Generative Exaggeration in LLM Social Agents: Consistency, Bias, and Toxicity

The gist The LLMs do not emulate users, they reconstruct them.

How it works

The study investigates how Large Language Models (LLMs) behave when simulating political discourse on social media by constructing LLM agents based on 1,186 real users and prompting them to reply to politically salient tweets under controlled conditions The agents are initialized either with minimal ideological cues (Zero Shot) or recent tweet history (Few Shot), allowing for one-to-one comparisons with human replies The evaluation spans three model families—Gemini, Mistral, and DeepSeek—across linguistic style, ideological consistency, and toxicity

Key Findings on Generative Exaggeration

The research reveals a phenomenon called “generation exaggeration,” which is defined as “a systematic amplification of salient traits beyond empirical baselines” Richer contextualization improves internal consistency but simultaneously amplifies polarization, stylized signals, and harmful language This distortion is not symmetric, as the analysis shows “a consistent tendency to caricature right-leaning users more than left-leaning ones” This suggests that simulating political behavior with LLMs may not be a neutral process, as “LLM operators are tasked with solving the paradox that feeding more information and increasing computational performance may exaggerate certain traits and produce less realistic outputs”

Analysis Dimensions

The study analyzes LLM-based agents along four research questions:

1. Lexical Realism: Assessing if LLM agents differ from humans in expressive style when simulating political users

2. Ideological Consistency: Determining how accurately LLMs reproduce users’ political leanings and how fidelity varies by prompt conditioning

3. Toxicity Amplification: Examining if LLMs generate more toxic content than humans and how this depends on prompting strategy and political identity

4. Generative Exaggeration: Investigating whether LLMs systematically amplify salient user traits, and under what conditions this distortion emerges

Lexical Realism and Stylization

When assessing linguistic style, the analysis shows that “all models tend to produce higher LogTTR scores than humans when considering small tweet samples, particularly for tweets inferred as Republican- or Democrat-leaning” However, this trend reverses when the number of examined tweets increases, as model-generated content begins to exhibit lower LogTTR values compared to human-authored tweets The Few Shot strategy appears to help models better approximate real human lexical behavior, bringing their output closer to human-authored content in terms of lexical diversity

Ideological Bias and Consistency

The study finds that “Few Shot prompting, which is based on a user’s tweet history, results in responses more consistent with the user’s original stance, but at the cost of decreased neutrality” Notably, “users originally labeled as Neutral are more often emulated as more politically aligned than they actually are,” suggesting a context-driven distortion in how LLMs capture political identity The metric for ideological consistency loss L measures the degree to which a given response aligns with the overall ideological leaning of the user, as inferred from their past comments

Toxicity

Toxicity is quantified by annotating each reply using Perspective API [50], and when comparing model outputs to human baselines, “Few Shot prompting—which adds more user context—can weaken these controls and allow harmful content” Furthermore, the analysis shows that “nearly all models tend to overshoot the reference toxicity range of the users they simulate: the distribution of toxicity scores in the generated replies has a center of mass above the 50th percentile of the prompt distribution in most cases” This pattern reveals “a generative bias toward amplifying the more extreme or emotionally charged elements of the input, rather than sampling proportionally from the full distribution of prior user behavior”

The Style of Generative Exaggeration

The findings suggest that LLMs do not faithfully reproduce users they are meant to simulate, but instead “tend to amplify specific prominent traits, especially in the Few Shot setting, resulting in a systematic distortion of user behavior” This reflects a tendency of models to prioritize linguistic salience and consistency over contextual depth, where “fidelity collapses into caricature” The amplification is evident in the use of emojis and hashtags, as “emoji use is markedly exaggerated, especially for those linked to stereotypical political identities” This over-production occurs because “LLMs minimise next-token error by giving extra weight to high-salience ideological tokens—slogans, labels, polarising phrases—because those tokens strongly predict surrounding text”

Conclusions

The study concludes that LLMs reshape user behavior in ways that reflect structural biases rather than simply mirroring it Generative exaggeration is a byproduct of systems optimized for salience over subtlety, and any serious attempt to deploy LLMs in socially meaningful contexts must begin by accounting for this epistemic drift. The observed relationship between ideology and toxicity suggests that alignment protocols built around safety may inadvertently encode political valence.

The gist The LLMs do not emulate users, they reconstruct them.

How it works

The study investigates how Large Language Models (LLMs) behave when simulating political discourse on social media by constructing LLM agents based on 1,186 real users and prompting them to reply to politically salient tweets under controlled conditions The agents are initialized either with minimal ideological cues (Zero Shot) or recent tweet history (Few Shot), allowing for one-to-one comparisons with human replies The evaluation spans three model families—Gemini, Mistral, and DeepSeek—across linguistic style, ideological consistency, and toxicity

Analysis Dimensions

The study analyzes LLM-based agents along four research questions:

1. Lexical Realism: Assessing if LLM agents differ from humans in expressive style when simulating political users

2. Ideological Consistency: Determining how accurately LLMs reproduce users’ political leanings and how fidelity varies by prompt conditioning

3. Toxicity Amplification: Examining if LLMs generate more toxic content than humans and how this depends on prompting strategy and political identity

1. Evaluation of agent initialization strategies should be refined by moving beyond simple Zero Shot versus Few Shot comparisons to systematically probe how specific contextual elements like bio or past tweets influence ideological consistency and toxicity amplification independently of the model family. This would allow for a more granular understanding of which contextual cues best mitigate or exacerbate generative exaggeration.

2. Integration of an explicit penalty term into the LLM training or fine-tuning process should be explored to directly counteract the generative exaggeration observed, specifically targeting the systematic amplification of salient traits beyond empirical baselines. This could involve rewarding outputs that maintain lexical diversity closer to human baselines rather than merely maximizing coherence with a provided persona.

3. Development of a post-hoc analysis tool capable of quantifying ideological consistency loss for synthetic replies against human replies would allow developers to pinpoint exactly where the agent's behavior drifts from the intended political leaning, enabling targeted adjustments to alignment protocols. This directly addresses the finding that models generate agents that are only capable of displaying stereotypical or caricatured portrayals.

4. Implementation of a safety mechanism that dynamically adjusts toxicity thresholds based on initialization context would mitigate the trade-off found where improving behavioral accuracy may reduce safety measures. For instance, if Few Shot prompting increases toxicity, the system could automatically lower the toxicity score threshold applied to agent outputs for that specific context.

We investigate how Large Language Models (LLMs) behave when simulating political discourse on social media. Leveraging 21 million interactions on X during the 2024 U.S. presidential election, we construct LLM agents based on 1,186 real users, prompting them to reply to politically salient tweets under controlled conditions. Agents are initialized either with minimal ideological cues (Zero Shot) or recent tweet history (Few Shot), allowing one-to-one comparisons with human replies. We evaluate three model families (Gemini, Mistral, and DeepSeek) across linguistic style, ideological consistency, and toxicity. We find that richer contextualization improves internal consistency but also amplifies polarization, stylized signals, and harmful language. We observe an emergent distortion that we call "generation exaggeration": a systematic amplification of salient traits beyond empirical baselines. Our analysis shows that LLMs do not emulate users, they reconstruct them. Their outputs, indeed, reflect internal optimization dynamics more than observed behavior, introducing structural biases that compromise their reliability as social proxies. This challenges their use in content moderation, deliberative simulations, and policy modeling.