1
On the Role of Artificial Intelligence in Human-Machine Symbiosis
arXiv:2605.00440v1 [cs.AI] 1 May 2026
Ching-Chun Chang, Yuchen Guo, Hanrui Wang, Timo Spinde, and Isao Echizen
Abstract—The evolution of artificial intelligence (AI) has rendered the boundary between humanity and computational machinery increasingly ambiguous. In the presence of more interwoven relationships within human-machine symbiosis, the very notion of AI-generated information becomes difficult to define, as such information arises not from either humans or machines in isolation, but from their mutual shaping. Therefore, a more pertinent question lies not merely in whether AI has participated, but in how it has participated. In general, the role assumed by AI is often specified, either implicitly or explicitly, in the input prompt, yet becomes less apparent or altogether unobservable when the generated content alone is available. Once detached from the dialogue context, the functional role may no longer be traceable. This study considers the problem of tracing the functional role played by AI in natural language generation. A methodology is proposed to infer the latent role specified by the prompt, embed this role into the content during the probabilistic generation process and subsequently recover the nature of AI participation from the resulting text. Experimentation is conducted under a representative scenario in which AI acts either as an assistive agent that edits human-written content or as a creative agent that generates new content from a brief concept. The experimental results support the validity of the proposed methodology in terms of discrimination between roles, robustness against perturbations and preservation of linguistic quality. We envision that this study may contribute to future research on the ethics of AI with regard to whether AI has been used fairly, transparently and appropriately.
I. I NTRODUCTION RTIFICIAL intelligence (AI) originates in the pursuit of imitating the human mind through computational machinery whose behaviour evokes the belief that a thinking mind resides within it [1]–[5]. The boom in the field of AI, accompanied by rapidly growing capability and accessibility, has created unprecedented opportunities and possibilities, yet it has also brought forth a wide range of risks and concerns [6]– [11]. Hallucinations represent the tendency of AI models to fabricate factual assertions that are in reality false and misleading [12]–[15]. Toxicity occurs when harmful, offensive and unsafe content is elicited from AI models, thereby breaking alignment with ethical principles and social norms [16]–[20]. Algorithmic bias arises when skewed training data lead AI models to reinforce stereotypical and unfair associations across demographics, such as age, gender, income, education, race and cultural ethnicity [21]–[25]. Synthetic media generated
A
C.-C. Chang, Y. Guo, H. Wang, T. Spinde and I. Echizen are with the Information and Society Research Division, National Institute of Informatics, Tokyo, Japan. Y. Guo, and I. Echizen are also with the Graduate School of Information Science and Technology, University of Tokyo, Tokyo, Japan. Correspondence: C.-C. Chang (email: [email protected])
An Essay on Can Machines Think?
HUMAN write This should begin with definitions of the meaning of the terms machine and think...
AI assist A natural starting point is to define the terms machine and think...
create The question invites not only a technical evaluation of artificial intelligence but also a deeper reflection on...
Is the content human-written or AI-generated? If AI plays a role, what role does it play?
Fig. 1. An essay on a topic may be written by a human or generated by AI, with AI assuming assistive, creative or other roles, yet which form of human–machine collaboration occurred may no longer be transparent from the observed essay.
by AI models may be maliciously used to impersonate individuals, fabricate events and forge evidence, causing financial fraud, political manipulation, social panic and other forms of defamation and deception [26]–[29]. Against this backdrop, the identification of AI-generated information has become a matter of increasing significance. The evolution of AI has rendered the boundary between human beings and computational machines increasingly ambiguous in practical and societal contexts. It may be envisaged that human beings and intelligent machines are evolving towards an increasingly interwined relationship resembling a form of symbiosis, in which each complements the limitations and extends the capabilities of the other [30]–[35]. In this sense, the very notion of AI-generated information becomes difficult to define, for such information emerges not from either humans or machines in isolation, but from their mutual shaping. Therefore, a more pertinent question may not be whether a piece of information is generated in the mere presence of AI, but rather how AI has participated in its generation. This concept is captured in Philosophical Investigations by Ludwig Wittgenstein [36]: The meaning of a word is its use in the language. This dictum suggests that words derive their meaning from their functional role within a context. By analogy, the nature of AI lies in the role it assumes within human-machine symbiosis. In this view, AI is defined not merely by a binary notion of presence or absence. Rather, it is understood through the form of its participation in relation to humans, shifting the focus towards a taxonomy of participation. In scientific discovery, for instance, it is less informative to state simply that an AI agent participated than to situate it within a
2
spectrum of roles. Such a taxonomy may include theorists who contribute hypotheses and reasoning, experimentalists who undertake experimental design and empirical testing, and critics who evaluate and falsify the resulting claims. Likewise, in a piece of writing produced through humanmachine collaboration, AI may assume a variety of roles, including supportive assistance, principal creation and other forms of participation, as illustrated in Figure 1. In this study, we consider the problem of tracing the functional role assumed by AI in human-machine collaboration, with a focus on natural language generation. Such roles may leave little, if any, discernible evidence in the generated textual content, depending upon the taxonomy of roles and the nature of the task. In general, the role assumed by AI is often specified, either implicitly or explicitly, in the input prompt, yet becomes less apparent or altogether unobservable in the resulting output. In other words, the role may become obscured once detached from its dialogue context. This observation motivates our methodology, which analyses the role specified in the prompt and embeds it within the response at the time of generation. The role can later be extracted for resolving the nature of AI participation. II. BACKGROUND The following preliminaries provide the foundation for this study. We begin by describing the typical generative process of language models and then formulate the problem of inferring the role played by AI from the observed content. We subsequently review prior work on AI content detection in order to contextualise and position the present research. A. Language Generation Contemporary language models typically generate text by repeatedly predicting the next token in a sequence, proceeding one token at a time in an autoregressive loop [37]–[40]. Given an input prompt x, the model f autoregressively generates an output sequence y by iteratively predicting the next token conditioned on the prompt and all previously generated tokens. For notational simplicity, we assume that each text is represented as a sequence of tokens, as language models operate on tokens as the atomic units of text. The tokenisation and detokenisation procedures are therefore omitted throughout this paper. At each generation step t, the language model produces a logit vector z = f (x, y<t ) ∈ R∥V∥ ,
(1)
where V denotes the vocabulary, namely the set of all possible tokens associated with a language model. Applying the softmax function, the logits are transformed into a probability distributiony over the next token exp(zv ) . u∈V exp(zu )
P(yt = v | x, y<t ) = softmax(z)t = P
(2)
The next token is then sampled according to yt ∼ softmax(z).
(3)
This process repeats until the end-of-sequence (EOS) token is generated.
B. Problem Formulation The prompt x may implicitly or explicitly specify a functional role under which the model is expected to operate. The research problem is to infer, from the observable output y, the underlying role specified by the input x. Let r ∈ R denote a role (expessed as a sequence of tokens), where R is a predefined role space. Formally, the most plausible latent role r∗ induced by the input prompt x is defined as r∗ = arg max P(r | x).
(4)
r∈R
In practice, however, the source prompt is often unavailable, and only the derived content y can be observed. The objective is therefore to infer the underlying role from the content alone, that is, r̂ = arg max P(r | y), (5) r∈R
where r̂ is expected to coincide with the true role r∗ . A role space may include an assistive role, which edits existing content, and a creative role, which generates new content from a concept, together with other role definitions depending on the application context. We define r̂ = ∅ when the language model does not play a role in producing the observed content, in which case the content is regarded as human-written. C. AI Content Detection Prior work on AI content detection can broadly be divided into reactive and proactive paradigms. The reactive paradigm attempts to uncover, after release, identifiable signs left by language models within the content. By contrast, the proactive paradigm attempts to insert, before release, verifiable marks within content generated by language models, such that its origin may later be authenticated. For the reactive paradigm, existing methods are typically based on either statistical or data-driven models. Statistical models rely on the uncertainty of a language model with respect to the observed content, since text generation is founded upon the probabilistic sampling of tokens [41]–[44]. Detection methods based on statistical models are often interpretable and require little, if any, additional training cost. However, such methods may fail when the identifiable signs reside not in observable statistical patterns, but instead in deeper latent representations. Data-driven models learn discriminative features directly from large corpora of human-written and machine-generated texts [45]. While such high-dimensional features may be more effective in capturing subtle artefacts left by language models, they are often less interpretable and require substantial labelled data and training effort. In addition, detection methods based on data-driven models may fail to generalise to unseen content that differs from the examples encountered during training. The proactive paradigm is represented primarily by watermarking methods [46]. Rather than searching for discriminative patterns, these methods explicitly embed a mark into the content during the generation process and detection is subsequently performed by verifying its presence or absence. Although watermarking methods may provide stronger reliability, they depend upon the premise that the party operating
3
the language model is cooperative and subject to regulation. Consequently, they are not applicable to arbitrary or alreadyexisting content originating from an uncooperative source. The problem of AI content detection has been studied extensively in prior work, in which the task is typically framed as a binary classification between human-written content and machine-generated content. However, in the presence of increasingly interwoven relationships within human-machine collaboration, the question of how AI has participated in the generation process may extend far beyond such a binary classification. This study seeks to trace the role played by AI from the observed content. The proposed method is founded upon the premise of a cooperative generating party. In particular, it follows the proactive paradigm and is developed on the basis of watermarking techniques. III. M ETHODOLOGY The methodology of this study consists of the following procedures: role classification, role encoding and role decoding. To begin with, the latent role implied by the input prompt is inferred through meta-prompting. The identified role is then encoded within the probabilistic generation process, leaving statistical traces in the generated content, from which the role can subsequently be decoded. An overview of the proposed methodology is summarised in Algorithm 1. A. Role Classification To operationalise role classification, the input prompt is reformulated into a meta-prompt that treats the original instruction as the object of inference and asks the model to identify the latent role it implies. An example meta-prompt, which defines the candidate roles and asks the model to identify the appropriate one, is given below: Definitions: Assistive — edit given content. Creative — generate content about given concept. Classify this instruction: [x] In-context examples may optionally be incorporated into the meta-prompt to further illustrate how different prompts correspond to distinct roles. Given the meta-prompt xΩ , the latent role is inferred by selecting the candidate role with the highest length-normalised log-probability r∗ = arg max r∈R
1 log P(r | xΩ ) ∥r∥ ∥r∥
1 X = arg max log P(rt | xΩ , r<t ) ∥r∥ t=1 r∈R
(6)
Since the candidate roles may consist of different numbers of tokens, length normalisation is applied to avoid favouring shorter role descriptions. B. Role Encoding Role encoding introduces statistical traces associated with the latent role r∗ into the probabilistic generation process, such that the generated content y carries evidence of the underlying
role. To this end, each candidate role r ∈ R is associated with a designated subset of the vocabulary, denoted by Wr ⊂ V. During generation, the language model is biased towards tokens corresponding to r∗ , thereby leaving role-specific traces in the generated sequence. Each role-specific token subset is randomly sampled from the vocabulary with equal cardinality ∥Wr ∥ = q∥V∥,
(7)
where q ∈ (0, 1) denotes the vocabulary coverage rate, controlling the proportion of vocabulary tokens assigned to each role. At generation step t, the language model produces a logit vector z ∈ R∥V∥ . To increase the likelihood of generating tokens associated with r∗ , the logits corresponding to the token subset Wr∗ are increased by a constant bias δ, yielding ( zv + δ, if v ∈ Wr∗ , ′ zv = (8) zv , otherwise. Consequently, tokens associated with r∗ become more likely to be sampled. Applying the softmax function, the biased logits are converted into a probability distribution P(yt = v | x, y<t ) = softmax(z′ )v = P
exp(zv′ ) , (9) ′ u∈V exp(zu )
from which the next token is sampled and appended to the generated sequence. Repeating this procedure throughout generation creates statistical evidence that can later be used to infer the latent role from the generated content alone. C. Role Decoding Role decoding seeks to determine the latent role from an observed query sequence y. The decoding procedure evaluates whether y contains an unusually large number of tokens from any vocabulary subset Wr . Under the null hypothesis H0 , the observed sequence is assumed not to exhibit any preference for Wr . Consequently, each generated token yt independently falls into Wr with probability q, that is, H0 : P(yt ∈ Wr ) = q =
∥Wr ∥ . ∥V∥
(10)
Let N denote the random variable representing the number of tokens that belong to Wr under H0 . It then follows that N ∼ Binomial(∥y∥, q). Its probability mass function is therefore ∥y∥ n P(N = n) = q (1 − q)∥y∥−n . n
(11)
(12)
For an observed generated sequence y, the number of tokens belonging to Wr is computed as nr =
∥y∥ X
I yt ∈ Wr ,
(13)
t=1
The statistical significance of the role r is then quantified by the upper-tail probability ∥y∥ X ∥y∥ n pr = P N ≥ n r = q (1 − q)∥y∥−n (14) n n=n r
4
// --- role encoding --Function Encoder(x, r∗ ): t←1 while yt−1 ̸= EOS do z ← f (x, y<t ) for v ∈ V do if v ∈ Wr∗ then zv′ ← zv + δ else zv′ ← zv sample yt ∼ softmax(z′ ) append yt to y t←t+1 return y // --- role decoding --Function Decoder(y): for r ∈ R do P∥y∥ nr ← t=1 I yt ∈ Wr P∥y∥ ∥y∥ n pr ← n=nr n q (1 − q)∥y∥−n if minr∈R pr ≥ θ then r̂ ← ∅ else r̂ ← arg minr∈R pr return r̂ // --- main program --define role space R define vocabulary coverage ratio q define vocabulary subsets Wr for all r ∈ R define statistical significance threshold θ r∗ ← Classifier(x) y ← Encoder(x, r∗ ) r̂ ← Decoder(y)
A smaller pr indicates stronger evidence that y contains more tokens from Wr than would be expected by chance, and is therefore more likely to have been generated under role r. Finally, the role is determined according to the minimum pvalue among all candidates. If no candidate yields statistically significant evidence, that is, if the minimum p-value exceeds the significance threshold θ, no role is assigned. The resulting decision rule is given by ( ∅, if minr∈R pr ≥ θ, r̂ = (15) arg minr∈R pr , otherwise. IV. E VALUATION The experimentation in this study focuses on a restricted but well-defined scenario in which AI acts either as an
Number of Words (Concept)
// --- role classification --Function Classifier(x): formulate meta-promptP xΩ ∥r∥ 1 r∗ ← arg maxr∈R ∥r∥ t=1 log P(rt | xΩ , r<t ) ∗ return r
Concept Content
500
359±93
80
400 247±94
215±99
60
300
41±14 159±44
40 20
9±3
8±2
0
Number of Words (Content)
100
Algorithm 1: Role-Aware AI
IMDb
200 100
4±2
CNN/DailyMail
Wikipedia
arXiv
0
Fig. 2. Statistics of concept and content lengths across datasets.
assistive agent that edits human-written content or as a creative agent that generates text about target concepts. The proposed method is evaluated from three complementary perspectives: discriminability, robustness and perplexity. A. Experimental Setup Datasets: Experiments are conducted on four benchmark datasets spanning diverse domains, including entertainment, journalism, general knowledge, and scientific academia: • IMDb dataset consists of movie reviews annotated with binary sentiment labels (positive or negative) [47]. We define ‘<sentiment> review of the film <title>’ as the concept and the corresponding <review> as the content. • CNN/DailyMail dataset consists of journalistic articles paired with summarised highlights [48]. We define the <highlight> as the concept and the corresponding <article> as the content. • Wikipedia dataset consists of encyclopaedic topics accompanied by descriptive summaries [49]. We define the <topic> as the concept and the corresponding <description> as the content. • arXiv dataset consists of scientific research papers with titles and abstracts [50]. We define the <title> as the concept and the corresponding <abstract> as the content. A total of 1,000 evaluation samples are randomly selected from each dataset, subject to the constraint that each concept contains fewer than 100 words and its corresponding content contains 100 to 500 words. The word distribution for each dataset is shown in Figure 2. Models: For natural language generation, two opensource models are employed: • GPT-2 (base version with 124 million parameters) developed by OpenAI. • LLaMA-3 (instruction-tuned version with 3 billion parameters) developed by Meta. Using the evaluation samples from each dataset, each model generates 1,000 texts acting as an assistive agent conditioned on the human-written contents and 1,000 texts acting as a creative agent conditioned on the human-written concepts. For assitive writing, the prompts follow the template: Please <verb> the following paragraph: <content>,
5
TABLE I B INARY C LASSIFICATION ACCURACY ACROSS M ODELS AND DATASETS T IERS : ≤ 0.6 ≤ 0.7 ≤ 0.8 ≤ 0.9 ≤ 1.0
GPT-2 (124 million parameters)
All Datasets
arXiv
Wikipedia
CNN/DailyMail
IMDb
Entropy
LogLikelihood
LogRank
Curvature
LLaMA-3-Instruct (3 billion parameters) RoBERTa
Ours
Entropy
LogLikelihood
LogRank
Curvature
RoBERTa
Ours
AUC ACC AUC
ACC
AUC ACC AUC ACC AUC ACC AUC ACC AUC ACC AUC
ACC
AUC ACC AUC ACC AUC ACC AUC ACC
H vs A
0.82
0.74
0.71
0.64
0.96
0.91
0.89
0.81
0.95
0.90
0.93
0.86
0.97
0.91
0.98
0.93
0.98
0.92
0.79
0.74
0.86
0.81
1.00
0.97
H vs C
0.91
0.83
0.95
0.88
1.00
0.97
0.94
0.87
0.96
0.92
1.00
1.00
0.99
0.98
1.00
1.00
1.00
0.99
0.91
0.86
0.98
0.95
0.99
0.95
A vs C
0.32
0.51
0.89
0.82
0.13
0.51
0.61
0.58
0.58
0.56
0.90
0.85
0.09
0.50
0.95
0.92
0.04
0.50
0.82
0.78
0.75
0.69
1.00
0.98
Average 0.68
0.70
0.85
0.78
0.69
0.80
0.81
0.75
0.83
0.79
0.94
0.90
0.68
0.80
0.98
0.95
0.67
0.80
0.84
0.79
0.86
0.82
0.99
0.97
H vs A
0.41
0.51
0.14
0.51
0.67
0.61
0.39
0.50
0.99
0.96
0.96
0.95
0.75
0.67
0.85
0.77
0.84
0.76
0.63
0.62
0.71
0.71
0.99
0.94
H vs C
0.36
0.50
0.45
0.51
0.94
0.87
0.50
0.53
0.99
0.96
0.97
0.90
0.70
0.69
0.86
0.83
0.85
0.83
0.49
0.53
0.96
0.93
0.97
0.92
A vs C
0.53
0.53
0.86
0.79
0.15
0.51
0.60
0.57
0.50
0.51
0.86
0.85
0.51
0.56
0.64
0.62
0.35
0.53
0.39
0.52
0.78
0.73
0.99
0.97
Average 0.43
0.52
0.48
0.60
0.59
0.66
0.50
0.53
0.83
0.81
0.93
0.90
0.65
0.64
0.78
0.74
0.68
0.70
0.50
0.56
0.82
0.79
0.98
0.94
H vs A
0.55
0.52
0.24
0.50
0.84
0.77
0.62
0.59
0.94
0.88
0.96
0.91
0.77
0.69
0.82
0.74
0.80
0.72
0.71
0.65
0.72
0.70
0.96
0.90
H vs C
0.51
0.49
0.68
0.63
0.97
0.92
0.77
0.71
0.95
0.90
1.00
0.97
0.97
0.92
1.00
0.97
0.99
0.96
0.83
0.74
0.96
0.91
0.95
0.89
A vs C
0.54
0.53
0.88
0.80
0.13
0.50
0.66
0.60
0.57
0.54
0.90
0.88
0.11
0.50
0.97
0.91
0.04
0.50
0.67
0.63
0.86
0.77
0.99
0.96
Average 0.53
0.51
0.60
0.64
0.65
0.73
0.68
0.63
0.82
0.77
0.95
0.92
0.62
0.71
0.93
0.87
0.61
0.73
0.74
0.67
0.85
0.79
0.97
0.92
H vs A
0.85
0.79
0.64
0.61
0.96
0.90
0.76
0.70
0.92
0.85
0.88
0.79
0.89
0.83
0.91
0.87
0.90
0.85
0.86
0.81
0.65
0.61
1.00
0.96
H vs C
0.87
0.81
0.89
0.80
0.99
0.96
0.87
0.78
0.92
0.85
0.99
0.96
1.00
0.98
1.00
0.99
1.00
0.99
0.99
0.95
0.90
0.82
1.00
0.97
A vs C
0.46
0.49
0.81
0.73
0.22
0.50
0.66
0.59
0.51
0.49
0.64
0.60
0.12
0.50
0.98
0.95
0.02
0.50
0.89
0.80
0.81
0.74
1.00
0.99
Average 0.73
0.70
0.78
0.71
0.72
0.79
0.76
0.69
0.78
0.73
0.83
0.79
0.67
0.77
0.96
0.94
0.64
0.78
0.91
0.85
0.79
0.72
1.00
0.98
H vs A
0.65
0.64
0.43
0.57
0.86
0.80
0.66
0.65
0.95
0.90
0.93
0.88
0.84
0.78
0.89
0.83
0.88
0.81
0.75
0.70
0.73
0.71
0.98
0.94
H vs C
0.66
0.66
0.74
0.71
0.97
0.93
0.77
0.72
0.95
0.91
0.99
0.96
0.92
0.89
0.96
0.95
0.96
0.94
0.80
0.77
0.95
0.90
0.98
0.93
A vs C
0.46
0.51
0.86
0.78
0.16
0.50
0.63
0.59
0.54
0.52
0.83
0.80
0.21
0.51
0.89
0.85
0.11
0.51
0.69
0.68
0.80
0.73
0.99
0.98
Average 0.59
0.61
0.68
0.69
0.66
0.74
0.69
0.65
0.81
0.78
0.91
0.88
0.66
0.73
0.91
0.88
0.65
0.75
0.75
0.72
0.83
0.78
0.99
0.95
where <verb> is drawn at random from edit, improve, proofread, revise and summarise. For creative writing, the prompts are instead constructed according to the form: Please <verb> a paragraph about: <concept>, where <verb> is chosen randomly from compose, create, draft, generate and write. Hyperparameters: This study focuses on the scenario in which AI acts as either an assistive or a creative agent; accordingly, the role space R is defined as {assistive, creative}. The vocabulary coverage rate q is set to 0.5, such that half of the vocabulary is assigned higher sampling probability. The bias strength δ is set to 1.5 for GPT-2 and 3.0 for LLaMA3-Instruct. The significance threshold θ is selected adaptively via multi-fold cross-validation. Baselines: In our comparative study, the baselines include four zero-shot statistical methods, Entropy [41], LogLikelihood [42], LogRank [43], and Curvature [44], together with a data-driven detector based on the RoBERTa model [45]. These methods were originally developed for the binary task of distinguishing human-written from machine-generated text rather than for AI-role attribution. Nevertheless, in the present task setting, the assistive and creative roles differ partly in the degree of machine involvement: assistive writing remains more closely tied to the original human-written text, whereas creative writing is generated more independently by the language
model. The baseline methods therefore remain applicable insofar as they capture a notion of machine-likeness. B. Discriminability Analysis We analyse the discriminability of the proposed system across different datasets and language models in comparison to the baselines. Our experimentation is designed to examine three complementary aspects: performance on pairwise binary classification, performance on joint ternary classification and the distribution of discriminative features. Pairwise analysis: Table I reports the performance on the three pairwise binary classification tasks, namely human versus assistive (H vs A), human versus creative (H vs C), assistive versus creative (A vs C) and average. For each method, the decision threshold is determined automatically via 10-fold cross-validation. In each round, one fold is held out for testing, while the remaining nine folds are used to select the threshold that yields the best performance. The selected threshold is then applied to the held-out fold. This process is repeated for all ten folds, and the results are averaged across folds. The baseline methods each produce a single scalar score, as they were originally designed for the binary task of distinguishing human-generated from AIgenerated content. By contrast, our method produces multiple features since it was originally designed for multi-class role
6
CNN/DailyMail
CNN/DailyMail
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
arXiv LogLikelihood RoBERTa
LogRank Ours
GPT-2 (124 million parameters)
IMDb
IMDb
Entropy Curvature
0.2
Wikipedia
Wikipedia
0.2
Entropy Curvature
arXiv LogLikelihood RoBERTa
LogRank Ours
LLaMA-3-Instruct (3 billion parameters)
Fig. 3. Ternary classification accuracy across models.
attribution, in which a separate p-value is computed for each AI-related role. To adapt our method to the pairwise analysis, these multiple features are decomposed into distinct decision variables for each binary subtask. Specifically, pA is used for distinguishing between H and A, pC is used for distinguishing between H and C, pC − pA is used for distinguishing between A and C. Overall, the proposed method consistently achieves the best performance across all datasets and both language models. It attains an average AUC of 0.91 and an average ACC of 0.88 under GPT-2, and the performance further improves to 0.99 and 0.95, respectively, under LLaMA-3-Instruct due to the model’s ability in understanding and classifying the role. The advantage of the proposed method is particularly pronounced on the more challenging role-attribution task (A vs C), in contrast to the more conventional task of detecting AI-generated text (H vs A and H vs C). The baseline methods exhibit substantial degradation when both classes correspond to AI-generated content and differ only in the role played by the model. This observation suggests that existing reactive detectors primarily capture the presence of machine generation, but are limited in their ability to distinguish between different modes of generation. These results therefore highlight the need for a role-attribution framework in this context.
by exhaustively searching the ordered feature axis so as to partition the feature axis into three regions that maximise the three-class accuracy on the nine training folds. In contrast, the proposed method involves only one threshold θ, applied to the p-value of each role. If none of the role-specific pvalues is smaller than θ, the sample is regarded as lacking sufficient statistical significance for any role. In other words, a text is classified as human-written if all p-values exceed θ; otherwise, it is assigned to the role with the smallest p-value. The optimal threshold θ is selected by searching over the ordered values of min(pA , pC ) on the training folds. A similar trend is observed for the ternary classification task, where the proposed method generally attains the highest performance, or performance comparable to the strongest baseline, across the datasets and both language models. The advantage is particularly pronounced on CNN/DailyMail, whose concepts and contents contain the largest number of words. In this dataset, the prompts for creative writing are based on news highlights rather than brief concepts, similar to the prompts used to generate the assistive writing. Consequently, the prompts for both creative and assistive writing contain substantial amounts of human-written text, making both AI-content detection and AI-role attribution more challenging.
Joint analysis: Figure 3 evaluates the performance on joint ternary classification. For each method, the threshold selection procedure is likewise performed by 10-fold crossvalidation. For each baseline method, the single scalar feature is assumed to be ordered such that samples from class H tend to yield the smallest values, followed by A, and then C, where larger values indicate stronger evidence of machine generation. The optimal threshold pair is obtained
Distributional analysis: Figure 4 and Figure 5 visualise the feature distributions across different classes for the baseline methods and the proposed method, respectively. The results from different datasets are aggregated, whereas different language models are shown separately. For the baseline methods, the assistive samples tend to occupy an intermediate region between the human and creative samples along the feature axis. This observation supports the assumption that
LLaMA-3-Instruct GPT-2 Probability Density Probability Density
7
2.0
2.0
3.5
6.5
3.5
1.6
1.6
2.8
5.2
2.8
1.2
1.2
2.1
3.9
2.1
0.8
0.8
1.4
2.6
1.4
0.4
0.4
0.7
1.3
0.7
0.0
0.0
0.0
0.0 0.0 0.0 6.5
2.0
6.0
4.8
3.6
2.4
1.2
0.0
2.0
6.0
4.8
3.6
2.4
1.2
0.0
3.5
3.0
2.4
1.8
1.2
0.6
0.3
0.6
0.9
1.2
1.5
0.0 3.5
1.6
1.6
2.8
5.2
2.8
1.2
1.2
2.1
3.9
2.1
0.8
0.8
1.4
2.6
1.4
0.4
0.4
0.7
1.3
0.7
0.0
0.0
0.0
0.0 0.0 0.0
6.0
4.8
3.6
2.4
Entropy
1.2
0.0
6.0
4.8
3.6
2.4
1.2
+ LogLikelihood
0.0
3.0
Human
2.4
1.8
1.2
LogRank Assistive
0.6
0.3
0.6
0.9
+ Curvature
1.2
1.5
0.0
5.0
2.5
0.0
2.5
5.0
5.0
2.5
+ RoBERTa
0.0
2.5
5.0
Creative
1.00
1.00
0.75
0.75
p-value (creative)
p-value (creative)
Fig. 4. Distribution of baseline features across models.
0.50
0.25
0.00 0.00
0.50
0.25
0.25
0.50
p-value (assistive) Human Assistive
0.75
1.00
Creative
GPT-2 (124 million parameters)
0.00 0.00
0.25
0.50
p-value (assistive) Human Assistive
0.75
1.00
Creative
LLaMA-3-Instruct (3 billion parameters)
Fig. 5. Distribution of role-specific p-values across models.
larger values indicate stronger evidence of machine generation. While the human and creative samples are often reasonably separable, the assistive samples still overlap substantially with both. The data-driven detector yields scores that are highly concentrated near the two extremes, yet still provides only limited separation of the assistive samples. These results suggest that the baseline methods primarily capture a notion of machine-likeness, which is insufficient to distinguish different roles in AI generation. In contrast, the proposed method forms a two-dimensional feature space. Human-written samples are concentrated near the upper-right region, where both p-values are large, indicating that neither the assistive nor the creative role is statistically significant. Assistive and creative samples are instead concentrated near the left and lower boundaries, respectively. Compared with the results associated with GPT-2, those associated with LLaMA-3-Instruct exhibit
clearer separation between the creative and assistive samples. These observations suggest that the proposed method captures role-specific evidence beyond a one-dimensional notion of machine-likeness, thereby enabling more reliable discrimination between different forms of AI involvement. C. Robustness Analysis Figure 6 evaluates the robustness of the proposed method under synonym substitution. In practice, AI-generated text may be humanised to evade or confuse detection. Synonym substitution is adopted because it constitutes one of the simplest and most controllable forms of such post-editing. Compared with stronger forms of paraphrasing, synonym substitution largely preserves the original semantics and sentence structure, thereby providing a standardised perturbation that isolates the
0.95
0.95
0.90
0.90
0.85
0.85
Accuracy
Accuracy
8
0.80
0.80
0.75
0.75
0.70
0.70
0.65
0
IMDb
10
20
30
40
50
60
70
Synonym Substitution Rate CNN/DailyMail Wikipedia
80
90
0.65
100
arXiv
0
IMDb
GPT-2 (124 million parameters)
10
20
30
40
50
60
70
Synonym Substitution Rate CNN/DailyMail Wikipedia
80
90
100
arXiv
LLaMA-3-Instruct (3 billion parameters)
Fig. 6. Evaluation of robustness against synonym substitution across models. TABLE II E VALUATION OF P ERPLEXITY ACROSS M ODELS AND DATASETS GPT-2
LLaMA-3-Instruct
H
A (unbiased)
A (biased)
C (unbiased)
C (biased)
H
A (unbiased)
A (biased)
C (unbiased)
C (biased)
IMDb
47.15
34.69
46.19
21.76
27.09
28.02
8.47
9.19
2.93
4.69
CNN/DailyMail
23.14
36.18
46.53
23.95
32.75
12.84
8.39
7.70
8.21
8.89
Wikipedia
26.10
33.87
44.81
21.67
32.13
9.71
6.41
6.04
2.56
3.36
arXiv
44.09
36.07
49.29
24.11
33.50
20.11
8.71
9.56
2.98
3.87
Overall
35.12
35.20
46.71
22.87
31.37
17.67
7.99
8.12
4.17
5.20
effect of lexical variation. In our experiments, a proportion of eligible words in each text is randomly replaced with synonyms obtained from the synsets of WordNet, a large lexical database of English in which words are grouped into synonym sets interlinked through semantic relations. Only content words with a valid part-of-speech tag (nouns, verbs, adjectives and adverbs) are considered, whereas function words and punctuation marks are preserved. The synonym substitution rate denotes the percentage of eligible words that are replaced. As the substitution rate increases, the accuracy decreases gradually for all datasets and both language models. Nevertheless, the degradation remains relatively limited under GPT-2, for which the accuracy decreases by less than approximately 0.06 even when all replaceable words are substituted. Under LLaMA3-Instruct, the decrease is slightly more pronounced, but still remains below approximately 0.14. Despite this degradation, the proposed method retains reasonably high accuracy over a broad range of substitution rates. In particular, the performance remains above 0.80 for most datasets and above 0.70 for the worst-performing dataset even at high substitution rates. These results suggest that the role information is not carried solely by the content words, but is also distributed across the remaining parts of the text.
D. Perplexity Analysis Table II reports the perplexity of the human-written texts, the unbiased assistive and creative texts, and their biased counterparts in order to evaluate the impact of biasing the generation distribution on text quality. Perplexity measures the uncertainty of a language model when predicting a given text. A lower perplexity score indicates that the model is less confused and better at predicting the words in a given sample. Although perplexity does not directly correspond to fluency or coherence as perceived by humans, it is often used as a proxy for text quality. We measure self-perplexity, in which the same model is used for both text generation and text evaluation. It can be seen that the biased texts exhibit noticeably higher perplexity than the corresponding unbiased texts, indicating that the biasing process affects the quality of the generated text to some extent. Creative writing often yields the lowest perplexity because it is mostly created by the language model without substantial human-written prompts and thus conforms most closely to the model’s internal distribution. By contrast, human-written text generally exhibits higher perplexity because it is not produced by the model and may therefore deviate from its preferred writing patterns. Assistive writing lies between these two extremes, since it is generated by modifying a human-written text and thus retains characteristics of both human-written and machine-generated language.
9
V. C ONCLUSION This study has investigated the problem of tracing the role played by AI in human-machine collaboration, moving beyond a binary classification between human-written and machine-generated content. A methodology has been proposed to infer the latent role specified by the prompt, encode this role into the generation process and subsequently decode it from the resulting text. In this manner, the functional role of AI can remain observable even after the original prompt and dialogue context are no longer available. The experimentation focuses on a representative scenario in which AI acts either as an assistive or a creative agent, yet the implications of this study extend beyond these particular forms of participation. The experimental results demonstrate that the proposed method achieves effective discrimination between assistive and creative roles, remains reasonably robust under lexical perturbations and preserves acceptable text quality despite the introduction of lexical bias during language generation. This study reflects the view that the significance of AI lies not merely in whether it has participated, but in how it has participated. Several directions remain for future research. First, a more intricate form of collaboration may involve a substantially richer taxonomy of roles, together with hierarchical relationships between them. In addition, an extension beyond natural language to other modalities, including visual and auditory content, is expected. Furthermore, tracing roles over multiturn interaction merits further investigation, since the role of AI may evolve dynamically according to the conversational context. We envision that this study may contribute towards a deeper understanding of human-machine symbiosis, in which the role assumed by AI becomes more transparent and traceable. ACKNOWLEDGEMENTS This work was supported in part by the Japan Society for the Promotion of Science (JSPS) under KAKENHI Grants JP21H04907 and JP24H00732, and in part by the Japan Science and Technology Agency (JST) under the CREST Grants JPMJCR20D3 and JPMJCR2562, including the AIP Challenge Program, and under the AIP Acceleration Grant JPMJCR24U3 and under the K Program Grant JPMJKP24C2. R EFERENCES [1] A. M. Turing, “Computing machinery and intelligence,” Mind, vol. 59, no. 236, pp. 433–460, 1950. [2] C. E. Shannon, “Presentation of a maze solving machine,” in Proc. Josiah Macy Jr. Found. Conf. Cybern., New York, NY, USA, 1951, pp. 173–180. [3] Y. LeCun, Y. Bengio, and G. E. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015. [4] V. Mnih et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015. [5] D. Ha and J. Schmidhuber, “Recurrent world models facilitate policy evolution,” in Proc. Int. Conf. Neural Inf. Process. Syst. (NeurIPS), vol. 31, Montréal, QC, Canada, 2018, pp. 2455–2467. [6] L. Floridi and J. W. Sanders, “On the morality of artificial agents,” Minds Mach., vol. 14, no. 3, pp. 349–379, 2004. [7] S. Russell, D. Dewey, and M. Tegmark, “Research priorities for robust and beneficial artificial intelligence,” AI Mag., vol. 36, no. 4, pp. 105– 114, 2015.
[8] M. Brundage et al., “The malicious use of artificial intelligence: Forecasting, prevention, and mitigation,” Apollo – University of Cambridge Repository, Tech. Rep., 2018. [9] L. Floridi et al., “AI4People–An ethical framework for a good AI society: Opportunities, risks, principles, and recommendations,” Minds Mach., vol. 28, no. 4, pp. 689–707, 2018. [10] E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell, “On the dangers of stochastic parrots: Can language models be too big?” in Proc. ACM Conf. Fairness Account. Transpar. (FAccT), Virtual Event, Canada, 2021, pp. 610–623. [11] L. Weidinger et al., “Taxonomy of risks posed by language models,” in Proc. ACM Conf. Fairness Account. Transpar. (FAccT), Seoul, Korea, 2022, pp. 214–229. [12] J. Maynez, S. Narayan, B. Bohnet, and R. McDonald, “On faithfulness and factuality in abstractive summarization,” in Proc. Annu. Meet. Assoc. Comput. Linguist. (ACL), Virtual Event, 2020, pp. 1906–1919. [13] Z. Ji et al., “Survey of hallucination in natural language generation,” ACM Comput. Surv., vol. 55, no. 12, pp. 1–38, 2023. [14] M. Dahl, V. Magesh, M. Suzgun, and D. E. Ho, “Large legal fictions: Profiling legal hallucinations in large language models,” J. Leg. Anal., vol. 16, no. 1, pp. 64–93, 2024. [15] Y. Zhang et al., “Siren’s song in the AI ocean: A survey on hallucination in large language models,” Comput. Linguist., vol. 51, no. 4, pp. 1373– 1418, 2025. [16] S. Gehman, S. Gururangan, M. Sap, Y. Choi, and N. A. Smith, “RealToxicityPrompts: Evaluating neural toxic degeneration in language models,” in Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), Virtual Event, 2020, pp. 3356–3369. [17] D. Hendrycks et al., “Aligning AI with shared human values,” in Proc. Int. Conf. Learn. Represent. (ICLR), Virtual Event, 2021, pp. 1–29. [18] J. Welbl et al., “Challenges in detoxifying language models,” in Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), Punta Cana, Dominican Republic, 2021, pp. 2447–2469. [19] A. Wei, N. Haghtalab, and J. Steinhardt, “Jailbroken: How does LLM safety training fail?” in Proc. Int. Conf. Neural Inf. Process. Syst. (NeurIPS), vol. 36, no. 80079–80110, New Orleans, LA, USA, 2023. [20] C. Anil et al., “Many-shot jailbreaking,” in Proc. Int. Conf. Neural Inf. Process. Syst. (NeurIPS), vol. 37, Vancouver, BC, Canada, 2024, pp. 129 696–129 742. [21] C. O’Neil, Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. New York, NY, USA: Crown Publishing Group, 2016. [22] T. Bolukbasi, K.-W. Chang, J. Zou, V. Saligrama, and A. Kalai, “Man is to computer programmer as woman is to homemaker? Debiasing word embeddings,” in Proc. Int. Conf. Neural Inf. Process. Syst. (NeurIPS), vol. 29, Barcelona, Spain, 2016, pp. 4356–4364. [23] A. Caliskan, J. J. Bryson, and A. Narayanan, “Semantics derived automatically from language corpora contain human-like biases,” Science, vol. 356, no. 6334, pp. 183–186, 2017. [24] S. Barocas and A. D. Selbst, “Big data’s disparate impact,” Calif. Law Rev., vol. 104, no. 3, pp. 671–732, 2016. [25] K. Crawford, Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven, CT, USA: Yale University Press, 2021. [26] B. Chesney and D. Citron, “Deep fakes: A looming challenge for privacy, democracy, and national security,” Calif. Law Rev., vol. 107, no. 6, pp. 1753–1820, 2019. [27] S. Greengard, “Will deepfakes do deep damage?” Commun. ACM, vol. 63, no. 1, pp. 17–19, 2019. [28] C. Vaccari and A. Chadwick, “Deepfakes and disinformation: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news,” Soc. Media + Soc., vol. 6, no. 1, pp. 1–13, 2020. [29] Y. Mirsky and W. Lee, “The creation and detection of deepfakes: A survey,” ACM Comput. Surv., vol. 54, no. 1, pp. 1–41, 2021. [30] J. C. R. Licklider, “Man-computer symbiosis,” IRE Trans. Hum. Factor Electron., vol. 1, no. 1, pp. 4–11, 1960. [31] L. Floridi, The Fourth Revolution: How the Infosphere is Reshaping Human Reality. Oxford, UK: Oxford University Press, 2014. [32] E. Kamar, “Directions in hybrid intelligence: Complementing AI systems with human intelligence,” in Proc. Int. Jt. Conf. Artif. Intell. (IJCAI), New York, NY, USA, 2016, pp. 4070–4073. [33] I. Seeber et al., “Machines as teammates: A research agenda on AI in team collaboration,” Inf. & Manag., vol. 57, no. 2, pp. 1–22, 2020. [34] D. Wang et al., “From human-human collaboration to human-AI collaboration: Designing AI systems that can work together with people,” in Proc. ACM Conf. Hum. Factors Comput. Syst. (CHI), Honolulu, HI, USA, 2020, pp. 1–6.
10
[35] A. Fügener, D. D. Walzner, and A. Gupta, “Roles of artificial intelligence in collaboration with humans: Automation, augmentation, and the future of work,” Manag. Sci., vol. 72, no. 1, pp. 538–557, 2025. [36] L. Wittgenstein, Philosophical Investigations. Oxford, UK: Basil Blackwell, 1953. [37] A. Vaswani et al., “Attention is all you need,” in Proc. Int. Conf. Neural Inf. Process. Syst. (NeurIPS), vol. 30, Long Beach, CA, USA, 2017, pp. 6000–6010. [38] T. Brown et al., “Language models are few-shot learners,” in Proc. Int. Conf. Neural Inf. Process. Syst. (NeurIPS), vol. 33, Virtual Event, 2020, pp. 1877–1901. [39] L. Ouyang et al., “Training language models to follow instructions with human feedback,” in Proc. Int. Conf. Neural Inf. Process. Syst. (NeurIPS), vol. 35, New Orleans, LA, USA, 2022, pp. 27 730–27 744. [40] E. Malach, “Auto-regressive next-token predictors are universal learners,” in Proc. Int. Conf. Mach. Learn. (ICML), Vienna, Austria, 2024, pp. 1–15. [41] T. Lavergne, T. Urvoy, and F. Yvon, “Detecting fake content with relative entropy scoring,” in Proc. Int. Conf. Uncovering Plagiarism Authorship Soc. Softw. Misuse (PAN), vol. 377, Patras, Greece, 2008, pp. 27–31. [42] S. Gehrmann, H. Strobelt, and A. Rush, “GLTR: Statistical detection and visualization of generated text,” in Proc. Annu. Meet. Assoc. Comput. Linguist. (ACL), Florence, Italy, 2019, pp. 111–116. [43] J. Su, T. Zhuo, D. Wang, and P. Nakov, “DetectLLM: Leveraging log rank information for zero-shot detection of machine-generated text,” in Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), Singapore, 2023, pp. 12 395–12 412. [44] E. Mitchell, Y. Lee, A. Khazatsky, C. D. Manning, and C. Finn, “DetectGPT: Zero-shot machine-generated text detection using probability curvature,” in Proc. Int. Conf. Mach. Learn. (ICML), vol. 202, Honolulu, HI, USA, 2023, pp. 24 950–24 962. [45] A. Uchendu, Z. Ma, T. Le, R. Zhang, and D. Lee, “TURINGBENCH: A benchmark environment for Turing test in the age of neural text generation,” in Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), Punta Cana, Dominican Republic, 2021, pp. 2001–2016. [46] J. Kirchenbauer, J. Geiping, Y. Wen, J. Katz, I. Miers, and T. Goldstein, “A watermark for large language models,” in Proc. Int. Conf. Mach. Learn. (ICML), vol. 202, Honolulu, HI, USA, 2023, pp. 17 061–17 084. [47] A. L. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y. Ng, and C. Potts, “Learning word vectors for sentiment analysis,” in Proc. Annu. Meet. Assoc. Comput. Linguist. (ACL), Portland, Oregon, USA, 2011, pp. 142– 150. [48] D. Chen, J. Bolton, and C. D. Manning, “A thorough examination of the CNN/Daily Mail reading comprehension task,” in Proc. Annu. Meet. Assoc. Comput. Linguist. (ACL), Berlin, Germany, 2016, pp. 2358–2367. [49] S. Merity, C. Xiong, J. Bradbury, and R. Socher, “Pointer sentinel mixture models,” in Proc. Int. Conf. Learn. Represent. (ICLR), Toulon, France, 2017, pp. 1–15. [50] C. B. Clement, M. Bierbaum, K. P. O’Keeffe, and A. A. Alemi, “On the use of arXiv as a dataset,” in Proc. Int. Conf. Learn. Represent. (ICLR) Workshop, New Orleans, LA, USA, 2019, pp. 1–7.
Ching-Chun Chang received the PhD in Computer Science from the University of Warwick, UK, in 2019. He is currently affiliated with the National Institute of Informatics, Japan, as a Project Assistant Professor. He also serves as a Visiting Researcher at Peking University, China, and a Distinguished Professor at Hangzhou Dianzi University, China. He participated in the ShortTerm Scientific Mission supported by European Cooperation in Science and Technology Actions at the Faculty of Computer Science, Otto von Guericke University of Magdeburg, Germany, in 2016. He was granted the MarieCurie Fellowship and participated in the Research and Innovation Staff Exchange supported by Marie Skłodowska-Curie Actions at the Department of Electrical and Computer Engineering, New Jersey Institute of Technology, USA, in 2017. He was a Visiting Scholar at the School of Computing and Mathematics, Charles Sturt University, Australia, in 2018, and at the School of Information Technology, Deakin University, Australia, in 2019. He was a Research Fellow at the Department of Electronic Engineering, Tsinghua University, China, in 2020. His research interests include artificial intelligence, biometrics, cryptography, cybersecurity, evolutionary computation, forensics, information theory, steganography, and watermarking.
Yuchen Guo received the BS degree from the Faculty of Informatics, Beijing University of Technology in 2022 and the MS degree from the University of Tokyo in 2025. He is currently pursuing the PhD degree at the University of Tokyo, while also working as a Research Assistant at the National Institute of Informatics, Japan. His research focuses on AI security and NLP security, with particular interests in the security, robustness, and trustworthiness of intelligent systems.
Hanrui Wang received the BS degree in Electronic Information Engineering from Northeastern University (China) in 2011. After working in the IT industry and rising to a director-level position, he transitioned to academic research in 2019 and earned the PhD in Computer Science from Monash University, Australia, in January 2024. He is currently an Assistant Professor at the National Institute of Informatics, Japan. His research focuses on AI security and privacy, with emphasis on adversarial machine learning, model inversion, and trustworthy multimedia systems. He has published multiple papers in leading journals such as IEEE TIFS, IEEE TDSC, ACM TOMM, and J-STARS, as well as in reputable international conferences including WWW, WACV, FG, ICPR, ICASSP, and ICMI.
Timo Spinde received the PhD in Computer Science from the University of Göttingen, Germany, in 2024, graduating summa cum laude. He also holds a master’s degree in Social and Economic Data Analysis from the University of Konstanz and completed dual bachelor’s degrees in Computer Science and Media and Communication at the University of Passau. He is currently a Postdoctoral Researcher with the National Institute of Informatics, Tokyo, Japan, and the University of Göttingen, Germany. He is also the initiator and coordinator of the Media Bias Group, an international research network dedicated to interdisciplinary research on media bias, and has founded multiple startups. His research interests include media bias detection, natural language processing, artificial intelligence, computational social science, and information quality.
Isao Echizen received BS, MS, and DE degrees from the Tokyo Institute of Technology, Japan, in 1995, 1997 and 2003, respectively. He joined Hitachi, Ltd. in 1997 and until 2007 was a Research Engineer in the company’s systems development laboratory. He is currently a Director and Professor of the Information and Society Research Division, as well as a Director of the Global Research Center for Synthetic Media, at the National Institute of Informatics; a Professor in the Department of Information and Communication Engineering, Graduate School of Information Science and Technology, the University of Tokyo; and a Professor in the Graduate University for Advanced Studies (SOKENDAI), Japan. He was a Visiting Professor at the Tsuda University, Japan; at the University of Freiburg, Germany; and at the University of HalleWittenberg, Germany. He is currently engaged in research on AI security, multimedia security and multimedia forensics, serving as a Research Director for the CREST FakeMedia project and the K Program SYNTHETIQ X project of the Japan Science and Technology Agency (JST). He received the Commendation for Science and Technology by the Minister of Education, Culture, Sports, Science and Technology (Research Category) in 2025. He also received the IEICE Best Paper Award in 2023; the IPSJ Best Paper Awards in 2005 and 2014; the IPSJ Nagao Special Researcher Award in 2011; the DOCOMO Mobile Science Award in 2014; the IISEC Information Security Cultural Award in 2016; and the IEEE WIFS Best Paper Award in 2017. He is an IEICE Fellow, an IPSJ Fellow, an IEEE Senior Member, an IFIP Japanese Representative, and an APSIPA Vice President.