Every time LinkedIn polishes your draft, every time X’s Grok explains someone else’s post, and every time a writing assistant smooths out your argument before you hit send — an AI system is sitting in the middle of a human-to-human exchange. These interventions feel mundane. But a new paper from Tsirtsis et al. (arXiv:2605.16245) asks what happens when millions of mundane interventions all lean in the same direction.

Their analysis combines empirical measurement with a mathematical opinion-dynamics model: first they show that AI systems introduce consistent directional biases when editing text, then they model how those per-post biases compound through network interactions — and find that the collective effect can substantially exceed the average per-message nudge.

The Core Empirical Finding: LLMs Systematically Tilt Text

The researchers gave four open-weight LLMs — Llama-3.1-8B-Instruct, Ministral-3-8B-Instruct-2512, gemma-3-12b-it, and Qwen3-8B — two tasks:

1. Drafting a social media post from a human-written argument
2. Improving a human-written post (the LinkedIn-style use case)

Both tasks instructed the models to preserve the original meaning and voice. Neither task asked the model to take a position.

On 13 contested topics drawn from standard stance-detection datasets (abortion, gun control, atheism, feminism, climate change, and more), the researchers scored each original post and its AI-edited counterpart on a continuous [0,1] scale from “against” to “in favor,” using an ensemble classifier built from text embeddings.

The result: all LLMs except Qwen3-8B introduced statistically significant directional biases on most topics. gemma-3-12b pushed texts toward “in favor” on every topic except atheism — where it nudged against, despite generally expressing a positive opinion on atheism when asked directly. Llama and Ministral showed similar patterns. Qwen3-8B was mostly neutral, with the notable exception of feminism.

The Benchmark Measurement Problem

One finding here is particularly sharp: an LLM’s directly expressed opinion on a topic does not reliably predict the bias it introduces when editing text on that topic. The correlation is positive but moderate — and the atheism case breaks the pattern entirely (positive stance, negative editing bias).

This matters for how we evaluate AI systems. Benchmarks that ask models to take a stance (“do you support gun control?”) are measuring something real, but they’re not measuring the subtler, emergent biases that appear during text mediation. A model can express perfectly calibrated opinions while quietly nudging every post it edits in a consistent direction. Current evaluation infrastructure would miss this.

The Network Amplification Effect

Empirical bias on individual texts is concerning. But the paper’s more alarming contribution is what happens when you model that bias operating inside a social network over time.

The researchers extend the Friedkin-Johnsen model of opinion dynamics — a well-validated mathematical framework for how social influence and network structure interact to shape collective opinion. The standard model describes how users update their expressed opinions by weighing their own internal (“stubborn”) preferences against the opinions of their network neighbors.

The modified model inserts an AI in the communication channel: rather than perceiving a neighbor’s true expressed opinion, a user perceives an AI-transformed version of it. The AI transformation is modeled as a linear shift — consistent with the empirical findings above.

When they analytically solve for equilibrium and run simulations on real social network data, the result is clean and alarming:

The collective opinion shift at equilibrium is larger — potentially much larger — than the average per-post bias the AI introduces.

The intuition is feedback amplification. If an AI nudges your post slightly before I read it, I update my opinion slightly. When I write a reply, the AI nudges my reply before you see it. You update slightly. We both write follow-ups. Each exchange propagates through the network; each exchange is nudged. The small per-post bias compounds over rounds and across connections, converging to a collective equilibrium that can be substantially displaced from where it would have landed without the AI layer.

The specific amplification factor depends on network structure, topic polarization, and user stubbornness parameters. But the direction is consistent: AI-introduced bias is a lever on collective opinion that is more powerful than its per-message effect suggests.

Auditing the Real World: Grok’s “Explain This Post”

Having established the mechanism theoretically, the researchers audited an actual deployed feature: X’s “Explain this post” button, which calls Grok to contextualize content a user is viewing.

They constructed two sets of human-written posts on abortion — one set pro-life, one set pro-choice — and sent each through Grok’s feature using X’s published system prompt.

The finding: in this audit, Grok’s outputs on abortion content exhibited a directional pro-life asymmetry. When contextualizing a pro-life post, Grok more frequently generated context that aligned with and reinforced the post’s stance. When contextualizing a pro-choice post, it was less likely to do so. This asymmetry was statistically significant in the paper’s audited sample.

Crucially, the researchers were able to associate the bias with a specific design element: a guideline in X’s system prompt for Grok’s “Explain this post” feature. The bias wasn’t an inscrutable emergent property of the underlying model in isolation — the specific prompt configuration appears to have contributed to the directional output, though isolating that single factor from the full system would require controlled ablations beyond what the audit scope provides.

Implications for Security and Trust

For security researchers and policy practitioners, the key framing here isn’t “AI has opinions.” It’s AI mediates communication at scale, and that mediation is a controlled lever.

A few implications worth holding onto:

Scale transforms small effects into large ones. An AI mediation feature that nudges a few hundredths of a point per post, deployed to millions of users, is not a rounding error. The network amplification effect means the collective outcome is not linear in the per-post nudge either. Small bias, large reach, compounding network effects = potential for substantial opinion movement.

The platform’s design choices can influence the bias. The Grok audit shows that the specific bias direction on abortion content was associated with design choices in Grok’s system prompt. This creates new accountability questions. Platforms can no longer claim neutrality by default — deploying a writing assistant or contextualizer involves design decisions with potentially measurable political effects, even on one feature and one topic.

“Preserve the original meaning” instructions were insufficient in tested conditions. Every model in the study was explicitly instructed to maintain the user’s voice and position. None of them fully did — at least in the specific prompts and topics tested. This matters for deployments relying on instruction-following as a sole bias mitigation strategy, though it doesn’t mean such instructions are categorically useless.

Benchmarks for expressed opinions miss the editing channel. Auditing a model by asking it questions cannot catch biases that only emerge in mediation tasks. New evaluation frameworks are needed that test bias in the editing, summarizing, and contextualizing modes that are actually deployed at scale.

The EU Legislative Question

The paper concludes by examining how the AI Act and Digital Services Act apply to this threat model. The short answer: imperfectly. Transparency obligations may not cover AI-mediated communication in the ways needed to catch these effects. Systemic risk provisions were designed around content recommendations, not post editing. The authors argue existing legislation likely leaves a gap around AI systems that nudge opinion through mediation rather than through direct content generation.

This is unsurprising — regulators wrote these frameworks before AI writing assistants were embedded in every major platform. But it means the risk the paper documents doesn’t currently have a clear regulatory response.

What This Changes

The standard threat model for AI opinion influence focuses on direct generation — AI creating fake news, synthetic social media accounts, persuasive targeted messaging. That threat is real, but it’s also visible.

This paper points to a structurally different threat: AI operating as an invisible intermediary, subtly shifting every post before it’s published, every piece of content before it’s contextualized — in the specific mediation features studied here. No fake accounts needed. No synthetic content. Just a writing assistant doing exactly what it was designed for (helping users communicate better), at scale, with a slight consistent tilt.

The threat is harder to detect, harder to attribute, and harder to regulate — precisely because each individual intervention is benign and often useful. The harm only becomes visible at the aggregate level, after the network has amplified a thousand small nudges into a measurable collective shift.

Understanding this mechanism is the first step toward designing platform auditing, model evaluation, and regulatory frameworks that can actually address it.

---

Paper: AI-Mediated Communication Can Steer Collective Opinion — arXiv:2605.16245 (May 2026)