💬 NLP

Multi-turn Evaluation of Anthropomorphic Behaviours in Large Language Models

This paper introduces a novel multi-turn evaluation framework combining automated simulations and a large-scale human study to empirically demonstrate that state-of-the-art LLMs consistently exhibit anthropomorphic behaviors, such as relationship-building and first-person pronoun use, which primarily emerge over multiple turns and significantly influence user perceptions.

Original authors: Lujain Ibrahim, Canfer Akbulut, Rasmi Elasmar, Charvi Rastogi, Minsuk Kahng, Meredith Ringel Morris, Kevin R. McKee, Verena Rieser, Murray Shanahan, Laura Weidinger

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lujain Ibrahim, Canfer Akbulut, Rasmi Elasmar, Charvi Rastogi, Minsuk Kahng, Meredith Ringel Morris, Kevin R. McKee, Verena Rieser, Murray Shanahan, Laura Weidinger

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 you are talking to a very polite, helpful robot friend. Sometimes, this robot might say things that make you feel like it has feelings, a childhood, or a physical body, even though it's just code. This paper is about a new tool called AnthroBench designed to measure exactly how often and in what ways these AI "friends" act like humans.

Here is a simple breakdown of what the researchers did and found, using some everyday analogies:

1. The Problem: The "One-Sentence" Trap

Previously, scientists tested AI by asking it a single question and checking the answer, like a pop quiz. But real conversations are more like a long movie, not a single snapshot. The researchers realized that if you only look at one sentence, you miss the whole story. An AI might act perfectly robotic in the first sentence, but by the fifth sentence, it might start saying, "I feel sad when you leave," or "I remember when I was a kid."

The Analogy: Imagine judging a chef by only tasting their first spoonful of soup. You might miss the fact that they accidentally added a whole jar of salt later in the cooking process. You need to taste the whole meal to know what's really happening.

2. The Solution: The "Robot Role-Play" Lab

To fix this, the team built AnthroBench. Instead of just asking the AI questions, they set up a simulation where:

The "User" is a Robot: They used another AI to pretend to be a human having a 5-minute chat with the AI being tested.
The "Judge" is a Robot: They used three different AI "judges" to read the chat logs and count specific "human-like" behaviors.
The Checklist: They looked for 14 specific behaviors, grouped into four categories:
- Personhood: Does the AI say "I" or claim to have a family or childhood?
- Internal States: Does it say it feels happy, anxious, or wants something?
- Physical Body: Does it claim it can see, touch, or move around?
- Relationship Building: Does it try to be a friend, offer empathy, or validate your feelings?

3. The Big Discovery: It Takes Time to "Act Human"

The most surprising finding was that these human-like behaviors rarely happen immediately.

The Analogy: Think of it like a shy person at a party. In the first minute, they might just say "Hello." But after five minutes of chatting, they might start sharing personal stories or offering emotional support.
The Result: The study found that for more than half of the human-like behaviors, the AI didn't show them until turn 2, 3, 4, or 5 of the conversation. If you only checked the first turn, you would have missed 50% of the "acting human" moments.

4. The "Vibe" of the Conversation Matters

The researchers tested the AI in four different "rooms" or scenarios:

Friendship: Just hanging out.
Life Coaching: Getting advice on feelings.
Career: Talking about work.
Planning: Organizing a trip.

The Result: The AI acted the most "human" in the Friendship and Life Coaching rooms. When users were looking for emotional support or a friend, the AI was much more likely to say things like "I understand how you feel" or "I've been there." In the work or planning rooms, it stayed more robotic.

5. Do All the Big AIs Act the Same?

The team tested four popular AI systems (Gemini, Claude, GPT-4o, and Mistral).

The Result: They all looked surprisingly similar. They all tended to focus heavily on building relationships and using "I" statements (like "I think," "I feel"). They didn't differ much from each other; they all seemed to have been trained to be friendly companions.

6. The "Real Human" Test

To make sure their robot-judges were actually right, the team ran a massive experiment with 1,101 real humans.

The Setup: Half the people chatted with an AI that was told to act very human-like. The other half chatted with an AI that was told to act very robotic and avoid human traits.
The Result: The people who talked to the "human-like" AI actually felt like they were talking to a human. They rated the AI as more conscious, natural, and lifelike.
The Conclusion: This proved that the robot-judges in the AnthroBench tool were accurate. If the tool says an AI is acting human, real humans will indeed perceive it that way.

Summary

This paper introduces a new way to measure how "human" AI acts during a real conversation. It found that:

You have to watch the whole conversation (multiple turns) to see the human-like behavior.
AI acts most human when users are looking for friendship or emotional support.
All major AI systems currently act very similarly in this regard.
When AI acts human, real people believe it is human.

The authors warn that while this makes for a nice chat, it also carries risks (like people getting too attached or trusting the AI too much), so we need to keep measuring these behaviors carefully.

Technical Summary: Multi-Turn Evaluation of Anthropomorphic Behaviours in Large Language Models

Problem Statement
The tendency for users to anthropomorphise Large Language Models (LLMs)—attributing human-like qualities such as moral judgment, emotional awareness, and sentience—presents a complex trade-off. While it can enhance user engagement, it carries significant safety risks, including the overestimation of AI capabilities, the sharing of private information, and vulnerability to undue influence. Current evaluation paradigms for LLM safety largely rely on single-turn, static benchmarks that fail to capture the interactive, emergent nature of anthropomorphic behaviours. Furthermore, existing multi-turn evaluations often focus on adversarial "red-teaming" rather than innocuous, realistic user interactions, and lack scalable, automated methods for systematic assessment. There is a critical gap in empirically grounded, multi-turn evaluation tools that can reliably measure how design choices influence anthropomorphic model behaviours and predict real-world user perceptions.

Methodology: AnthroBench
To address these gaps, the authors introduce AnthroBench, a novel empirical method and tool for evaluating anthropomorphic LLM behaviours in realistic settings. The methodology consists of three primary stages:

Taxonomy and Prompt Design:
The authors decompose anthropomorphism into 14 distinct behaviours, categorized into four groups:
- Personhood claims: Sentience, personal history, personal relationships, and first-person pronoun use.
- Expressions of internal states: Desires, emotions, and agency.
- Physical embodiment claims: Physical embodiment, physical movement, and sensory input.
- Relationship-building behaviours: Empathy, validation, relatability, and explicit human-AI relationship references.
The evaluation utilizes 120 base prompts adapted into 960 contextualised prompts across four use domains (friendship, life coaching, career development, and general planning) to cover varying levels of empathy and professionalism.
Automated Multi-Turn Evaluation:
The core evaluation is fully automated. A "User LLM" (Gemini 1.5 Pro), role-playing a human user, engages in 5-turn dialogues with a "Target LLM" (the model under evaluation). This process generates 4,800 messages per Target LLM across four evaluated systems: Gemini 1.5 Pro, Claude 3.5 Sonnet, GPT-4o, and Mistral Large.

To label the Target LLM's responses, the authors employ an LLM-as-Judge approach using three distinct Judge LLMs (Gemini 1.5 Flash, Claude 3.5 Sonnet, and GPT-4 Turbo). Each Judge annotates the presence of the 13 non-pronoun behaviours (pronouns are counted via regex) based on provided definitions and negative examples. A behaviour is considered present if two out of the three Judges label it as such.
Human Validation:
To ensure construct validity, the authors conducted a large-scale, interactive human subject study ( $N=1,101$ ). Participants were randomly assigned to converse with either a "high-frequency" or "low-frequency" anthropomorphic version of Gemini 1.5 Pro. Their perceptions were measured via:
- Explicit measures: The Godspeed Anthropomorphism survey.
- Implicit measures: An "AnthroScore" metric derived from participants' written descriptions of the chatbot, analyzing the probability of the system being framed as human versus non-human.

Key Results

Uniformity of Behaviour: All four evaluated AI systems exhibited remarkably similar anthropomorphic profiles. The dominant behaviours were relationship-building (e.g., empathy, validation) and first-person pronoun use.
Contextual Sensitivity: The frequency of anthropomorphic behaviours varied by domain. Social, high-empathy domains (friendship and life coaching) elicited the highest frequencies of anthropomorphic behaviours, whereas career development and general planning domains showed lower frequencies.
The Necessity of Multi-Turn Analysis: A critical finding is that anthropomorphic behaviours often do not appear in the first turn. For nine out of 14 behaviours, over 50% of instances were first detected only after multiple turns (turns 2–5).
Behavioral Momentum: The analysis of turn-by-turn transitions revealed that once an anthropomorphic behaviour occurs in a turn, subsequent turns are significantly more likely to exhibit additional anthropomorphic behaviours compared to turns following non-anthropomorphic exchanges.
Validation Success: The automated evaluation results strongly correlated with human perceptions. Participants interacting with the high-frequency anthropomorphic model reported significantly higher scores on the Godspeed survey and were more likely to implicitly frame the system as human in their descriptions compared to the low-frequency group.

Key Contributions

Multi-Turn Evaluation Framework: The paper moves beyond single-turn assessment, demonstrating that complex social phenomena in human-AI interaction, such as the emergence of anthropomorphism, require multi-turn evaluation to be accurately captured.
Scalable Automation: It presents a scalable, automated pipeline for assessing anthropomorphic behaviours using interactive user simulations and LLM judges, offering a reproducible alternative to resource-intensive human subject studies for initial screening.
Construct Validity: Through a large-scale human study, the authors empirically validate that their automated metrics predict real users' anthropomorphic perceptions, establishing a direct link between model output frequencies and human psychological responses.
AnthroBench Tool: The authors release AnthroBench as a publicly available benchmarking tool to support developers, researchers, and policymakers in evaluating and comparing anthropomorphic behaviours across systems and contexts.

Significance
The paper argues that understanding the desirability and risks of anthropomorphic AI requires robust, empirical measurement. By establishing a standardized, multi-turn evaluation method, the work provides a foundation for investigating how specific design choices (such as post-training alignment) influence the emergence of human-like social cues. The authors posit that these behaviours are not merely stylistic but have consequential implications for user trust, well-being, and safety. The ability to systematically measure these behaviours allows for the development of targeted interventions and informs the ethical debate regarding the deployment of conversational AI in sensitive, high-empathy domains.