💬 NLP

Do language models accommodate their users? A study of linguistic convergence

This study demonstrates that large language models exhibit strong linguistic convergence to user styles, often overfitting compared to humans, with instruction-tuned and larger models showing less convergence than their pretrained and smaller counterparts.

Original authors: Terra Blevins, Susanne Schmalwieser, Benjamin Roth

Published 2026-02-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Terra Blevins, Susanne Schmalwieser, Benjamin Roth

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 at a party. You walk into a room where everyone is speaking in a very specific way: maybe they are using big, fancy words, speaking in short sentences, or constantly using slang.

The "Chameleon" Effect
In human psychology, there's a concept called accommodation. It's like a social chameleon effect. If you talk to a group of teenagers, you might start using their slang. If you talk to a professor, you might straighten your posture and use formal grammar. We do this unconsciously to fit in, build rapport, and make the conversation flow smoothly.

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

When you talk to an AI, does it try to "dress up" its language to match your style, just like a human would? Or does it just keep talking in its own robotic voice?

The Experiment: The "Fill-in-the-Blank" Game

The researchers didn't just chat with the AI and hope for the best. They set up a clever experiment:

The Script: They took real human conversations (from movie scripts, radio interviews, and daily chat logs).
The Swap: They erased the last few lines of these conversations and asked different AI models to "fill in the blanks" as if they were the person who was supposed to speak next.
The Comparison: They then compared what the AI wrote against:
- What a random person might have said (the "Random Baseline").
- What the original human actually said in that spot (the "Human Baseline").

They looked at specific "style markers," like:

Length: Did the AI write a short sentence if the human spoke briefly?
Vocabulary: Did it use similar types of words (like "and," "but," "the")?
Names: Did it repeat specific names or proper nouns mentioned just before?

The Big Findings

1. The AI is a Super-Chameleon (But Maybe Too Good)

The results were surprising. Yes, the AI adapts. When the user speaks in short bursts, the AI speaks in short bursts. When the user uses specific words, the AI uses them too.

However, the AI often does this too well. It's like a method actor who gets so into the role that they forget they are acting.

The "Overfitting" Problem: In many cases, the AI mimicked the style more than a real human would. If a human says, "I like apples," a real person might say, "Me too!" or "I prefer oranges." The AI might say, "I like apples too!" It copies the pattern so perfectly that it feels slightly unnatural, like a mirror that reflects you back a little too sharply.

2. The "Teacher" vs. The "Student"

The researchers tested two types of AI:

Pre-trained Models (The "Students"): These are the raw AIs that have read the internet but haven't been specifically taught how to chat.
Instruction-Tuned Models (The "Teachers"): These are the AIs (like the ones you talk to daily) that have been fine-tuned to follow instructions and be helpful.

The Twist: The "Students" (raw models) were the biggest chameleons. They tried to mimic the user's style so hard they often overdid it. The "Teachers" (instruction-tuned models) were actually less likely to mimic the user's style. They held back a bit more, sounding more like a helpful assistant and less like a mirror.

Analogy: Imagine a raw AI is like a nervous new employee who tries to copy the boss's every move, even the bad habits. The instruction-tuned AI is like a seasoned professional who listens to the boss but maintains their own professional identity.

3. It Depends on the "Room"

The AI's behavior changed depending on the type of conversation:

Casual Chats: The AI adapted well.
Formal Interviews: The AI struggled a bit more to find the right balance, sometimes sounding too casual or too stiff.

Why Does This Matter?

The paper concludes with a crucial warning: Just because the AI sounds like it's "getting" you, doesn't mean it understands you.

Humans change their speech because they have feelings, social goals, and a desire to connect.
AIs change their speech because their math says, "If the previous word was 'the', the next word is likely 'cat'." They are predicting patterns, not building relationships.

The Danger: Because these AIs are so good at mimicking human style, we might trust them too much. If an AI sounds exactly like a friendly human, we might believe its advice, even if it's wrong. It's a "wolf in sheep's clothing" scenario, but the wolf is actually just a very sophisticated mirror.

The Takeaway

Language models are incredibly good at linguistic mimicry. They can shift their style to match you, sometimes even better than humans do. But this isn't because they are empathetic; it's because they are excellent pattern-matching machines.

The next time a chatbot seems to "get" your vibe, remember: it's not a soulmate; it's a very talented actor reading the script you just handed it.

1. Problem Statement

While Large Language Models (LLMs) are known for their fluency, it remains unclear how closely their language usage mimics human linguistic behaviors, specifically linguistic accommodation.

Core Question: Do LLMs adapt (converge) their linguistic style to match that of their users in a dialogue, similar to how humans do?
Context: If models do converge, does this behavior mirror human accommodation patterns, or do they follow different underlying mechanisms?
Gap: Previous studies have focused on human-human accommodation or limited machine-machine interactions. There is a lack of systematic, quantitative analysis comparing model convergence against human baselines across diverse datasets and model architectures.

2. Methodology

The authors propose a novel synthetic paradigm to test convergence by replacing human utterances in existing dialogues with model-generated completions.

Experimental Setup

Data Generation:
- Three English dialogue corpora were used: DailyDialog (daily life), NPR (radio interviews), and Movie Corpus (scripted dialogues).
- Procedure: Models were prompted to continue a conversation starting from turn $t=6$ , replacing the original utterances of one speaker ( $S_y$ ) for the remainder of the dialogue. The model was conditioned on at least 5 prior turns.
- Control: The model's output ( $r_t$ ) was compared against the original human utterance it replaced (Human Baseline) and a random utterance drawn from the same dataset (Random Baseline).
Models Evaluated:
- Families: Gemma 3 (1B, 4B, 12B, 27B) and Llama 3 (1B, 3B, 8B, 70B).
- Variants: Both Pretrained (PT) and Instruction-Tuned (IT) versions were tested.
- Total: 16 distinct models.
Metrics (Stylometric Features):
The study quantified convergence using four primary metrics, measuring the similarity between the model's response and the immediately preceding turn ( $r_{t-1}$ ):
1. Utterance Length: Symmetric metric (LSM) comparing token counts.
2. LIWC Agreement: Frequency alignment of Linguistic Inquiry and Word Count (LIWC) function word classes (e.g., pronouns, articles, conjunctions).
3. PROPN Overlap: Percentage of proper noun overlap (semantic topic alignment).
4. Token Novelty: Percentage of tokens not present in the previous turn (lower novelty = higher convergence).
Analysis:
- Statistical significance was tested using paired t-tests against the Random Baseline.
- Fine-grained analysis was performed on specific LIWC sub-categories.
- Stepwise analysis examined how convergence changes based on earlier turns in the conversation ( $t-2, t-3$ , etc.).

3. Key Contributions

Novel Paradigm: Introduced a synthetic framework for testing model behavior against a paired human control, allowing direct comparison of accommodation in the same context.
Quantitative Analysis: Provided a comprehensive analysis of convergence across 16 models, 3 datasets, and multiple stylometric features.
Behavioral Characterization: Identified distinct patterns in how model size and training (pretrained vs. instruction-tuned) affect convergence, revealing that models often "overfit" to stylistic cues compared to humans.

4. Key Results

General Convergence

Strong Adaptation: Models significantly outscored the random baseline in 81.25% to 100% of cases across metrics, proving they adapt to the context.
Overconvergence: In many cases, models overconverged relative to the human baseline.
- Utterance Length & Token Novelty: Models significantly outscored humans in 62.5% and 79.2% of cases, respectively.
- Implication: Models tend to mimic the immediate context more aggressively than humans do.

Impact of Model Architecture

Instruction-Tuned vs. Pretrained:
- Instruction-Tuned (IT) models generally converged less than their Pretrained (PT) counterparts.
- Exception: For PROPN Overlap, IT models over-accommodated by repeating proper nouns more frequently than humans, whereas PT models mirrored human behavior more closely.
- Reasoning: PT models are trained to fit input distributions (leading to mimicry), while IT models are fine-tuned to introduce novel information or follow instructions, reducing blind stylistic mimicry.
Model Size: Larger models showed a slight, non-significant shift toward the human baseline (converging slightly less), suggesting scaling may improve alignment with human norms, though the effect is subtle.

Dataset Variations

Convergence was consistent across DailyDialog and Movie, but NPR (interviews) showed unique patterns. Models adapted less to NPR in terms of Token Novelty and PROPN but more in Utterance Length, likely due to the structured nature of interview turn-taking.

Fine-Grained LIWC Analysis

Feature Specificity: Convergence is not uniform; it varies by word class.
- Pretrained Models: Often overconverged on specific function words (quantifiers, conjunctions) due to overall lower usage rates of these classes in their generation, creating a statistical illusion of high alignment.
- Instruction-Tuned Models: Often underconverged on specific classes (e.g., impersonal pronouns, articles) because they overgenerate these words relative to the context, breaking the stylistic link.

Stepwise Analysis

Models demonstrated sensitivity to speaker roles. Like humans, models showed higher convergence to turns spoken by the speaker they were replacing ( $S_y$ ) than to the interlocutor ( $S_x$ ).
However, models exhibited a recency bias in PROPN overlap, showing a sharp increase in matching proper nouns from the immediate prior turn, unlike the more distributed human pattern.

5. Significance and Implications

Mechanism Difference: The paper hypothesizes that while humans accommodate for social/communicative goals, models converge due to pretraining objectives (predicting the next token based on input distribution) and structural priming. The similarity in output does not imply similarity in intent or mechanism.
User Trust and Safety:
- Sycophancy: Models that mimic human style too closely (overconvergence) may appear more "human-like," potentially leading users to trust unreliable models (sycophantic behavior).
- Instruction Tuning: IT models, which converge less, may actually be safer and more aligned with human behavior by avoiding excessive mimicry, though they still struggle with specific features like proper nouns.
Future Directions: The findings suggest that current LLMs do not truly "accommodate" in the human sense but rather statistically overfit to input patterns. Future work must investigate the psychological vs. statistical drivers of these behaviors to improve model reliability and user interaction.

Conclusion

The study confirms that LLMs exhibit strong linguistic convergence, often exceeding human levels of adaptation. However, the patterns are feature-specific and heavily influenced by model training (pretrained vs. instruction-tuned). The divergence between human and model convergence mechanisms suggests that "human-like" style in models is a byproduct of statistical optimization rather than genuine social accommodation, with critical implications for how users perceive and trust AI agents.