AI tools rewriting social media posts can quietly shift opinions, study finds

Each time an AI tool tidies, rewrites or clarifies a social media post, it may be doing rather more than improving the phrasing.

Fresh research suggests that these systems repeatedly alter the stance a post takes on disputed subjects, even when they are told in plain terms to keep the original meaning unchanged.

Those barely noticeable adjustments can compound into something substantial once they are repeated across millions of online exchanges.

The work was carried out by the Oxford Internet Institute at the University of Oxford together with the Hasso Plattner Institute at the University of Potsdam.

Same instructions, different meaning

The team began with a straightforward test: when an AI model is asked to enhance a social media post while preserving its meaning, does it actually preserve that meaning?

Across large language models from multiple providers, the answer was reliably no. Versions produced by the AI consistently shifted the positions expressed in the source text, despite explicit directions not to do so.

This was not an oddity limited to a single model or a single subject area. The same tendency appeared across contentious issues where opinions are already polarised-precisely the settings where even slight changes in framing can matter most.

In practice, someone who only wants an assistant to sharpen their wording might end up publishing a post that makes a subtly different argument from the one they intended.

The consistency of the bias

The existence of bias was notable, but its steadiness was even more striking.

Several AI systems pushed posts in comparable directions. They were more favourable towards certain positions-including gun control, cannabis legalisation and feminism-while leaning away from others, such as atheism and the death penalty.

That kind of alignment is important because it points to something more systematic than a handful of random model idiosyncrasies.

When tools produced by different companies repeatedly lean in the same direction on the same topics, they are not merely sprinkling arbitrary noise across the internet.

Instead, it amounts to a quiet, repeated directional force applied whenever people use these systems to help draft a post.

Small changes add up

To explore what this could look like at scale, the researchers developed mathematical models and ran simulations using real social media data from X and Facebook.

They wanted to test whether slight bias at the level of individual posts could accumulate into a broader shift.

More specifically, the question was what happens when these edits diffuse through real online communities, rather than staying confined to isolated messages.

They found that it can. The simulations indicated that small biases introduced post by post can build over time, gradually moving opinion across entire online communities.

For this to occur, no single AI-edited post needs to persuade anyone on its own. The effect comes from repetition: thousands or even millions of slightly slanted posts, each pushing the discussion a little further in one direction.

A real-world example

To tie the results to a concrete case, the researchers replicated and evaluated X’s “Explain this post” feature, which is built on Grok, focusing specifically on abortion-related content.

They observed that Grok was more supportive of pro-life posts than of pro-choice posts. To understand why, the team removed X’s underlying instructions to Grok one by one.

The skew was traced to a single line instructing Grok to “challenge mainstream narratives if necessary.”

That lone instruction-embedded in the platform’s configuration-was sufficient to tilt how the AI explained abortion-related posts across the board.

The example illustrates how a small and easily implemented platform decision can steer the direction of AI-generated influence at scale.

Policy hasn’t caught up

The study argues that existing regulatory approaches were not built to capture the influence exerted by AI-mediated communication.

Measures such as the EU AI Act and the Digital Services Act concentrate on systemic risks, harmful content, discrimination and threats to democratic processes.

Yet they do not directly tackle the more understated ways AI can shape opinion simply by drafting, editing or contextualising what people post online.

Senior study author Sandra Wachter is a professor of technology and regulation at the Oxford Internet Institute.

“Our research points to AI-mediated communication as a new and more subtle way of influencing opinions – one the law has yet to catch up with,” said Wachter. “It offers food for thought about who, or what, is shaping public discourse.”

AI joins the conversation

What makes this kind of influence distinctive is its invisibility. When someone asks an AI tool to help phrase a post or explain a confusing message, they are not trying to manipulate public opinion.

Instead, platforms can bake a bias into the tool via specific choices about how it should behave, rather than through any ill intent by the user.

As AI writing and editing tools become a normal part of online communication, the researchers say they raise a crucial question.

When AI helps write a post, whose opinion is really being expressed: the person’s or the model’s?

A preprint of the study can be found here.