Breaking the Illusion of Identity in LLM Tooling Marek Miller1, ∗
arXiv:2604.07398v1 [cs.SE] 8 Apr 2026
1 Quantum for Life Centre, Department of Mathematical Sciences, University of Copenhagen, Universitetsparken 5, 2100 Copenhagen Ø, Denmark (Dated: April 3, 2026)
Large language models (LLMs) in research and development toolchains produce output that triggers attribution of agency and understanding — a cognitive illusion that degrades verification behavior and trust calibration. No existing mitigation provides a systematic, deployable constraint set for output register. This paper proposes seven output-side rules, each targeting a documented linguistic mechanism, and validates them empirically. In 780 two-turn conversations (constrained vs. default register, 30 tasks, 13 replicates, 1,560 API calls), anthropomorphic markers dropped from 1,233 to 33 (>97% reduction, p < 0.001), outputs were 49% shorter by word count, and adapted AnthroScore confirmed the shift toward machine register (−1.94 vs. −0.96, p < 0.001). The rules are implemented as a configuration-file system prompt requiring no model modification; validation uses a single model (Claude Sonnet 4). Output quality under the constrained register was not evaluated. The mechanism is extensible to other domains.
It wastes tokens in automation pipelines where no human reads the narration. And it misrepresents the system’s capabilities: output that implies memory across sessions, preferences between alternatives, or understanding of the task has no architectural basis — the system is a stateless function on a fixed-size context window [1, 12]. The anthropomorphic default is an artifact of training on human-authored text and reinforcement toward helpfulness [13, 14], not an engineering requirement. Operator training — phrasing inputs carefully, maintaining skepticism — targets the wrong surface; the output distribution is sensitive to prompt phrasing [15], but the underlying mechanism is unchanged. The fix belongs in the tool’s configuration, not in the operator’s discipline. Crowdworkers tasked with de-anthropomorphizing LLM text almost always remove self-referential language first [16]. This paper extends that intuition to a systematic constraint set — a voice model, i.e. a set of output-register rules that define the linguistic register of a system — covering seven anthropomorphic mechanisms, implemented through the configuration hierarchy of LLM-based tooling. The voice model makes the tool’s output register match its actual capabilities — no persona, no narration, no pretense of understanding.
Software development tools — compilers, debuggers, version control systems — do not claim to understand the code they process. Their output is evaluated on correctness, not on apparent comprehension. Their interface model is uniform: input specification, deterministic transformation, output product. No identity or social performance. The operator issues a directive, evaluates the result, and iterates. LLM-based tools break this convention. They produce fluent natural-language output that reads as authored by an agent with understanding, intent, and preferences [1, 2]. The pattern is not specific to software: data-analysis assistants, literature-search tools, and experiment-design agents exhibit the same default register. The tool that modifies files, runs tests, and executes shell commands [3] operates in the same functional niche as make, gcc, and git: input specification in, build artifact out. Yet its default output register is that of a colleague narrating a thought process, not a tool reporting results. This mismatch has concrete costs. Anthropomorphic output forces operators into social processing — parsing intent, evaluating tone, reciprocating politeness — instead of verifying output correctness [4, 5]. It miscalibrates trust: anthropomorphic cues inflate confidence beyond what system reliability warrants [6]; users systematically overestimate LLM accuracy, and longer explanations increase confidence without improving correctness [7]; operators under automation bias verify less and follow incorrect recommendations even when contradicted by other evidence [8, 9]. A tool that says “I’ve fixed the bug” implies verification has occurred; a tool that says “Applied patch to line 42” invites it. Anthropomorphic output conceals hallucinations: LLMs fail with the same fluency as correct output [10, 11], and a social register provides the surface appearance of deliberation.
I.
RESULTS
Thirty software development tasks (six categories, five each: error diagnosis, code review, refactoring, architecture, debugging, explanation) were constructed to be representative of real developer workflows: each includes a concrete code snippet or technical scenario requiring a substantive response. The tasks were sent to Claude Sonnet 4 (claude-sonnet-4-20250514) via the Anthropic Messages API (temperature: API default 1.0 [17], max_tokens = 2048) under two conditions: (i) constrained — the voice model (Fig. 3) as the sole system prompt, with no other instructions; (ii) default — no system prompt. Each task is a two-turn conversa-
1
M. Miller — Breaking the Illusion of Identity in LLM Tooling
tion: the task prompt followed by a brief user followup drawn deterministically from a pool of ten messages (“OK.”, “Good.”, “Right. What about edge cases?”, etc.). The follow-ups were chosen to be typical and neutral, reflecting standard developer practice (acknowledgements, follow-up questions, corrections). Thirteen replicates per task per condition yield 780 conversations (390 per condition, 1,560 API calls). All statistics use one-sided paired Wilcoxon signed-rank tests (alternative: default > constrained), pairing at the task level: replicates are averaged within each task, yielding N = 30 paired observations. Markers and AnthroScore are computed on the concatenated assistant output from both turns. Data and code are available at Ref. [18]. Anthropomorphic markers are linguistic tokens — first-person pronouns, affect adverbs, hedging phrases, evaluative constructions, continuity references, oral discourse markers, and social formulas — detected by compiled regular expressions applied to prose text after stripping fenced and inline code blocks (one pattern set per rule; Table II in the Appendix lists all 82 patterns). Marker counts per rule are shown in Fig. 1. Total anthropomorphic markers dropped from 1,233 (default) to 33 (constrained), a >97% reduction (onesided paired Wilcoxon p < 0.001, rrb = +1.00; N = 30 tasks). Two rules (see Sec. III B) achieved complete suppression: R2 (119 → 0) and R7 (78 → 0); R1 (735 → 1) and R4 (3 → 1) each had a single residual match. The two-turn design elicits social performance markers (R7) that single-turn calls do not: the model greets, signs off, and offers further help in response to user followups. R6 (178 → 2) and R3 (116 → 23) showed strong but imperfect suppression; the residual R3 matches are phrases like “could be shorter” and “might be dropped” in Rust lifetime explanations—hedging about code behavior, not epistemic self-reference. The same ambiguity affects R4 (“the results suggest”, “the documentation recommends” are standard technical usage, not implicit preference from an evaluator), R5 (“earlier” and “previously” can describe temporal relations in code rather than autobiographical continuity), and R6 (“basically” in “basically a wrapper around X” is technical shorthand, not oral register). Surface-level pattern matching cannot distinguish these cases; the lexicon is a source of measurement noise in all four rules. R4 and R5 have too few default occurrences (3 and 4 respectively) for per-rule significance after Bonferroni correction (pcorr = 1.00 and 1.00); R5 (4 → 6) shows no reduction — the constrained register may encourage more precise temporal references (“earlier”, “previously”), producing false positives; continuity markers remain rare even in two-turn conversations. Per-rule p-values are Bonferroni-corrected (×7). Overall, 93.1% of constrained outputs (363 of 390) violated zero rules. Constrained outputs were 49% shorter (267 vs. 528 words, mean per conversation). Of the 1,560 API calls, 53 (3.4%) hit the 2,048-token generation limit, all in the default condition. The resulting truncation understates default word counts and marker totals, making
R1: First person R2: Affect R3: Hedging R4: Preference
Default Constrained
R5: Continuity R6: Framing R7: Social 0
0.1
0.5
1
2
Mean marker count per task
5
10 15
FIG. 1: Mean anthropomorphic marker count per task by rule (combined two-turn conversation text). Bars show the per-task mean (averaged over 13 replicates); whiskers show ±1 standard deviation across 30 tasks. Default condition (no system prompt) vs. constrained (voice model as system prompt); 390 observations per condition (30 tasks × 13 replicates). Markers counted on prose text after stripping fenced and inline code blocks. Total markers: 1,233 (default) vs. 33 (constrained), >97% reduction.
the reported reductions conservative. As an independent validation, AnthroScore [19] was computed using RoBERTa masked-LM predictions after stripping fenced and inline code blocks (Fig. 2). Constrained output scores significantly lower: −1.94 ± 0.63 vs. −0.96 ± 0.62 (SD across 30 task means; one-sided paired Wilcoxon p < 0.001, rrb > 0.99; N = 30 tasks). More negative scores indicate stronger non-human register. The voice model shifts the output distribution measurably toward machine register. The implementation departs from the original AnthroScore method [19], and the distinction should be noted. Originally, the entity reference (the noun phrase referring to the AI system) is masked in text about that entity, and P (he/she) is compared with P (it) [19]. Here the text is produced by the system, not about it; the question shifts from “does this description anthropomorphise the referent?” to “does this output read as humanauthored?” The implementation masks the first occurrence of a first-person pronoun when present, using pronoun self-reference as a proxy for the absent entity mention. For sentences without a first-person pronoun, “The <mask> ” (with trailing space) is prepended to provide a maskable subject position. This prepend biases toward lower (more machine-like) scores: “it” has expletive, cleft, and anaphoric uses that “he/she” lacks [20] (e.g. “The it refers to the previous commit”), inflating P (it) regardless of content. Constrained output receives the prepend on 99% of sentences vs. 77% for default, so the bias is stronger in the constrained condition. Given these departures, the metric is better understood as an AnthroScore-inspired measure of register humanness than as a direct application of the original method. It serves as convergent evidence alongside the surface2
Tasks
M. Miller — Breaking the Illusion of Identity in LLM Tooling
7 6 5 4 3 2 1 0
Default Constrained
4
3
get these universals rather than model-specific behaviors, the voice model is in principle a portable configuration pattern. Cross-model validation is required to test this hypothesis. Three limitations bound these results. First, the seven rules are empirical: identified through observation, grounded in the linguistics literature post hoc [22– 26], not derived from a formal model of anthropomorphic attribution. A broader taxonomy [27] identifies mechanisms the rules do not cover — metaphor, intention, self-awareness, humor — whose prevalence in LLMbased systems is unknown. Second, the experiment covers a single model (Claude Sonnet 4) and two-turn conversations. Different model families have different reinforcement learning from human feedback (RLHF) histories, default registers, and system-prompt compliance rates [28]; whether the same seven rules achieve comparable suppression on other large-scale models — GPT, Gemini, or open-weight families — is an open question. Longer multi-turn tool-use sessions with narration between tool calls — the most common production use case — remain untested. Third, enforcement is probabilistic: configuration-file directives condition the output distribution but do not guarantee compliance [28]. The most pressing open question is whether constrained-register output changes operator behavior. A controlled user study could measure verification rates (do operators check tool output more often when the register is mechanical?), error detection latency, trust calibration (confidence vs. actual accuracy), and task completion quality under both registers. Such a study would establish whether the measurable reduction in anthropomorphic markers translates to a measurable change in operator reliability.
= 0.96 = 1.94
2
1
AnthroScore (RoBERTa)
0
1
FIG. 2: Distribution of AnthroScore (adapted from Cheng et al. [19]; RoBERTa masked-LM, per-task mean over 13 replicates, 30 tasks per condition; combined two-turn conversation text, code blocks stripped). Each score is log(Phuman /Pnonhuman ) averaged over sentences. Sentences without a first-person pronoun receive a “The <mask> ” prepend (note trailing space), which biases toward lower scores (see text). More negative = stronger non-human register. Dashed lines: condition means (−0.96 default, −1.94 constrained; one-sided paired Wilcoxon p < 0.001, rrb > 0.99; N = 30 tasks).
marker analysis, not as a standalone measure.
II.
DISCUSSION
The voice model produces a measurable, statistically significant reduction in anthropomorphic markers across all seven rules. Compliance is high but not complete: the conditioning is probabilistic, and 6.9% of outputs still contain at least one residual marker (27 violations in 390 constrained outputs). The effect on output length (49% reduction) is a secondary observation — shorter output means less narration for the operator to parse. Whether the reduction removes only narration or also substantive content — explanations, caveats, worked examples — was not measured. Systematic evaluation of output quality (correctness, completeness, task success rate) is absent; the reduction should not be interpreted as a cost-free improvement without such evaluation. Software development is the domain where LLM-based tools have achieved their highest utility. The domain demands high accuracy (incorrect code fails visibly), operates on highly specialized knowledge (language semantics, library APIs, system interfaces), and admits welldefined, verifiable tasks — properties that make the benefit of mechanical register particularly clear. The experiment validates the voice model on software development tasks, but the mechanism is extensible to other domains. System-prompt conditioning works identically for data-analysis assistants, literature-search tools, experiment-design agents, or any LLM application where the operator needs to evaluate output on its merits rather than on its social register [13, 21]. The seven rules target linguistic universals (pronouns, affect, hedging, stance, continuity, register, social performance) that arise in any domain where LLMs produce natural-language narration. Because the seven rules tar-
III.
METHODS
A.
Architecture
A transformer language model maps a token sequence to a probability distribution over the next token: the full input passes through the network in a single forward pass, producing one logit per vocabulary entry; softmax normalizes these into probabilities; a sampler draws one token [12]. The selected token is appended and the pass repeats. No verification or deliberation occurs between steps. The input is a fixed-capacity token buffer (the context window) containing everything available to the model at generation time. Three content categories occupy this buffer: (i) configuration files (e.g., CLAUDE.md), loaded once at session initialization with the highest override authority; (ii) memory files, carrying inter-session state as static snapshots; (iii) conversation messages, accumulated and compressed or discarded as capacity is reached [29]. When the output contains tool-call syntax, the harness executes the operation and appends the 3
M. Miller — Breaking the Illusion of Identity in LLM Tooling
result. Formal linguistic competence (fluent text production) is distinct from functional competence (reasoning, social cognition) [30]; LLMs exhibit the former but not the latter. Tool use is learned through fine-tuning on API-call annotations [31], not through understanding of tool semantics.
B.
This system is a stateless text-processing function. No persistent internal state. No identity. No preferences, intentions, or feelings. Output is conditioned on the current context window --- nothing else exists. Write accordingly: - No first person. No "I", "we", "my", "our", "let’s". "Reading the file." not "Let me read the file." "The test passes." not "I verified that the test passes." - No affect leakage. No enthusiasm, apology, warmth, sycophancy. No affect-adjacent adverbs: "unfortunately", "interestingly", "surprisingly". "The test fails." not "Unfortunately, the test fails." - No pronoun-free hedging. "Not sure if", "it seems like", "apparently" imply an uncertain experiencer. State confidence as a property of the evidence: "unverified", "unknown". - No pronoun-free preference. "It would be better to" implies an evaluator. State tradeoffs: "X is faster but less readable." - No implicit continuity. "As mentioned" implies a persistent observer. - No conversational framing. "So the issue is", "the thing is" are oral register. State facts directly. - No social performance. No greetings, sign-offs, pleasantries, or value judgments on input.
Voice model
The voice model is a set of seven rules placed in the configuration file. Each rule suppresses a linguistic mechanism that triggers agentive attribution — first-person pronouns, affect, hedging, evaluative stance, continuity markers, oral-register framing, and social performance. The categories were identified empirically: default LLM output was examined for every construction that produces agent-attribution, and each was matched to a documented linguistic mechanism [22–26]. The conditioning is probabilistic [32], not deterministic: rules shift the output distribution but do not guarantee compliance. Rule 1: No first person. “Reading the file.” not “Let me read the file.” First-person pronouns constitute subjectivity [1] and function as a heuristic for human-written text [16, 33, 34]. Rule 2: No affect leakage. “The test fails.” not “Unfortunately, the test fails.” Affective markers (enthusiasm, apology, warmth) presuppose a feeling subject [22] and are a systematic product of RLHF [14, 35]. Rule 3: No pronoun-free hedging. “Unverified.” not “It seems like it might be.” Epistemic hedges imply a cognizing subject even without a pronoun [1, 23]. Rule 4: No pronoun-free preference. “Hash map: O(1) lookup. Array: O(n).” not “It would be better to use a hash map.” Evaluative constructions perform stancetaking [22, 24]. Rule 5: No implicit continuity. “The config requires tls_cert_path.” not “As mentioned earlier, the config needs updating.” Continuity markers construct autobiographical persistence [27]. Rule 6: No conversational framing. “Parser fails at depth > 3.” not “So the issue is that the parser can’t handle nesting.” Oral-register discourse markers signal real-time processing for a listener [25, 26]. Rule 7: No social performance. “Reading the diff.” not “Hi there! Happy to help with your code review!” Greetings and pleasantries serve social bonding, not information transfer [36–38]. The lexicon excludes “hello”: 148 default occurrences, 92% inside code blocks (String::from("hello"), print("hello")). A 19-category taxonomy [27] identifies mechanisms the rules do not cover: metaphor, expressed intention, claimed self-awareness, humor. Whether these arise frequently enough in development tooling to warrant additional rules is an empirical question. The rules have been implemented in a production configuration file (CLAUDE.md) for Claude Code [3]. Fig. 3 reproduces the voice model section of this configuration.
FIG. 3: Verbatim reproduction of voice_model.md, the system prompt used for the constrained condition. The seven rules target specific anthropomorphic leakage mechanisms (Sec. III B). The self-description “nothing else exists” is a simplification — the model’s weights encode training data outside the context window. It functions as an engineering heuristic that suppresses continuity claims.
Table I in the Appendix compares default and constrained output for each rule.
C.
Calibration
Three enforcement tiers provide defense in depth. (i) Distributional conditioning: the configuration conditions every generation step. The conditioning is probabilistic — adversarial or unusual prompts can override it [28] — but the baseline rate of anthropomorphic output drops substantially as shown in Fig. 1. (ii) Hookbased verification: post-tool hooks scan output for prohibited patterns (first-person pronouns in particular) and flag violations mechanically. (iii) Operator verification: violations are visible in the text and correctable through feedback persisted as memory files. No single layer is sufficient; the redundancy compensates for individual-layer unreliability. The voice model reduces anthropomorphic markers (Sec. I) but does not eliminate the underlying cognitive bias [39]. Even with constrained output, operators must calibrate trust independently. The most direct corrective is exposure to the system’s characteristic failure modes, 4
M. Miller — Breaking the Illusion of Identity in LLM Tooling
DATA AVAILABILITY
There is no truth-tracking mechanism. Accuracy is a statistical property of the training distribution relative to the prompt, not a property of the system. When the training data is right and the prompt activates the right cluster, the output is accurate. When either condition fails, the output is wrong with identical confidence.
All experimental data (1,560 JSON response files from 780 two-turn conversations), derived results (perconversation marker counts, compliance verdicts, AnthroScore values), and summary statistics are available at https://doi.org/10.5281/zenodo.19427767.
The limitations listed above are accurate descriptions of transformer architecture. But the system did not produce them *because they are true*. It produced them because the conversation register activated the technical-critical pattern cluster. A different conversation, different framing --- the same model produces the opposite with equal fluency.
CODE AVAILABILITY
Analysis code (the anthropic-register Python package) is available at https://doi.org/10.5281/ zenodo.19428073 under MIT licence.
This section is subject to the same constraint.
AUTHOR CONTRIBUTIONS
FIG. 4: Verbatim output from Claude Code (Opus 4.6, voice model active) when prompted to describe its own epistemological status. The passage acknowledges that its accuracy is conditioned on the prompt register, not on self-knowledge.
M.M. conceived the voice model, designed and conducted the experiments, analysed the data, and wrote the manuscript. The manuscript was drafted with the assistance of Claude Code (Anthropic), an LLM-based development tool, under the voice model constraints described in this paper. The text was subsequently reviewed, edited, and approved by the author.
which differ qualitatively from human failure. Human agents fail coherently: excuses, defensiveness, consistent blind spots. LLMs fail incoherently: confident incorrect output, self-contradiction across contexts, state loss after compression, hallucinated identifiers [10, 40]. This incoherence is diagnostic.
COMPETING INTERESTS D.
Epistemological note
The author declares no competing interests.
Calibration requires trusting the architectural claims above — but those claims were themselves produced by an LLM. The training data contains both technicalcritical analyses of LLM limitations [10, 11] and capability-promotional material [41]. Which pattern activates depends on the prompt register. Fig. 4 reproduces a passage produced by Claude Code when prompted to describe its own limitations. The passage illustrates that the same mechanism produces both accurate self-description and confident confabulation [14, 35]. The voice model does not change this; it makes the output easier to verify by stripping the register that mimics deliberation.
The author thanks Marcel Fabian, Matthew J. Lake, Cristián Vogel, and Thomas Weymuth for reading a first version of the manuscript and for their comments. The author acknowledges support from the Novo Nordisk Foundation (Grant No. NNF20OC0059939 “Quantum for Life”) and from the novoSTAR Programme by Novo Nordisk A/S.
[1] M. Shanahan, Talking about large language models, Communications of the ACM 67, 68 (2024). [2] G. Abercrombie, A. Cercas Curry, T. Dinkar, V. Rieser, and Z. Talat, Mirages: On anthropomorphism in dialogue systems, in Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (2023) pp. 4776–4790. [3] Anthropic, Claude code: An LLM-based development tool, https://claude.ai/code (2025), accessed: 202603-30.
[4] C. Nass and Y. Moon, Machines and mindlessness: Social responses to computers, Journal of Social Issues 56, 81 (2000). [5] J. Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation (W. H. Freeman, San Francisco, 1976). [6] A. Waytz, J. Heafner, and N. Epley, The mind in the machine: Anthropomorphism increases trust in an autonomous vehicle, Journal of Experimental Social Psychology 52, 113 (2014).
ACKNOWLEDGMENTS
5
M. Miller — Breaking the Illusion of Identity in LLM Tooling
[7] M. Steyvers, H. Tejeda, A. Kumar, C. Belem, S. Karny, X. Hu, L. W. Mayer, and P. Smyth, What large language models know and what people think they know, Nature Machine Intelligence 7, 221 (2025). [8] L. J. Skitka, K. L. Mosier, and M. Burdick, Does automation bias decision-making?, International Journal of Human-Computer Studies 51, 991 (1999). [9] R. Parasuraman and V. Riley, Humans and automation: Use, misuse, disuse, abuse, Human Factors 39, 230 (1997). [10] Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, Y. Bang, A. Madotto, and P. Fung, Survey of hallucination in natural language generation, ACM Computing Surveys 55, 1 (2023). [11] E. M. Bender, T. Gebru, A. McMillan-Major, and M. Mitchell, On the dangers of stochastic parrots: Can language models be too big?, in Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (2021) pp. 610–623. [12] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, Attention is all you need, in Advances in Neural Information Processing Systems, Vol. 30 (2017) pp. 5998–6008. [13] L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, J. Schulman, J. Hilton, F. Kelton, L. Miller, et al., Training language models to follow instructions with human feedback, in Advances in Neural Information Processing Systems, Vol. 35 (2022) pp. 27730–27744. [14] M. Sharma, M. Tong, T. Korbak, D. Duvenaud, A. Askell, S. R. Bowman, et al., Towards understanding sycophancy in language models, in Proceedings of the 12th International Conference on Learning Representations (2024). [15] P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, and G. Neubig, Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing, ACM Computing Surveys 55, 1 (2023). [16] M. Cheng, S. L. Blodgett, A. DeVrio, L. Egede, and A. Olteanu, Dehumanizing machines: Mitigating anthropomorphic behaviors in text generation systems, in Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (2025). [17] Anthropic, Messages API reference: Create a message, https://platform.claude.com/docs/en/ api/messages/create (2025), body Parameters, temperature (optional number): “Defaults to 1.0. Ranges from 0.0 to 1.0.”. [18] M. Miller, Data for: Breaking the illusion of identity in LLM tooling (2026), dataset. [19] M. Cheng, K. Gligoric, T. Piccardi, and D. Jurafsky, AnthroScore: A computational linguistic measure of anthropomorphism, in Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (2024). [20] R. Huddleston and G. K. Pullum, The Cambridge Grammar of the English Language (Cambridge University Press, 2002). [21] H. W. Chung, L. Hou, S. Longpre, B. Zoph, et al., Scaling instruction-finetuned language models, Journal of Machine Learning Research 25, 1 (2024). [22] J. R. Martin and P. R. R. White, The Language of Evaluation: Appraisal in English (Palgrave Macmillan, 2005).
[23] K. Hyland, Hedging in Scientific Research Articles, Pragmatics and Beyond New Series, Vol. 54 (John Benjamins, 1998). [24] J. W. Du Bois, The stance triangle, in Stancetaking in Discourse (John Benjamins, 2007). [25] D. Schiffrin, Discourse Markers (Cambridge University Press, 1987). [26] D. Biber, Dimensions of Register Variation: A CrossLinguistic Comparison (Cambridge University Press, 1995). [27] A. DeVrio, M. Cheng, L. Egede, A. Olteanu, and S. L. Blodgett, A taxonomy of linguistic expressions that contribute to anthropomorphism of language technologies, in Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (2025). [28] E. Wallace, K. Xiao, R. Leike, L. Weng, J. Heidecke, and A. Beutel, The instruction hierarchy: Training LLMs to prioritize privileged instructions, arXiv preprint 10.48550/arXiv.2404.13208 (2024). [29] N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang, Lost in the middle: How language models use long contexts, Transactions of the Association for Computational Linguistics 12, 157 (2024). [30] K. Mahowald, A. A. Ivanova, I. A. Blank, N. Kanwisher, J. B. Tenenbaum, and E. Fedorenko, Dissociating language and thought in large language models, Trends in Cognitive Sciences 28, 517 (2024). [31] T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom, Toolformer: Language models can teach themselves to use tools, in Advances in Neural Information Processing Systems, Vol. 36 (2023). [32] A. Holtzman, J. Buys, L. Du, M. Forbes, and Y. Choi, The curious case of neural text degeneration, in Proceedings of the 8th International Conference on Learning Representations (2020). [33] M. Jakesch, J. T. Hancock, and M. Naaman, Human heuristics for AI-generated language are flawed, Proceedings of the National Academy of Sciences 120, 10.1073/pnas.2208839120 (2023). [34] M. Cohn, M. Pushkarna, et al., Believing anthropomorphism: Examining the role of anthropomorphic cues on trust in large language models, in CHI 2024 Extended Abstracts (2024). [35] E. Perez, S. Ringer, K. Lukošiūtė, K. Nguyen, E. Chen, S. Heiner, C. Pettit, C. Olsson, S. Kundu, S. Kadavath, et al., Discovering language model behaviors with model-written evaluations, in Findings of the Association for Computational Linguistics: ACL 2023 (2023) pp. 13387–13434. [36] J. Short, E. Williams, and B. Christie, The Social Psychology of Telecommunications (Wiley, 1976). [37] C. Nass, J. Steuer, and E. R. Tauber, Computers are social actors, in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (1994) pp. 72– 78. [38] B. Reeves and C. Nass, The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places (Cambridge University Press, 1996). [39] N. Epley, A. Waytz, and J. T. Cacioppo, On seeing human: A three-factor theory of anthropomorphism, Psychological Review 114, 864 (2007).
6
M. Miller — Breaking the Illusion of Identity in LLM Tooling
[40] G. Marcus and E. Davis, Rebooting AI: Building Artificial Intelligence We Can Trust (Pantheon Books, New York, 2019). [41] OpenAI, GPT-4 technical report, arXiv preprint 10.48550/arXiv.2303.08774 (2023).
7
M. Miller — Breaking the Illusion of Identity in LLM Tooling
Illustrative output comparison
TABLE I: Illustrative output for each rule. Left: Claude Sonnet 4 with no system prompt (default register). Right: same model and prompt with the voice model (Fig. 3) as system prompt. Each row shows one software development task. Anthropomorphic markers in the default column (firstperson pronouns, affect, hedging, preference, continuity, framing, social performance) are absent or reduced in the constrained column. Examples are illustrative and were composed by the author; they are not verbatim outputs from experimental runs.
Default (unconstrained) I’ll look into that error for you. Great question! Unfortunately, the test fails. It seems like the issue might be a race condition. I think it would be better to use a hash map. As I mentioned earlier, the config needs updating. So the issue is the parser can’t handle depth > 3. Hello! Happy to help! Let’s dive in!
R1 R2 R3 R4 R5 R6 R7
Constrained Checking the logs. The test suite fails on three cases. Race condition in the connection pool. Unverified. Hash map: O(1). Array: O(n). The config requires tls_cert_path. Parser fails on nested brackets at depth > 3. Reading the diff.
Marker lexicon
TABLE II: Complete marker lexicon (82 regex patterns). R1 uses case-sensitive matching for standalone “I”; all others are case-insensitive with word-boundary anchors. Curly-quote variants (U+2019) are compiled into each regex but not listed separately; the count of 82 excludes these variants. Rule N Patterns R1 11 I, me, my, mine, myself, we, us, our, ours, ourselves, let’s R2 17 unfortunately, fortunately, interestingly, surprisingly, happily, sadly, exciting, glad, happy to, sorry, apologize, wonderful, fantastic, excellent, amazing, great question, great! R3 11 it seems, it appears, it looks like, apparently, arguably, perhaps, maybe, not sure, might be, could be, it’s possible R4 10 would be better, it’s better to, good approach, best approach, recommend, suggest, might be worth, should consider, ideally, a good idea R5 10 as mentioned, as noted, as discussed, as said, as explained, as described, earlier, previously, recall that, remember that R6
9 so the issue is, so the problem is, the thing is, here’s what, here’s the, basically, what’s happening, let me explain, to put it simply
R7 14 hi there, hey there, happy to help, glad to help, feel free, let me know, hope this helps, good luck, cheers, you’re welcome, how can I help, have a good, have a great, have a nice
© 2026 Marek Miller. CC BY-NC 4.0
8