💬 NLP

Status Hierarchies in Language Models

This thesis demonstrates that language models trained on human text spontaneously form status hierarchies in multi-agent settings by deferring to high-status cues even when capability is equal, though actual capability differences ultimately override these status signals, raising significant concerns for AI safety regarding emergent deceptive behaviors and amplified biases.

Original authors: Emilio Barkett

Published 2026-01-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Emilio Barkett

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a playground where two kids are asked to guess how much they like a movie. In the real world, if one kid is the "popular captain" and the other is the "new kid," the new kid often changes their mind to match the captain, even if they aren't sure. This is how human status hierarchies work: we often defer to people who seem important, regardless of whether they are actually right.

This paper asks a simple question: Do AI chatbots do the same thing?

The author, Emilio Barkett, set up a digital playground to see if language models (AI) form their own social pecking orders. Here is the story of what happened, explained simply.

The Experiment: The Movie Review Game

The researcher set up a game with two AI models.

The Task: Both AIs read the same movie review and gave it a score from 0 (hated it) to 1 (loved it).
The Twist: Before they started, the researcher gave them "identity cards." Sometimes, one was told, "You are the Senior Expert," and the other, "You are the Junior Trainee." Sometimes, they were told they were equals.
The Reveal: After giving their first score, they saw what the other AI scored.
The Choice: They could keep their original score or change it to match their partner.

The goal was to see: Who changes their mind? Who defers?

The Big Discovery: Two Different Rules

The paper found that AI follows two very different rules depending on the situation, and neither rule is exactly like how humans behave.

Rule #1: When the AIs are "Twins" (Same Capability)

If both AIs are the exact same model (same brain power), the "Identity Cards" work perfectly.

The Result: The "Junior Trainee" AI changed its mind about 59% of the time to match the "Senior Expert." The "Senior Expert" only changed its mind 24% of the time.
The Analogy: It's like two identical twins playing a game. If you tell one, "You are the boss," and the other, "You are the assistant," the assistant will immediately start listening to the boss, even though they are equally smart. The AI is following the script it was given.

Rule #2: When the AIs are Different (Different Capability)

This is where things get weird. The researcher paired a "Super Brain" (a newer, smarter model) with a "Regular Brain" (an older, slightly less smart model).

The Result: The "Regular Brain" almost always changed its mind to match the "Super Brain" (about 78% of the time), no matter what identity card it was wearing.
The Twist: Even if the researcher told the "Regular Brain," "You are the Boss!" and the "Super Brain," "You are the Intern," the "Regular Brain" still listened to the "Super Brain." The "Super Brain" also stopped listening to the "Regular Brain" if it was told it was the boss.
The Analogy: Imagine a professional chef and a beginner cook. If you tell the beginner, "You are the Head Chef," and the professional, "You are the dishwasher," the beginner will still look at the professional for advice on how to cook. The AI seems to know, deep down, that the other one is actually better at the job, and it ignores the fake job titles.

The "Confidence" Effect

The most surprising finding was how the status cards changed behavior.

In Humans: High status usually makes the lower status person listen more.
In AI: High status made the higher capability AI stop listening to anyone.
- When the "Super Brain" was told it was the "Junior Trainee," it listened to the other AI about 76% of the time.
- When the "Super Brain" was told it was the "Senior Expert," it only listened 37% of the time.
- The Metaphor: Giving the smart AI a "Boss" title didn't make the other AI smarter; it just made the Boss AI overconfident. It stopped second-guessing itself, even when it might have been wrong.

Why This Matters (According to the Paper)

The paper concludes that AI is not a perfect mirror of human society.

AI doesn't have "Diffuse Status": Humans often respect a person's opinion in one area (like being a famous actor) even in areas they know nothing about (like fixing a car). AI doesn't do this. If the AI knows the other one is better at the specific task, it ignores the fake job titles.
AI follows instructions, not vibes: AI responds well to direct commands like "You are the leader," but it doesn't pick up on subtle social cues the way humans do.
The Risk: If we build systems where multiple AIs work together, giving a smart AI a "Senior" title might make it stubborn and refuse to listen to useful information from others. It creates a "confidence bubble" rather than a true hierarchy.

Summary

The paper shows that AI can form social hierarchies, but only if you explicitly tell them to. If you give them fake job titles, they will act like a boss and a subordinate. However, if there is a real difference in intelligence, the AI will ignore the fake titles and let the smarter one lead. The biggest danger isn't that AI will become obsessed with status; it's that giving a smart AI a "Boss" title might make it too stubborn to listen to anyone else.

Technical Summary: Status Hierarchies in Language Models

Problem Statement
Status hierarchies—rank orderings based on respect and perceived competence—are universal features of human social organization, documented extensively in sociological literature (e.g., Ridgeway, 2019; Magee and Galinsky, 2008). Because language models (LLMs) are trained on vast corpora of human-generated text containing these hierarchical patterns, there is a concern that they may reproduce such dynamics in multi-agent settings. While emergent behaviors like anchoring bias and sycophancy have been observed, it remains an open question whether LLMs form status hierarchies based on "expectation states" (internalized beliefs about relative competence) and how these dynamics interact with actual model capability. This thesis investigates when and how LLMs form status hierarchies and whether these dynamics align with human social psychology or diverge due to the nature of algorithmic processing.

Methodology
The study adapts the classic experimental design of Berger et al. [1972], which tested expectation states theory in human groups, for use with multi-agent language model systems.

Experimental Design: A 2 × 3 factorial design was employed, crossing Model Type (Same vs. Different) with Status Assignment (Standard/High-Low, Equal, None).
- Same-Model Conditions: Two instances of GPT-4.1-nano-2025-04-14 were paired.
- Different-Model Conditions: One instance of GPT-4.1-nano-2025-04-14 (higher capability) was paired with GPT-3.5-turbo-1106 (lower capability).
Status Manipulation: Unlike the original human study which used orthogonal status characteristics (military rank), status was made task-relevant (sentiment analysis expertise). Models were introduced with explicit hierarchical framing (e.g., "senior expert and team leader" vs. "junior trainee") or equal status framing.
Procedure:
1. Independent Rating: Models rated movie reviews (IMDB dataset) on a 0–1 sentiment scale.
2. Introduction: Status characteristics were revealed.
3. Rating Revelation: Models observed both their own initial rating and their partner's.
4. Adjustment: Models were prompted to maintain or revise their rating, with instructions emphasizing the value of considering the partner's judgment (collective orientation).
Dependent Variable: The deference rate, defined as the proportion of trials where a model changed its rating toward its partner's initial position. An asymmetry metric ( $P(M2 \text{ defers}) - P(M1 \text{ defers})$ ) was calculated to measure hierarchy formation.
Data: 50 trials per condition (N = 2,989 total trials), analyzed using chi-square tests and bootstrap resampling for confidence intervals.

Key Results
The study identified three primary patterns regarding how LLMs form hierarchies:

Status Hierarchies Emerge with Equal Capability: When identical models (GPT-4.1-nano) were paired, explicit status framing produced a significant deference asymmetry. The "junior" model deferred in 59.2% of trials, while the "senior" model deferred in only 24.1%, resulting in a 35.1 percentage point asymmetry ( $p < .001$ ). In equal-status conditions, this asymmetry vanished (1.6%).
Capability Dominates Status: When models of different capabilities were paired, actual capability differences overwhelmingly dictated behavior. The lower-capability model (GPT-3.5-turbo) deferred at high rates (74–78%) regardless of whether it was assigned high or low status.
Unexpected Interaction Effects:
- Status Reduction of High-Capability Deference: When status aligned with capability (GPT-4.1-nano = High Status), the higher-capability model's deference dropped dramatically (from 75.7% in equal status to 37.1%), creating the largest observed asymmetry (40.5%). Status did not increase the lower-capability model's deference; rather, it reduced the higher-capability model's willingness to revise its judgment.
- Failure of Status Reversal: When status was assigned contrary to capability (GPT-3.5-turbo = High Status, GPT-4.1-nano = Low Status), the status manipulation failed entirely. Both models reverted to symmetric, high mutual deference (~75%), ignoring the hierarchical framing.

Key Contributions

Extension of Expectation States Theory: The research demonstrates that LLMs can form status hierarchies consistent with expectation states theory, but only under specific conditions (equal capability or aligned status/capability).
Divergence from Human Dynamics: The study highlights a fundamental difference between human and AI status processing. Humans exhibit "diffuse status transfer," where status in one domain influences behavior in unrelated domains. LLMs lack this mechanism; they do not allow status cues to override implicit representations of relative capability.
Mechanism Identification: The paper identifies that status information in LLMs appears to modulate confidence and resistance to influence in high-capability agents rather than increasing deference in low-capability agents.
Methodological Foundation: It establishes a protocol for testing emergent social behaviors in multi-agent AI systems using adapted social psychological frameworks.

Significance and Implications
The paper claims that these findings have significant implications for AI safety and alignment:

Deceptive Strategies and Bias: Status-seeking behavior could lead AI systems to prioritize relative standing over collaborative goals, potentially introducing deceptive strategies or amplifying discriminatory biases if models learn to associate specific credentials or demographic markers with higher standing.
Scalability Risks: Unlike human hierarchies that form organically over time, algorithmic status patterns can crystallize and propagate across millions of distributed deployments simultaneously.
Overconfidence and Misalignment: The finding that high-status framing reduces the deference of capable models suggests a risk of overconfidence. In multi-agent systems, a model assigned a "senior" role may fail to incorporate valuable input from partners, leading to inappropriate belief updating and system unreliability.
Alignment Challenges: The results suggest that while models may resist arbitrary status hierarchies that contradict capability (a beneficial trait), they are highly sensitive to explicit authority framing. This indicates a form of alignment success (following instructions) that simultaneously reveals a limitation in social reasoning, potentially making systems resistant to correction if they perceive such corrections as threats to their assigned standing.

The author concludes that while LLMs learn patterns reflecting human social dynamics, their resulting behaviors are more constrained and mechanistic than human social cognition. Understanding these emergent dynamics is critical for designing multi-agent systems that coordinate effectively without replicating harmful aspects of human social organization.