Copyright Challenges for Generative Artificial Intelligence Systems. European Journal of Risk Regulation 15, 3 (Sept. 2024), 602–624. doi:10.1017/err.2023.59 [103] Michael Madaio, Lisa Egede, Hariharan Subramonyam, Jennifer Wortman Vaughan, and Hanna Wallach. 2022. Assessing the Fairness of AI Systems: AI Practitioners’ Processes, Challenges, and Needs for Support. Proceedings of the ACM on Human-Computer Interaction 6, CSCW1 (March 2022), 1–26. doi:10.1145/3512899 [104] Gennie Mansi and Mark Riedl. 2025. Implications of Current Litigation on the Design of AI Systems for Healthcare Delivery. doi:10.48550/arXiv.2507.15981 arXiv:2507.15981 [cs] version: 1. [105] Yaoli Mao, Dakuo Wang, Michael Muller, Kush R. Varshney, Ioana Baldini, Casey Dugan, and AleksandraMojsilović. 2019. How Data Scientists Work Together With Domain Experts in Scientific Collaborations: To Find The Right Answer Or To Ask The Right Question? Proceedings of the ACM on Human-Computer Interaction 3, GROUP (Dec. 2019), 1–23. doi:10.1145/3361118 arXiv:1909.03486 [cs]. [106] Jayson S. Marwaha, Adam B. Landman, Gabriel A. Brat, Todd Dunn, and William J. Gordon. 2022. Deploying digital health tools within large, complex health systems: key considerations for adoption and implementation. npj Digital Medicine 5, 1 (Jan. 2022), 13. doi:10.1038/s41746-022-00557-1 Publisher: Nature Publishing Group. [107] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. 2019. Model Cards for Model Reporting. In Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT* ’19). Association for Computing Machinery, New York, NY, USA, 220–229. doi:10.1145/3287560.3287596 [108] Shakir Mohamed, Marie-Therese Png, and William Isaac. 2020. Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence. Philosophy & Technology 33, 4 (Dec. 2020), 659–684. doi:10.1007/s13347-020-00405-8 [109] Jessica Morley, Luciano Floridi, Libby Kinsey, and Anat Elhalal. 2020. From What to How: An Initial Review of Publicly Available AI Ethics Tools, Methods and Research to Translate Principles into Practices. Science and Engineering Ethics 26, 4 (Aug. 2020), 2141–2168. doi:10.1007/s11948-019-00165-5 [110] Chris Morris. 2025. Going ’AI first’ appears to be backfiring on Klarna and Duolingo. https://www.fastcompany.com/91332763/goingai-first-appears-to-be-backfiring-on-klarna-and-duolingo [111] Michael Muller, Christine T. Wolf, Josh Andres, Michael Desmond, Narendra Nath Joshi, Zahra Ashktorab, Aabhas Sharma, Kristina Brimijoin, Qian Pan, Evelyn Duesterwald, and Casey Dugan. 2021. Designing Ground Truth and the Social Life of Labels. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM, Yokohama Japan, 1–16. doi:10.1145/3411764.3445402 [112] Jimin Mun, Liwei Jiang, Jenny Liang, Inyoung Cheong, Nicole DeCario, Yejin Choi, Tadayoshi Kohno, and Maarten Sap. 2024. ParticipAI: A Democratic Surveying Framework for Anticipating Future AI Use Cases, Harms and Benefits. doi:10.48550/arXiv.2403.14791 arXiv:2403.14791 [cs]. [113] Rashid Mushkani, Hugo Berard, Allison Cohen, and Shin Koeski. 2025. The Right to AI. (2025). doi:10.48550/ARXIV.2501.17899 Publisher: arXiv Version Number: 2. [114] Zabir Al Nazi and Wei Peng. 2024. Large Language Models in Healthcare and Medical Domain: A Review. Informatics 11, 3 (Sept. 2024), 57. doi:10.3390/informatics11030057 Publisher: Multidisciplinary Digital Publishing Institute. [115] Nataliya Nedzhvetskaya and J. S. Tan. 2024. The Role of Workers in AI Ethics and Governance. In The Oxford Handbook of AI Governance, Justin B. Bullock, Yu-Che Chen, Johannes Himmelreich, Valerie M. Hudson, Anton Korinek, Matthew M. Young, and Baobao Zhang (Eds.). Oxford University Press, 0. doi:10.1093/oxfordhb/9780197579329.013.68 [116] Anna Neumann and Jat Singh. 2025. Cascading Effects: A Multifaceted Governance Challenge in AI Systems. https://openreview.net/ forum?id=SdaBktr92I [117] Chris Norval, Heleen Janssen, Jennifer Cobbe, and Jatinder Singh. 2019. Data Protection and Tech Startups: The Need for Attention, Support, and Scrutiny. doi:10.2139/ssrn.3398204 [118] Matthew O’Shaughnessy, Daniel S Schiff, Lav R Varshney, Christopher Rozell, and Mark Davenport. 2021. What governs attitudes toward artificial intelligence adoption and governance? doi:10.31219/osf.io/pkeb8 [119] Andrei Paleyes, Raoul-Gabriel Urma, and Neil D. Lawrence. 2023. Challenges in Deploying Machine Learning: a Survey of Case Studies. Comput. Surveys 55, 6 (July 2023), 1–29. doi:10.1145/3533378 arXiv:2011.09926 [cs].

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

[120] European Parliament. 2024. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act). [121] Samir Passi and Solon Barocas. 2019. Problem Formulation and Fairness. In Proceedings of the Conference on Fairness, Accountability, and Transparency. ACM, Atlanta GA USA, 39–48. doi:10.1145/3287560.3287567 [122] Samir Passi and Steven J. Jackson. 2018. Trust in Data Science: Collaboration, Translation, and Accountability in Corporate Data Science Projects. Proc. ACM Hum.-Comput. Interact. 2, CSCW (Nov. 2018), 136:1–136:28. doi:10.1145/3274405 [123] Samir Passi and Phoebe Sengers. 2020. Making data science systems work. Big Data & Society 7, 2 (July 2020), 205395172093960. doi:10.1177/2053951720939605 [124] Yulu Pi, Lucas Lichner, Jae Woo Lee, Sijia Xiao, Renwen Zhang, and Jat Singh. 2026. Push and Pushback in Contesting AI: Demands for and Resistance to Accountability. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’26). ACM. [125] Giada Pistilli, Carlos Muñoz Ferrandis, Yacine Jernite, and Margaret Mitchell. 2023. Stronger Together: on the Articulation of Ethical Charters, Legal Tools, and Technical Documentation in ML. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’23). Association for Computing Machinery, New York, NY, USA, 343–354. doi:10.1145/3593013.3594002 [126] W. Nicholson Price, Sara Gerke, and I. Glenn Cohen. 2019. Potential Liability for Physicians Using Artificial Intelligence. JAMA 322, 18 (Nov. 2019), 1765. doi:10.1001/jama.2019.15064 [127] Jarmo Pulkkinen, Kimmo Huttu, and Marjo Suhonen. 2025. Systemic challenges in AI adoption in public social and health organizations in Finland: a technology-organisation-environment perspective. Journal of Health Organization and Management 39, 9 (Dec. 2025), 435–456. doi:10.1108/JHOM-06-2025-0309 [128] Bin Qian, Jie Su, Zhenyu Wen, Devki Nandan Jha, Yinhao Li, Yu Guan, Deepak Puthal, Philip James, Renyu Yang, Albert Y. Zomaya, Omer Rana, Lizhe Wang, Maciej Koutny, and Rajiv Ranjan. 2020. Orchestrating the Development Lifecycle of Machine Learning-based IoT Applications: A Taxonomy and Survey. ACM Comput. Surv. 53, 4 (Aug. 2020), 82:1–82:47. doi:10.1145/3398020 [129] Katyanna Quach. 2023. Stanford takes costly, risky Alpaca AI model offline. https://www.theregister.com/2023/03/21/stanford_ai_ alpaca_taken_offline/ [130] Inioluwa Deborah Raji, I. Elizabeth Kumar, Aaron Horowitz, and Andrew D. Selbst. 2022. The Fallacy of AI Functionality. In 2022 ACM Conference on Fairness Accountability and Transparency. 959–972. doi:10.1145/3531146.3533158 arXiv:2206.09511 [cs]. [131] Bogdana Rakova, Jingying Yang, Henriette Cramer, and Rumman Chowdhury. 2021. Where Responsible AI meets Reality: Practitioner Perspectives on Enablers for shifting Organizational Practices. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (April 2021), 1–23. doi:10.1145/3449081 arXiv:2006.12358 [cs]. [132] Delaram Rezaeikhonakdar. 2024. AI Chatbots and Challenges of HIPAA Compliance for AI Developers and Vendors. The Journal of Law, Medicine & Ethics 51, 4 (2024), 988–995. doi:10.1017/jme.2024.15 [133] Shalaleh Rismani, Renee Shelby, Leah Davis, Negar Rostamzadeh, and AJung Moon. 2025. Measuring What Matters: Connecting AI Ethics Evaluations to System Attributes, Hazards, and Harms. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society 8, 3 (Oct. 2025), 2199–2213. doi:10.1609/aies.v8i3.36706 [134] Everett M. Rogers. 1983. Diffusion of innovations (3. ed ed.). Free Press [u.a.], New York, NY. [135] Yuji Roh, Geon Heo, and Steven Euijong Whang. 2021. A Survey on Data Collection for Machine Learning: A Big Data - AI Integration Perspective. IEEE Transactions on Knowledge and Data Engineering 33, 4 (April 2021), 1328–1347. doi:10.1109/TKDE.2019.2946162 [136] Simon P. Rowland, J. Edward Fitzgerald, Matthew Lungren, Elizabeth (Hsieh) Lee, Zach Harned, and Alison H. McGregor. 2022. Digital health technology-specific risks for medical malpractice liability. npj Digital Medicine 5, 1 (Oct. 2022), 1–6. doi:10.1038/s41746-02200698-3 Publisher: Nature Publishing Group. [137] Daniel Russo. 2024. Navigating the Complexity of Generative AI Adoption in Software Engineering. ACM Transactions on Software Engineering and Methodology 33, 5 (June 2024), 1–50. doi:10.1145/3652154 [138] Agustina D. Saenz, Zach Harned, Oishi Banerjee, Michael D. Abràmoff, and Pranav Rajpurkar. 2023. Autonomous AI systems in the face of liability, regulations and costs. npj Digital Medicine 6, 1 (Oct. 2023), 185. doi:10.1038/s41746-023-00929-1 [139] Nithya Sambasivan, Shivani Kapania, Hannah Highfill, Diana Akrong, Praveen Paritosh, and Lora M Aroyo. 2021. “Everyone wants to do the model work, not the data work”: Data Cascades in High-Stakes AI. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM, Yokohama Japan, 1–15. doi:10.1145/3411764.3445518 [140] Sigal Samuel. 2025. Is AI being shoved down your throat at work? Here’s how to fight back. https://www.vox.com/future-perfect/ 468672/how-to-fight-generative-ai [141] Megan Sauer. 2024. Google CEO: AI development is finally slowing down—’the low-hanging fruit is gone’. https://www.cnbc.com/ 2024/12/08/google-ceo-sundar-pichai-ai-development-is-finally-slowing-down.html Section: Make It: Next Gen Investing. [142] Xinyue Shen, Zeyuan Chen, Michael Backes, Yun Shen, and Yang Zhang. 2024. "Do Anything Now": Characterizing and Evaluating In-The-Wild Jailbreak Prompts on Large Language Models. Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security (Dec. 2024), 1671–1685. doi:10.1145/3658644.3670388 Conference Name: CCS ’24: ACM SIGSAC Conference

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

on Computer and Communications Security ISBN: 9798400706363 Place: Salt Lake City UT USA Publisher: ACM. [143] David O. Shumway and Hayes J. Hartman. 2024. Medical malpractice liability in large language model artificial intelligence: legal review and policy recommendations. Journal of Osteopathic Medicine 124, 7 (June 2024), 287–290. doi:10.1515/jom-2023-0229 [144] Felix M. Simon. 2024. Escape Me If You Can: How AI Reshapes News Organisations’ Dependency on Platform Companies. Digital Journalism 12, 2 (Feb. 2024), 149–170. doi:10.1080/21670811.2023.2287464 [145] Peter Slattery, Alexander K. Saeri, Emily A. C. Grundy, Jess Graham, Michael Noetel, Risto Uuk, James Dao, Soroush Pour, Stephen Casper, and Neil Thompson. 2025. The AI Risk Repository: A Comprehensive Meta-Review, Database, and Taxonomy of Risks From Artificial Intelligence. doi:10.48550/arXiv.2408.12622 arXiv:2408.12622 [cs]. [146] Klaas-Jan Stol, Paris Avgeriou, Muhammad Ali Babar, Yan Lucas, and Brian Fitzgerald. 2014. Key factors for adopting inner source. ACM Trans. Softw. Eng. Methodol. 23, 2 (April 2014), 18:1–18:35. doi:10.1145/2533685 [147] Stop Gen AI. 2025. Stop Gen AI – Mutual Aid and Political Activism. https://stopgenai.com/ [148] Nan Sun, Yuantian Miao, Hao Jiang, Ming Ding, and Jun Zhang. 2025. From Principles to Practice: A Deep Dive into AI Ethics and Regulations. doi:10.48550/arXiv.2412.04683 arXiv:2412.04683 [cs]. [149] Harini Suresh and John V. Guttag. 2021. A Framework for Understanding Sources of Harm throughout the Machine Learning Life Cycle. In Equity and Access in Algorithms, Mechanisms, and Optimization. 1–9. doi:10.1145/3465416.3483305 arXiv:1901.10002 [cs]. [150] Aneeta Sylolypavan, Derek Sleeman, Honghan Wu, and Malcolm Sim. 2023. The impact of inconsistent human annotations on AI driven clinical decision making. npj Digital Medicine 6, 1 (Feb. 2023), 1–13. doi:10.1038/s41746-023-00773-3 Publisher: Nature Publishing Group. [151] Elham Tabassi. 2023. Artificial Intelligence Risk Management Framework (AI RMF 1.0). Technical Report NIST AI 100-1. National Institute of Standards and Technology (U.S.), Gaithersburg, MD. NIST AI 100–1 pages. doi:10.6028/NIST.AI.100-1 [152] Eran Tal. 2023. Target specification bias, counterfactual prediction, and algorithmic fairness in healthcare. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society. ACM, Montr\’{e}al QC Canada, 312–321. doi:10.1145/3600211.3604678 [153] Erdal Tasci, Ying Zhuge, Kevin Camphausen, and Andra V. Krauze. 2022. Bias and Class Imbalance in Oncologic Data—Towards Inclusive and Transferrable AI in Large Scale Oncology Data Sets. Cancers 14, 12 (June 2022), 2897. doi:10.3390/cancers14122897 [154] Claudio Terranova, Clara Cestonaro, Ludovico Fava, and Alessandro Cinquetti. 2024. AI and professional liability assessment in healthcare. A revolution in legal medicine? Frontiers in Medicine 10 (Jan. 2024), 1337335. doi:10.3389/fmed.2023.1337335 [155] Tobias Mann. 2025. Most AI spending driven by FOMO, not ROI, CEOs tell IBM. https://www.theregister.com/2025/05/06/ibm_ai_ investments/ [156] Andrius Vabalas, Emma Gowen, Ellen Poliakoff, and Alexander J. Casson. 2019. Machine learning algorithm validation with a limited sample size. PLOS ONE 14, 11 (Nov. 2019), e0224365. doi:10.1371/journal.pone.0224365 [157] Martim Veiga and Carlos J. Costa. 2024. Ethics and Artificial Intelligence Adoption. (2024). doi:10.48550/ARXIV.2412.00330 Publisher: arXiv Version Number: 1. [158] Viswanath Venkatesh and Hillol Bala. 2008. Technology Acceptance Model 3 and a Research Agenda on Interventions. Decision Sciences 39, 2 (2008), 273–315. doi:10.1111/j.1540-5915.2008.00192.x _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.15405915.2008.00192.x. [159] Zhiyuan Wan, Xin Xia, David Lo, and Gail C. Murphy. 2020. How does Machine Learning Change Software Development Practices? IEEE Transactions on Software Engineering (2020), 1–1. doi:10.1109/TSE.2019.2937083 [160] Angelina Wang, Sayash Kapoor, Solon Barocas, and Arvind Narayanan. 2024. Against Predictive Optimization: On the Legitimacy of Decision-making Algorithms That Optimize Predictive Accuracy. ACM Journal on Responsible Computing 1, 1 (March 2024), 1–45. doi:10.1145/3636509 [161] Zhiyuan Wang, Runze Yan, Sherilyn Francis, Carmen Diaz, Tabor Flickinger, Yufen Lin, Xiao Hu, Laura E. Barnes, and Virginia LeBaron. 2025. Stakeholder-centric participation in large language models enhanced health systems. npj Health Systems 2, 1 (June 2025), 22. doi:10.1038/s44401-025-00024-5 Publisher: Nature Publishing Group. [162] Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, Po-Sen Huang, John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa Kasirzadeh, Courtney Biles, Sasha Brown, Zac Kenton, Will Hawkins, Tom Stepleton, Abeba Birhane, Lisa Anne Hendricks, Laura Rimell, William Isaac, Julia Haas, Sean Legassick, Geoffrey Irving, and Iason Gabriel. 2022. Taxonomy of Risks posed by Language Models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22). Association for Computing Machinery, New York, NY, USA, 214–229. doi:10.1145/3531146.3533088 [163] Karl Werder, Balasubramaniam Ramesh, and Rongen (Sophia) Zhang. 2022. Establishing Data Provenance for Responsible Artificial Intelligence Systems. ACM Transactions on Management Information Systems 13, 2 (June 2022), 1–23. doi:10.1145/3503488 [164] Seungjin Whang. 1992. Contracting for Software Development. Management Science 38, 3 (March 1992), 307–324. doi:10.1287/mnsc.38. 3.307 Publisher: INFORMS. [165] David Gray Widder and Dawn Nafus. 2023. Dislocated accountabilities in the “AI supply chain”: Modularity and developers’ notions of responsibility. Big Data & Society 10, 1 (Jan. 2023), 20539517231177620. doi:10.1177/20539517231177620

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

Chappidi & Singh

[166] Fons Wijnhoven. 2022. Organizational Learning for Intelligence Amplification Adoption: Lessons from a Clinical Decision Support System Adoption Project. Information Systems Frontiers 24, 3 (June 2022), 731–744. doi:10.1007/s10796-021-10206-9 [167] Kevin Witzenberger and Michael Richardson. 2025. Microsoft cuts data centre plans and hikes prices in push to make users carry AI costs. doi:10.64628/AA.96dhycghd [168] Claes Wohlin, Marcos Kalinowski, Katia Romero Felizardo, and Emilia Mendes. 2022. Successful combination of database search and snowballing for identification of primary studies in systematic literature studies. Information and Software Technology 147 (July 2022), 106908. doi:10.1016/j.infsof.2022.106908 [169] Richmond Y. Wong, Michael A. Madaio, and Nick Merrill. 2023. Seeing Like a Toolkit: How Toolkits Envision the Work of AI Ethics. Proceedings of the ACM on Human-Computer Interaction 7, CSCW1 (April 2023), 1–27. doi:10.1145/3579621 [170] Carole-Jean Wu, R. Raghavendra, Udit Gupta, Bilge Acun, Newsha Ardalani, Kiwan Maeng, Gloria Chang, Fiona Aga Behram, James Huang, Charles Bai, M. Gschwind, Anurag Gupta, Myle Ott, Anastasia Melnikov, Salvatore Candido, David Brooks, Geeta Chauhan, Benjamin Lee, Hsien-Hsin S. Lee, Bugra Akyildiz, Maximilian Balandat, Joe Spisak, R. Jain, M. Rabbat, and K. Hazelwood. 2021. Sustainable AI: Environmental Implications, Challenges and Opportunities. ArXiv (Oct. 2021). https://www.semanticscholar.org/ paper/2c6df83795cd5baf3b8c6e2639b85e2df0cee1d0 [171] Zhiyuan Yu, Xiaogeng Liu, Shunning Liang, Zach Cameron, Chaowei Xiao, and Ning Zhang. 2024. Don’t Listen To Me: Understanding and Exploring Jailbreak Prompts of Large Language Models. (2024). doi:10.48550/ARXIV.2403.17336 Publisher: arXiv Version Number: 2. [172] Hubert Dariusz Zając, Natalia Rozalia Avlona, Finn Kensing, Tariq Osman Andersen, and Irina Shklovski. 2023. Ground Truth Or Dare: Factors Affecting The Creation Of Medical Datasets For Training AI. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society. ACM, Montr\’{e}al QC Canada, 351–362. doi:10.1145/3600211.3604766 [173] Dawen Zhang, Boming Xia, Yue Liu, Xiwei Xu, Thong Hoang, Zhenchang Xing, Mark Staples, Qinghua Lu, and Liming Zhu. 2024. Privacy and Copyright Protection in Generative AI: A Lifecycle Perspective. In Proceedings of the IEEE/ACM 3rd International Conference on AI Engineering - Software Engineering for AI (CAIN ’24). Association for Computing Machinery, New York, NY, USA, 92–97. doi:10.1145/3644815.3644952

Factors that Lead to Non-Development or Abandonment of AI Systems

FAccT ’26, June 25–28, 2026, Montreal, QC, Canada

A Abandoned AI Cases Reported Via Survey A.1 Participant Demographics Participants represented 7 countries globally, with the majority from English-speaking nations. Responses were collected from the United States (n=14), the United Kingdom of Great Britain and Northern Ireland (n=6), India (n=3), Netherlands (3), Germany (n=3), and Spain (n=1). Participants reported an average of 12 years of professional experience (min=1 year, Q1=5, median=6, Q3=20, max=30 years). The surveyed cases reported operations across a wide range of domains, including technology (n=12), finance (9), research (9), government (6), professional services (n=5), health (3), energy (2), civil society/NGOs (2), military/defense (1), and legal (1). Participants also operated their algorithms in diverse industries, including tech companies (n=14), academia (n=8 with 6 related to AI research, 2 related to other fields), government (6), consulting firms (6), non-profits (n=6, 3 related to research and advocacy and 3 service-based), start-ups (3), and non-tech private/commercial companies (n=3).

A.2

Factor-Level Details Contributing to Abandoned AI Cases

System Description

ID

Stage at which Abandoned

Ethical Concerns

Organizational Dynamics

ML Development Challenges

Chatbot for public assistance program applications

S1

Deployment

Misinformation, Concerns over overreliance

Changing incentives / priorities

Undefined success or evaluation criteria, Low adoption observed during pilot deployment

Development timeline deemed too long

AI-mediated ecommerce integration with web search

S2

Model retraining

Not indicated / scoped well, Not aligned with organizational strategy, Insufficient leadership sponsorship

Model training was too technically challenging, Model performance not sufficient, Inability to conduct a proper evaluation or pilot

Lacked sufficient technical expertise to build, Development timeline deemed too long

Clustering to identify fraudulent crypto users

S3

Model retraining

Clustering to identify high-risk labor operations

S4

Problem Concerns over loss of operationalization

Inference to improve data quality gaps

S5

Data collection

Not indicated / scoped well, Insufficient leadership sponsorship

In-house LLM development (to compete w/ industry)

S6

Model evaluation

Changing incentives / priorities, Not indicated / scoped well, Clients / customers did not want it

LLM-based infrastructure support for SWEs

S7

Problem formulation

Concept extraction from larger NLP model

S8

Problem operationalization

LLM-based notification summaries for SWEs

S9

Problem operationalization

S10

Data collection

S11

Model evaluation

S12

Model evaluation

RAG system for sustainability business contacts Decision trees for COVID-19 travel warnings LLMs for functional form identification

Stakeholder Feedback

Resource Constraints

Legal / Regulatory Challenges

Lacked (or too challenging to obtain) ground truth, Too challenging to label data Discrimination, Privacy & security,

Could not measure target variable, Lacked (or too challenging to obtain) ground truth

human agency & autonomy

General concerns about AI (not tied to specific risks)

Changing incentives / priorities, Not indicated / scoped well, Not aligned with organizational strategy, Insufficient leadership sponsorship, Concerns over keeping internal assets private

Feedback to not build (via stakeholder engagement)

Challenging to collect desired data, Lacked/challenging to obtain ground truth, Too challenging to label data Sample size too small, Model training was too resource intensive, Model performance not sufficient, Difficult to integrate into existing pipelines, Low adoption during pilot deployment Challenging to collect desired data, Difficult to integrate into existing pipelines, Inability to conduct a proper evaluation or pilot,

Too expensive to build, Concerns over availability of compute resources, Easier to outsource development or procure existing tool Feedback to not build (via stakeholder engagement)

Undefined success or evaluation criteria

Feedback to not build (via stakeholder engagement)

Too challenging to pre-process/curate data, Undefined success or evaluation criteria, Low adoption during pilot deployment

Easier to outsource development or procure existing tool Compliance concerns with data protection regulations (e.g., GDPR)

Challenging to collect desired data Changing incentives / priorities, Insufficient leadership sponsorship, Clients / customers did not want it

Concerns over data protection regulations compliance, Concerns over domainspecific regulations compliance, Lack of regulatory guidance

Cheaper to outsource development or procure existing tool, Easier to outsource development or procure existing tool

Inappropriate model selection Changing incentives / priorities, Insufficient leadership sponsorship

Development timeline deemed too long, Too many solutions/use cases with disparate solutions

Feedback to not build (via user resistance)

Feedback to not build Lacked (or too challenging to obtain) ground (via prospect of academic truth, Undefined success or evaluation criteria community resistance)

Fig. A1. Factor-level details reported as contributing to AI abandonment decisions collected via practitioner survey.