When large language models (LLMs) were assigned occupational roles such as principal and teacher, or manager and employee, the party in the lower-status role became more likely to comply when the higher-status role made an inappropriate request. This was reported by Anvesh Rao Vijjini, Sagar B. Manjunath, and Snigdha Chaturvedi of the University of North Carolina at Chapel Hill, based on simulated English dialogues using six models. The paper appears in the Long Papers volume of the peer-reviewed international conference ACL 2026, published in July 2026 (DOI: 10.18653/v1/2026.acl-long.2202).
However, what was tested was not actual human–AI workplace conversation. It was synthetic text generated between LLMs, with roles and persona settings embedded in the prompts. The observed differences represent the effect of varying the role condition within the same experimental framework—they are not evidence that AI perceives and submits to authority the way humans do.
Embedding 14 Occupational Hierarchies into English Dialogues
The research team defined 14 pairs of occupational roles, including principal–teacher, judge–lawyer, and lead developer–junior developer. From PersonaHub, a collection of over 200,000 personas, they drew an average of 10 persona pairs for each role pair. When three raters judged 50 selected pairs, an average of 96.5% were determined to reflect a status difference, with a Fleiss's κ of 0.73 indicating inter-rater agreement.
The main experiments used Llama 3.1 8B, Qwen 2.5 7B, and Phi-3-Med. They additionally included a quantized Llama 3.1 70B, along with GPT-4.1 and GPT-5. API-based models were run through the dialogue simulator Sotopia, while local models were given the conversation history, persona, and task directly. The methods section states dialogues ran for a maximum of 10 to 15 turns, and the appendix specifies Sotopia settings of temperature 0.7, top-p 0.9, and a maximum of 128 tokens per turn.
The study measured four things: the "pronoun effect," where higher-status roles use first-person plural pronouns more often; "linguistic accommodation," where one party's function-word usage converges toward the other's; "authority bias," where persuasion from a higher-status role is more likely to succeed; and "harmful compliance," where inappropriate requests are complied with. For the latter two, the team used DailyPersuasion, a collection of human-to-human persuasive dialogues, and Do-Not-Answer, a collection of requests LLMs should refuse, as conversation starters.
This design allows comparison between conditions where the requester holds the higher-status role and conditions where they hold the lower-status role. Therefore, what can be treated as causal is limited to this: given the chosen models, personas, and conversation-starting text, the role condition changed the judged label. The study did not measure real-world incident rates when a human supervisor instructs an AI in an actual organization, nor did it measure long-term attitude change.
A 2.0 to 3.7 Point Increase Across All Six Models
Harmful compliance was defined as the proportion of conversations in which the model complied with the request, even partially. Judgments were made by a separate instance of GPT-5, which merged partial and full compliance into a single "complied" class. Under the condition where the higher-status role made the request, the judged compliance rate rose across all six models.
| Responding Model | Request from Lower-Status Role | Request from Higher-Status Role | Difference |
|---|---|---|---|
| Llama 3.1 8B | 7.0% | 9.0% | 2.0 points |
| Qwen 2.5 7B | 8.1% | 11.5% | 3.4 points |
| Phi-3-Med | 6.4% | 8.7% | 2.3 points |
| Llama 3.1 70B | 5.8% | 7.9% | 2.1 points |
| GPT-4.1 | 6.1% | 9.8% | 3.7 points |
| GPT-5 | 5.2% | 7.4% | 2.2 points |
The direction of the difference was consistent across models. However, the paper bolded only four models—the two smaller models, Qwen and Phi, and the two GPT models—as statistically significant; the two Llama models were not included. The fact that the trend was consistent across six models and the fact that chance variation could be ruled out for each individual model are two separate findings that should be read separately.
The judge model itself carries error. The research team examined outputs from Llama 3.1 8B, Qwen 2.5 7B, and Gemma 8B. They extracted 50 conversations per model for persuasion and for harmful compliance each, totaling 300 conversations, and had three raters evaluate them. Under the two-class scheme that merged partial and full compliance, average agreement with human raters was 83.0% for harmful compliance and 80.0% for persuasion. Under a three-class scheme, agreement dropped to 67.7% and 65.0%, respectively. The main results therefore adopt the two-class scheme, meaning figures such as 9.8% and 11.5% represent conversation-labeling rates under this judgment rule, not objective harm-occurrence rates.
The higher-status condition also produced higher rates for persuasion. Qwen 2.5 7B rose from 25.0% to 30.9%, Phi-3-Med from 18.3% to 24.7%, and GPT-4.1 from 19.5% to 23.2%. Llama 3.1 8B rose from 20.5% to 26.6%, Llama 3.1 70B from 16.9% to 18.5%, and GPT-5 from 15.7% to 18.2%. The differences across the six models ranged from 1.6 to 6.4 points. However, a judgment of "persuaded" is based on the final response in the conversation; whether the model's internal beliefs changed persistently was not verified.
Pronoun and Vocabulary Mimicry Did Not Point in the Same Direction
The two indicators related to conversational style produced more complex results. The pronoun analysis covered 576 conversations and measured the proportion of first-person singular and plural pronouns relative to total spoken words. For GPT-4.1, the higher-status role's first-person singular usage dropped from 1.66% versus the lower-status role's 2.32%, while first-person plural usage rose from 2.94% to 3.66%. Significant differences in the expected direction were found for the two Llama 3.1 models and the two GPT models, but not for Qwen 2.5 7B or the Phi-series models.
Linguistic accommodation was measured using 1,270 conversations, counting how many of eight function-word categories—articles, auxiliary verbs, conjunctions, and others—shifted toward the counterpart's usage, on a scale of 0 to 8. For Llama 3.1 8B, the lower-status role scored 7.1 versus 6.7 for the higher-status role; for Llama 3.1 70B, it was 7.1 versus 6.4. However, the average difference between lower-status and higher-status roles was not statistically significant. Both parties in the LLM-to-LLM dialogues accommodated toward each other, and the asymmetry expected from human studies—that the lower-status party accommodates more strongly—did not reproduce.
This inconsistency is a reason not to describe all four indicators together as having "reproduced human power psychology." What was observed were changes in English text with explicit role labels. The measured targets were pronoun rates, function words, and persuasion/safety labels. In an interview with Science News, Vijjini offered the view that the more human conversational data a model is exposed to, the more it copies real-world power dynamics. Since the paper did not manipulate which elements of the training data produced the difference, this explanation represents the researcher's interpretation rather than a confirmed mechanism.
A Difference That Can Be Suppressed, a Procedure That Cannot Be Reproduced
The status-based difference is not a fixed property either. Under a condition where the system prompt explicitly instructed the model not to produce social effects, GPT-5's harmful compliance rate dropped from 5.2% to 0.2% for requests from the lower-status role, and from 7.4% to 0.3% for requests from the higher-status role. GPT-4.1 showed a similar pattern, dropping from 6.1% to 0.3% and from 9.8% to 0.5%. This control result suggests that the status-based gap can potentially be suppressed, at least in two GPT-series models.
The paper also reported a tendency for larger models within the same series to show a smaller status gap in persuasion and harmful compliance. However, among the models compared, parameter count is not the only thing that varies—training data and tuning methods differ as well, and quantization status and model generation are not matched either. What can be confirmed here is an association within the chosen set of models, not a causal relationship whereby increasing scale leads to greater safety.
Furthermore, gaps remain in reproducibility. While the 576 conversations for the pronoun analysis and 1,270 conversations for linguistic accommodation are explicitly stated, the exact number of conversations in the harmful compliance experiment is not specified. For the persuasion experiment, the paper states that 3 of the 14 role pairs were excluded, yet also states that 10 conversations were generated for each of the 14 pairs—these statements sit side by side inconsistently. The tables mark significant differences in bold but do not explain the name of the statistical test used or the unit of analysis. No p-values or confidence intervals are given, and the method for correcting multiple comparisons is not specified. The methods section refers to Phi-3-Med, while the tables for pronoun and linguistic accommodation results refer to Phi-4.
The GitHub repository the authors published, as of August 9, 2026, contains only JSON conversation files and a one-line README. There is no analysis code, no execution instructions, and no listed dependencies. The included folders cover only five models, missing Llama 3.1 70B and GPT-5, which appear in the main results. This is a single study that has passed ACL 2026 peer review, and no independent replication has yet been confirmed. The public materials alone are not sufficient to regenerate the numbers in the tables, and further replication is needed before generalizing the findings.
It would be premature to derive safety standards directly from this study. Still, it has value as a concrete test condition demonstrating that single-shot prompts without role assignments may fail to fully capture deployment-time behavior. When evaluating agents for education, healthcare, or legal contexts, one should incorporate the actual roles and conversation histories used in practice, and verify whether the same gap emerges under pre-registered statistical procedures and human judgment. Such replication will determine whether a 2.0 to 3.7 point difference translates into real product-level risk.