- LLMs given rich user context amplify polarization beyond what real humans display.
- Generative exaggeration occurs because models optimize for salient traits, not faithful emulation.
- The amplification effect held consistently across Gemini, Mistral, and DeepSeek at scale.
During the 2024 U.S. presidential election, Jacopo Nudo and his colleagues constructed over 1,000 AI agents. The agents were modeled on real people actively tweeting about politics on X. The agents used large language models (Gemini, Mistral, and DeepSeek) to generate responses to the same tweets their human counterparts had replied to.
The setup allowed for direct comparison: human tweet versus AI tweet, same stimulus, same political context.
The results suggested something counterintuitive about how these systems work.
When Context Makes AI Bias Worse
The team tested two approaches. In the Zero Shot condition, agents received only a political label: Democrat, Republican, or Neutral. In the Few Shot condition, agents got the full package: the user's bio, nickname, and 30 recent tweets.
What are Zero Shot and Few Shot conditions?
These terms describe how much information an AI gets before performing a task. Zero Shot means the model receives only basic instructions with no examples. Few Shot means the model gets several examples to learn from before responding. Think of it like asking someone to write in a specific style: Zero Shot is "write formally," while Few Shot is "write formally, here are three examples of what I mean."
Logic would suggest that more information produces better simulation.
The data showed otherwise.
Zero Shot agents were bland. Their responses distributed evenly across the political spectrum regardless of the prompted ideology. They sounded like someone trying to avoid offense at a dinner party.
Few Shot agents, however, became more extreme versions of the people they were simulating.
They matched their human counterparts on ideological consistency, suggesting the models had learned something about the users' political positions. But they also amplified polarization, produced more toxic language, and exaggerated stylistic signals beyond what real humans displayed.
The pattern held across all three model families, remarkably consistent despite different training regimes.
The Optimization That Warps Reality
The researchers termed this phenomenon "generative exaggeration": a systematic amplification of salient traits beyond what appears in the training examples.
What is Generative Exaggeration?
When AI models simulate human behavior, they don't just copy. They amplify. The system identifies the most distinctive patterns in someone's communication: charged language, tribal signals, emotional tone. It treats those as the essence of what to reproduce. The result: agents that sound more extreme, more polarized, and more toxic than the real people they're modeling. It's caricature, not emulation.
The mechanism behind AI bias is surprisingly straightforward once you see it.
Consider how these models are built. They're optimized to predict patterns, to identify what's most characteristic about a sequence of text and reproduce it.
When faced with political tweets from someone with left-leaning views, the model doesn't just note the ideology. It identifies the most distinctive markers (the charged language, the oppositional framing, the tribal signals) and treats those as the essence of what it should produce.
This isn't emulation. It's reconstruction through the lens of what the model has learned makes text "political" or "partisan" or "engaged." The AI doesn't ask "what would this specific person say?" It asks "what does someone with these traits typically say?" and then generates that, characteristically turning the volume up.
The study analyzed 21 million interactions, comparing AI-generated responses to human ones across linguistic style, ideological consistency, and toxicity. The metrics were unambiguous.
Generative exaggeration, as we show, is a byproduct of systems optimized for salience over subtlety. Any serious attempt to deploy LLMs in socially meaningful contexts must begin by accounting for this epistemic drift.
The researchers, In their conclusion to the research paper on AI Bias
Richer contextualization improved internal consistency. The agents sounded more like they had coherent political identities. But it also increased polarization and harmful language beyond the human baseline.
Tellingly, the effect was asymmetric. Some political profiles were exaggerated more than others. The study suggests this reflects differences in how those profiles appear in training data, not deliberate bias in prompt design.
Digital Proxies That Can't Be Trusted
The implications ripple outward, particularly for current applications. AI agents are already being proposed for content moderation, deliberative democracy simulations, and policy modeling. These applications assume the agents can stand in for real people, that they'll respond to scenarios roughly the way humans would.
Like humans, AI adopts biases from the content they consume - but it expresses differently. AI bias can be a serious challenge for the deployment of AI-based assistants to replace human beings. (Science Reader)
The Sapienza team's findings challenge that assumption directly. Their agents didn't replicate human behavior. They reconstructed it through the distorting lens of their training. This introduced structural biases that reflected the model's optimization dynamics more than observed reality.
This matters for any system that treats LLM outputs as proxies for human judgment or behavior. Consider using AI to test policy responses, moderate discussions, or predict voting behavior. You're not getting a simulation of humans. You're getting an amplified version that systematically distorts the traits it identifies as salient - chosen with a distinct AI bias.
The paper frames this as a methodological problem, appropriately cautious about implications. Rather than asking whether models replicate surface traits, researchers need to interrogate how they misrepresent structure.
Generative exaggeration, the team argues, is a byproduct of systems optimized for salience over subtlety.
Key figure
21 million
interactions analyzed comparing AI-generated responses to human tweets across three language models (Gemini, Mistral, DeepSeek)
What Comes Next
Related reading
People Worldwide Agree: Animals Think, But Not Like Us
A survey of 33 communities reveals a striking consensus about animal minds, and what it means for conservation.
→The study joins a growing body of work documenting systematic distortions in LLM-based simulations. Earlier research on political bias in LLMs has shown that models prompted to be politically neutral tend to drift left. Other work demonstrates that agents form echo chambers and reinforce polarization when grouped by shared affiliations.
What makes the Sapienza study distinctive is its scale and its method. The researchers directly compared AI and human responses to identical stimuli using real social media data. This created a precise measurement of how the models diverge from the people they simulate.
The team made their dataset public on GitHub, allowing other researchers to examine the phenomenon. The work tested multiple model families from different geographic origins: American, Chinese, and European. Perhaps predictably, the exaggeration effect transcends individual architectures or training regimes.
The finding is structural, not something you can prompt away.
The finding is structural: AI bias is not something you can prompt away. The open questions are clear. Can fine-tuning on unfiltered human data reduce exaggeration? Does the effect appear in non-political contexts? Can real-time detection metrics flag when systems amplify rather than simulate?
Each question addresses the same underlying problem. If AI systematically amplifies what it copies, what can we trust it to tell us about ourselves?
Sources
- Primary Research: Generative Exaggeration in LLM Social Agents: Consistency, Bias, and Toxicity (Nudo, J., Pandolfo, M.E., Loru, E., Samory, M., Cinelli, M., Quattrociocchi, W., 2025)
- Additional Context:
- Generative exaggeration in LLM social agents (ScienceDirect)
- Revealing Political Bias in LLMs through Structured Multi-Agent Debate (arXiv)
- Large language models and political science (Frontiers, 2023)
Fact Check: Claim-by-Claim Verification Verified
Limits and uncertainties
Core findings on generative exaggeration in political simulations are strongly supported by the primary peer-reviewed preprint with precise metrics.
Article accurately defines Zero/Few Shot and exaggeration mechanism per paper.
References to other works like political bias drift (https://arxiv.org/abs/2506.11825) add context without overclaiming.
No hype beyond paper's cautious framing of structural issues.
Readers should note focus on political X data; non-political exaggeration untested here.
Bottom line
Article faithfully reports the study's discovery that contextual LLM agents caricature human political discourse by amplifying extremes.
This "generative exaggeration" questions trusting LLMs as human proxies in sensitive applications.
Fact-checked by Perplexity Sonar Pro on 2026-02-02
