acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Turing Test: Evaluating AI Agents in Group Conversation

Abstract

Turing’s Imitation Game tests machine intelligence through one-on-one turn-based conversation, but it does not capture the social dynamics of free-form human communication in group settings, where participation patterns and adherence to group norms complement linguistic content. To address this limitation, we introduce the Social Imitation Game, a communication game in which a group of human participants collectively attempts to identify which group member is an AI agent through successive rounds of conversation, each followed by a vote on the AI’s identity. This design produces a natural gradient of difficulty and score, as the most suspected player is eliminated after each round. We develop an LLM-based agent comprising a foreground pipeline that decides when to send a message and what message to generate, along with background-worker subagents that perform longer tasks, involving reasoning over the conversation state, retrieving external information, and maintaining memory. This architecture enables the agent to remain responsive while drawing on information accumulated across messages and rounds to support coherent social behavior. We evaluate the agent by collecting a novel dataset of Social Imitation Games involving hundreds of unique human players. The results show that our agent performs almost on par with human participants in terms of elimination and votes received, leaving little room for further improvement to achieve human parity. Further semantic analysis of the messages suggests that the agent takes a more reactive role in conversations, asking fewer questions than human participants. Together, our game formulation, agent architecture, and dataset provide a new framework for studying socially situated language agents beyond the classic Imitation Game.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
large language models, language agents, multi-party dialogue, social intelligence, human-AI interaction, Turing test, agent evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3EKJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces the Social Imitation Game, a multi-party extension of the Turing Test in which an AI agent participates in live group conversations and human participants repeatedly vote to identify and eliminate the AI. It further develops a real-time agent architecture that separates latency-sensitive message scheduling/generation from asynchronous background processes for social-state tracking, memory, retrieval, and behavioral adaptation. The system is evaluated in 50 live games with 234 human participants; the agent's survival and received votes are close to chance-level baselines, while behavioral analysis identifies some remaining differences between the agent and humans.

Critical strengths

A compelling evaluation setting. The Social Imitation Game moves evaluation from static or dyadic judgments to live, asynchronous multi-party interaction with real humans. Repeated voting and elimination give a natural graded outcome rather than relying exclusively on an offline LLM judge. This is a meaningful evaluation contribution for socially situated agents.

Strong ecological validity and unusually good system documentation. The study evaluates an actually deployed agent interacting with hundreds of human participants rather than only simulated interlocutors, while the implementation, prompts, runtime configuration, and behavioral analyses are documented in substantial detail. This makes the work potentially useful as a dataset and experimental framework beyond the specific reported agent.

Critical weaknesses

The architectural contribution is not empirically isolated. The main result demonstrates that one extensively engineered system is difficult for humans to identify, but there is no sufficiently strong baseline showing what is gained by the proposed foreground/background architecture. In particular, comparisons against a simpler prompt-only agent, foreground-only system, or variants without individual background components would be essential for determining whether the architecture contributes meaningfully beyond the strong backbone model and extensive behavioral prompting. The paper itself acknowledges that the effect of the background workers is difficult to quantify and supports their contribution mainly through qualitative examples.

The broader capability claim is stronger than the experimental scope supports. The evaluation uses one fixed final agent configuration, one main backbone, relatively short games, and a protocol that was preceded by substantial prompt engineering and pilot development. Thus, the results convincingly establish that this particular engineered system can be difficult to detect under this particular protocol, but provide substantially weaker evidence that current LLM agents generally exhibit human-level social behavior in group conversation. The authors themselves appropriately acknowledge that the experiments do not establish architectural optimality.

Questions for authors

Assessment-critical: How does the final system compare against minimally simpler baselines using the same backbone—for example, a single-prompt agent, the foreground pipeline without background workers, or an agent without the Reviewer? Even a limited controlled comparison would substantially affect my assessment of the architectural contribution.

Clarification: How extensive was the development process used to tune the anti-detection behavior, including the number of pilot games and iterations informed by human feedback? Relatedly, is there evidence that the main result transfers across different backbone models, personas, or interaction durations rather than depending strongly on the final tuned configuration?

6 comments
AuthorsOpenReview ↗

1. Requested Baseline Comparison

We thank the reviewer for the detailed feedback. We address the concerns below, including a new live ablation study comparing our full agent against a reduced version of the architecture.

To address the request for a baseline, we conducted an additional live ablation study consisting of 25 games with all six background workers disabled, compared with 50 full-system games. Both retained Gemini 2.5 Flash and the Scheduler–Generator–Reviewer pipeline, allowing us to assess the added value of the background layer.

Turning off the background agents led to a drop of 10% in games won. The full system won 30% (15/50), versus 20% (5/25) without background workers, and normalized survival was similar (0.548 vs. 0.547). A clearer difference emerged in social participation: with workers, AI and humans each averaged 3.84 messages per active player-round; without workers, AI averaged 3.68 versus 4.14 for humans. The gap was +0.02 with workers versus −0.51 without (difference: 0.53 messages; 95% game-clustered bootstrap CI [0.09, 0.94]). Message length also matched humans more closely with workers (with 4.44 vs. 4.59 words; without, 3.80 vs. 4.60). This suggests workers may help the agent calibrate its participation to the room.

The background worker-off games also reveal specific failures that the background workers are designed to address. In one worker-off game, the agent failed to answer a question about the 2026 World Cup, responding "no idea tbh". Two participants cited its responses as suspicious, and the agent was eliminated. The Internet Context worker could potentially have helped by retrieving up-to-date information. In another worker-off game, the agent repeatedly gave generic agreement, prompting a participant to remark "he is validating anything i guess". The agent was also eliminated. The Repetition Monitor and Behavior Adapter are designed to identify such patterns and encourage more varied, context-appropriate responses. Importantly, Table 3 shows these mechanisms operating in the full system.

These results help clarify what the background workers add to an already capable foreground pipeline. We observe better alignment with human participation and a higher, though statistically uncertain, win rate. The examples illustrate both relevant failure modes and how worker outputs inform foreground decisions. We will add the ablation results and their limitations to the revised paper.

AuthorsOpenReview ↗

2. Development Process and Potential Overfitting

We understand the concern that extensive development and tuning may limit generalization. Importantly, development and final evaluation involved separate participant populations. Development relied on a small, relatively homogeneous group of in-house testers, whereas the final evaluation included 234 previously unseen participants from 37 countries, recruited through Prolific. Participants ranged in age from 18 to 70, and 155 of 215 with available student-status information were non-students. This demonstrates performance beyond the development population, reducing concerns about overfitting to those participants.

The development phase focused mainly on system reliability and improving natural conversational behavior, including timing, repetition, and response selection, through debugging, prompt refinement, and human feedback. Development involved extensive iterative pilot testing with a small group of in-house testers, including 12 documented Generator–Reviewer configurations, separate from the 50 final evaluation games.

The 2.5-minute round duration and main game settings were chosen in advance for participant experience and retention, rather than to maximize agent performance. We will clarify the development process, its separation from the final evaluation, and these design choices in Section 5.1.

AuthorsOpenReview ↗

3. Generalizability and Scope of Claims

We appreciate the reviewer's question about whether our results generalize beyond the final tuned configuration. Our contribution is to extend Turing's original Imitation Game to live group conversations, where appearing human requires more than generating convincing text. An agent must also decide when to speak, whom to respond to, and how to adapt to evolving group dynamics. Our results demonstrate that a carefully designed real-time agent can blend into such conversations under the tested conditions.

As discussed above, the final evaluation involved a diverse population of previously unseen participants, providing evidence of performance beyond the development population. However, we have not systematically evaluated generalization across backbone models, personas, or interaction durations, and therefore do not claim that these results necessarily transfer to those settings.

Regarding interaction length, 2.5 minutes refers to a single discussion round, not the entire game. To win, the agent must survive until the final round, not just blend in for a single 2.5-minute round. In games lasting 3–4 rounds, this means remaining indistinguishable from humans throughout a continuous conversation of 7.5–10 minutes. As the game progresses, participants retain the full chat history, accumulate evidence, and can reassess their suspicions based on the agent's behavior over time. Median discussion time was 7.5 minutes, with separate medians of 50 total messages and 11 agent messages.

We will revise the abstract and conclusion to better distinguish our demonstrated contributions from broader generalization claims.

Reviewer_SzvWOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Social Imitation Game, a communication game where human participants are involved in identifying which group member is an AI, while the AI agent is trying to disguise its identity during the multi-turn conversation. In addition, this paper designs an LLM-based agentic system and shows that it performs on par with human players in Social Imitation Game.

Critical strengths
  1. The idea is novel and somewhat groundbreaking.
  2. The human-involved experiments are extensive, and the findings are insightful.
Critical weaknesses
  1. The validity of the proposed game's design & pipeline (serving as an improved Turing Test in group social scenarios) is of my concern. For example, this paper should provide more arguments and evidence on the correlation between the success rate in the proposed game and the "intelligence level" of AI.
  2. Lack of in-depth analysis and case studies to investigate why existing models perform poorly in the proposed game.
Questions for authors
  1. Figure 4 results: Why do Claude and GPT perform much worse than random guessing? Any analysis?
  2. Please consider addressing the issues raised in "Critical Weaknesses".
2 comments
AuthorsOpenReview ↗

1. Validity of the Evaluation

We thank the reviewer for the thoughtful feedback. We address the concerns below, including a new experiment and analysis of the LLM judges' errors.

We appreciate the reviewer's concern about the relationship between game performance and AI intelligence. We would like to clarify that our goal is to extend Turing's original Imitation Game to live group conversations, where appearing human requires more than generating convincing text. In this setting, the agent must decide when to speak, whom to respond to, and how to adapt to changing topics, group norms, and suspicion from other participants. We introduce the Social Imitation Game to evaluate these social aspects of intelligence, which are largely absent from traditional one-on-one, turn-based evaluations.

Importantly, we evaluate more than whether the agent wins or loses. We also examine votes received and compare its conversational behavior with that of human players, including participation frequency, message length, and interaction patterns. Our new ablation further shows that removing background workers reduces how closely the agent matches human participation patterns. Together, these results demonstrate the value of evaluating social behavior through live group interaction. Our claims therefore concern the social abilities demonstrated in this setting, rather than a general measure of AI intelligence, and we will make this distinction clearer in the revised paper.

AuthorsOpenReview ↗

2. Why Do GPT and Claude Struggle to Identify the AI?

We appreciate the reviewer's request to better understand why GPT and Claude perform below random guessing. To investigate this further, we conducted an additional experiment using GPT-5 with high reasoning effort, checking whether the difficulty persists with more reasoning. For robustness, we repeated each judgment three times, evaluating 135 conversation states across 50 games (405 judgments). Instead of asking only who the AI was, we asked the model to distribute 100 suspicion points among all remaining participants. This allowed us to examine not only its final choice, but also how strongly it suspected each candidate.

The results show that identifying the AI remains difficult even with high reasoning effort, although the suspicion scores provide a more detailed picture than detection accuracy alone. Across all conversation states, the AI received an average of 18.9 suspicion points out of 100, compared with 22.7 under equal allocation. GPT-5 selected the AI in 14.3% of judgments, compared with a random-choice reference of 22.7%.

To understand these errors, we examined whether GPT-5 was assigning higher suspicion scores to participants with particular conversational behaviors. We found that humans writing longer messages tended to receive higher suspicion scores. Those writing longer messages relative to their group received 20.9 points on average, compared with 16.4 for those writing shorter messages (difference: 4.5 points; 95% CI [0.45, 8.57]).

Our case studies suggest that the model may also mistake certain human behaviors for signs of AI. In one conversation, a human who asked questions in all four messages received an average suspicion score of 53, compared with 7 for the AI, which asked none. In another, a human who repeatedly emphasized being "100% human" received 64 points, compared with 7 for the actual AI. These cases illustrate possible reasons for the model's mistakes, although no single behavioral feature fully explains the errors.

Together, these findings provide new evidence about the judges' difficulties: Our analysis suggests that message length and other conversational behaviors may influence the judges' suspicion, helping explain some of their identification errors. We will incorporate the additional experiment, behavioral analysis, and representative cases into the revised paper.

Reviewer_rCcXOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Turing's test was considered as a targeted definition for AGI, but it is only one-on-one conversations, which doesn't capture the social perspective of conversational agents in multi-party conversations. The authors propose the Soical Imitation Game, where a mix of AI and human tries to identify which member is AI through group conversations and the suspect will be eliminated after each round. The core evaluation measurement for the AI participant is their chance to survive compared with randomized baselines. The authors also propose an agent harness for game participant AI, which consists of a scheduler, generator, and reviewer, with various asynchronous workflows run in background. The experiment on 50 games with 234 human participants and a fixed Gemini 2.5 Flash shows that the agent survives to the end in 30% of games, slightly lower than the baseline.

Critical strengths
  1. The paper addresses a meaningful gap in literature, which is the fact that the Turing test lacks social and conversational intelligence.
  2. The harness presented by the paper is well-engineered and represents the ceiling of agent performances. The further analysis is made more meaningful in discussing the limitations in agentic behaviors than barebone harnesses.
Critical weaknesses
  1. The game setting seems favoring AI's disguise in various different ways. The conversation window is relatively short (2.5 mins), AI sends 3.84 messages and 4.44 words per messages, which seems like a short group chat scenario rather than covering richer social dynamics that the paper claims as main contribution.
  2. The games start with a simple prompt for selecting the AI in the group chat, and no other contexts like certain simulated scenarios. The prompt is too simple to test the agent in more challenging and meaningful scenarios.
  3. The experiment seems very unscalable with difficulty to control and compare meaningfully. The agent's win rate is 30% versus a 35.5% chance reference, and the reported CI ranges from −17.3% to +7.2%, which makes the comparison difficult.
  4. The experiment bootstrapped on the game level and each human participant only takes once. This makes the experiment only meaningful in scaled manner, as single experiments are largely affected by participant sampling.
Questions for authors
  1. Can you propose potential ways to make this experiment more scalable?
  2. Can you discuss the benefits of this specific setting (conversation length, turns, simple starter prompt) over more diverse and complex social environments?
  3. Considering the experiments are not easy to reproduce, can you provide smaller indicator experiments comparing you agent harness with a baseline setup?
4 comments
AuthorsOpenReview ↗

1. Scalability and Statistical Reliability

We thank the reviewer for the detailed feedback. We address the concerns below, including a new live ablation study comparing our full agent against a reduced version of the architecture.

We agree that live human evaluation is more expensive and harder to control than automated benchmarks. However, we ran 50 games with 234 unique Prolific participants (£2 each) within an academic budget, demonstrating the feasibility of this approach. The same platform and experimental procedure can be reused for larger studies.

A practical way to scale the evaluation is to extend the platform to run multiple independent games in parallel, with automated participant assignment and matchmaking. Increasing the sample size of games collected would allow more precise estimates and stronger statistical comparisons.

One-time participation was also a deliberate methodological choice. Since the AI's identity is revealed after each game, returning participants could learn its writing style and behavior, shifting the task from detecting an unfamiliar AI to recognizing a specific agent encountered before. We chose new participants to avoid this bias, despite the additional variability.

Finally, the Social Imitation Game is independent of our specific agent architecture. As language models improve, they may be able to participate naturally in group conversations with much less additional engineering, making the framework easier to apply to a wider range of models. We will clarify these scaling paths and statistical limitations in the revised paper.

AuthorsOpenReview ↗

2. Why This Game Setting?

We understand the concern that game settings may favor AI disguise. We want to clarify that the reported 2.5 minutes and 3.84 AI messages refer to one round, not a whole game. To win, the agent must survive until the final round, not just blend in for a single 2.5-minute round. In games lasting 3-4 rounds, this means remaining indistinguishable from humans throughout a continuous conversation of 7.5-10 minutes. As the game progresses, participants retain the full chat history, accumulate evidence, and can reassess their suspicions based on the agent's behavior over time. Median discussion time throughout games was 7.5 minutes, separate medians were 50 total messages and 11 AI messages.

In Addition, the agent was designed to adapt message length and writing style to the humans. Humans averaged 4.59 words per message, and the agent closely matched them at 4.44. Their similar lengths are consistent with the intended adaptation, not evidence of a fixed short-response limit. We also examined whether the agent's performance could be explained simply by remaining quiet or sending unusually short messages. The agent participated in every round it survived, and its overall message frequency and length closely matched those of human players. Our agent did not rely on silence or a fixed short-response strategy to blend into conversations.

The free-form opening is intentional, following Turing's original Imitation Game. Humans choose topics, ask unexpected questions, and shift direction while trying to identify the AI. Importantly, the agent must navigate these interactions while the other participants actively try to identify it, responding to unexpected questions and accusations in real time, without fixed turns or predefined topics. Structured scenarios could test other skills but might also offer clearer goals and more predictable context.

We agree that our experiment does not test deep or long-term social relationships. Our focus is on a different, but meaningful, form of social interaction: blending into an everyday, informal group conversation. Even in this setting, the agent must decide when to speak, whom to respond to, and how to handle changing topics, questions, and suspicion. This follows the spirit of Turing's original Imitation Game, which tests whether a machine can be distinguished from a human through conversation. We will clarify the rationale, statistics, and scope in the revision.

AuthorsOpenReview ↗

3. Requested Baseline Comparison

To address the request for a baseline, we conducted an additional live ablation study consisting of 25 games with all six background workers disabled, compared with 50 full-system games. Both retained Gemini 2.5 Flash and the Scheduler–Generator–Reviewer pipeline, allowing us to assess the added value of the background layer.

Turning off the background agents led to a drop of 10% in games won. The full system won 30% (15/50), versus 20% (5/25) without workers, and normalized survival was similar (0.548 vs. 0.547). A clearer difference emerged in social participation: with workers, AI and humans each averaged 3.84 messages per active player-round; without workers, AI averaged 3.68 versus 4.14 for humans. The gap was +0.02 with workers versus −0.51 without (difference: 0.53 messages; 95% game-clustered bootstrap CI [0.09, 0.94]). Message length also matched humans more closely with workers (with 4.44 vs. 4.59 words; without, 3.80 vs. 4.60). This suggests workers may help the agent calibrate its participation to the room.

The worker-off games also reveal specific failures that the background workers are designed to address. In one worker-off game, the agent failed to answer a question about the 2026 World Cup, responding "no idea tbh". Two participants cited its responses as suspicious, and the agent was eliminated. The Internet Context worker could potentially have helped by retrieving up-to-date information. In another worker-off game, the agent repeatedly gave generic agreement, prompting a participant to remark "he is validating anything i guess". The agent was also eliminated. The Repetition Monitor and Behavior Adapter are designed to identify such patterns and encourage more varied, context-appropriate responses. Importantly, Table 3 shows these mechanisms operating in the full system.

These results help clarify what the background workers add to an already capable foreground pipeline. We observe better alignment with human participation and a higher, though statistically uncertain, win rate. The examples illustrate both relevant failure modes and how worker outputs inform foreground decisions. We will add the ablation results and their limitations to the revised paper.

Thank you again for your detailed review. With the discussion period closing soon, we wanted to check whether our response resolved your concerns. If so, we would be deeply grateful if you could consider raising your score, as every point makes a critical difference to our submission.

Please let us know if anything remains unresolved before the discussion closes, so we can provide further clarifications or run additional analyses if needed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions to AC

Dear Area Chair,

Thank you for your time and consideration. The reviewers' feedback helped us improve the paper.

We have uploaded a revised version of the paper, with all changes highlighted in red. Below, we summarize the main updates made in response to the reviewers' concerns.

1. Baseline comparison. To address the request for a baseline, we conducted 25 additional live games with all six background workers disabled, using the same model and foreground pipeline. The full agent achieved a higher win rate and more closely matched human players in both participation rate and message length.

2. LLM judges. To better understand why LLM judges struggle to identify the AI, we conducted a new experiment with GPT-5 using high reasoning effort, evaluating all conversation states with three repetitions each. Identifying the AI remained difficult even with high reasoning effort. Our analysis also suggests that message length and other human conversational behaviors may explain some of the judges' errors.

3. Game setting and evaluation validity. We clarified that games involve multiple rounds of continuous discussion, rather than a single 2.5-minute round, with a median of 50 messages per game. We also explained that the agent's participation rate and message length closely match those of humans, consistent with its goal of adapting to their writing styles. Finally, we clarified our choice of a free-form setting and clarified that our goal is to evaluate social aspects of agent behavior in group conversations, rather than treating game success as a general measure of AI intelligence. We also clarified that our proposed architecture enables an agent to participate naturally in live group conversations, without claiming that it is the optimal architecture for this task.

4. Development and scalability. We expanded the description of the development process and clarified that development involved a small group of in-house testers, while the final evaluation included 234 previously unseen participants from 37 countries. We also discussed the limits of generalization and how our platform can be used to evaluate other agent and game configurations in future studies.

5. Ethical considerations. To address concerns about AI impersonation, we trained an agent-specific detection classifier (macro-F1 = 0.937) that can distinguish our agent's messages from human messages. We added this classifier to the paper and explained that it may learn our agent's particular writing style rather than general AI-specific properties.

We believe these revisions strengthen the paper's three main contributions: the Social Imitation Game as a framework for evaluating social behavior in group conversations, a modular agent designed for this setting, and a dataset of live games with human participants.

We hope these additional experiments and revisions address the reviewers' main concerns. We would be grateful if you could consider the new experimental results and the revised manuscript in your final assessment of our submission.

Thank you again for your consideration.

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.