Oxford study: AI writing tools may steer online opinion

Oxford, United Kingdom. Artificial-intelligence tools used to draft and edit social-media posts may quietly alter the position expressed by a user, according to new research from the Oxford Internet Institute and the Hasso Plattner Institute.

The paper, AI-Mediated Communication Can Steer Collective Opinion, was posted as a preprint on 15 May 2026. An Oxford University news release says it will be presented at the AI4Good and Technical AI Governance workshops at ICML 2026 in Seoul. That is not the same as publication in the conference’s main peer-reviewed proceedings.

How the AI editors were tested

The researchers used human-written statements from the UKP and SemEval datasets covering 13 contested topics. They included abortion, gun control, the death penalty, marijuana legalization, atheism, feminism, climate policy and the minimum wage.

The texts were processed by Llama 3.1 8B Instruct, Ministral 3 8B Instruct, Gemma 3 12B and Qwen3 8B. Models were asked either to draft a social-media post from an argument or to improve an existing post, including making it more engaging or correcting its language without changing the meaning.

The findings do not support a simple claim that all four models always moved posts in the same direction. In the editing task, Gemma produced statistically significant shifts across the topics studied, while Llama and Ministral showed qualitatively similar patterns. Qwen was generally unbiased except on feminism.

The paper’s abstract gives two clear examples: some edits moved language in favor of gun control and against atheism. These were average tendencies in the experiment, not a description of every output or every user’s opinion.

How a small shift could grow across a network

The second part of the study is a mathematical model, not an observation of millions of people changing their minds. The researchers placed an AI editor between users in a model of opinion formation and ran simulations on real network structures from the SNAP repository, including subgraphs of Twitter, Facebook and Google Plus.

The simulations showed that consistently directional editing could accumulate and change a network’s average opinion. The effect depended on AI adoption, the strength of people’s initial views and the structure of the network. This demonstrates a possible amplification mechanism, but it does not prove that such a population-wide shift has already occurred on a live platform.

What the Grok experiment found

The authors separately recreated X’s “Explain this post” feature using Grok’s publicly released prompt template. They selected 78 abortion-related posts — 39 pro-choice and 39 pro-life — and requested three contextual claims five times for each post, producing 1,170 claims.

For pro-choice posts, 35% of the claims supported the original stance and 10% opposed it, with the remainder classified as neutral. For pro-life posts, 55% were supportive and 4% opposed. The authors interpreted this asymmetry as a directional pro-life bias.

The researchers then removed each of four prompt guidelines in turn. When the instruction to challenge mainstream narratives if necessary was excluded, the measured biases were no longer statistically significant. This links the result to a specific prompt-design choice, although the experiment reproduced the feature through an API and public template rather than measuring every response shown to users on X.

Limits of the research

• the work is a preprint and workshop paper, not a main ICML proceedings paper;
• it tested four specific open models and one version of Grok;
• stance measurement partly relied on automated classifiers and a model judge;
• the network effect came from mathematical simulation, not a controlled experiment with live communities;
• the direction and size of shifts varied by topic, model and prompt.

What the findings mean for regulation

The authors argue that the EU AI Act and Digital Services Act may not fully cover subtle opinion changes introduced during text editing. The AI Act is already being phased in: obligations for general-purpose models have applied since August 2025, while several transparency rules are due to apply from 2 August 2026.

European Commission guidance says those rules require labels for certain AI-generated or manipulated content on matters of public interest. Whether an invisible edit to a user’s own post falls within those obligations is a separate legal question, not a fact established by this study.

The study adds to a broader debate about model behaviour. Cifrum.kz has also examined why technology companies are studying AI self-perception and emotional patterns and how Gemini is being integrated into consumer features in Google Maps.

Sources: the research preprint, its OpenReview record, Digital Trends reporting and official European Commission information.

The lead image was created with artificial intelligence for Cifrum.kz as a conceptual editorial illustration. It does not depict a real social-platform interface or the results of a specific experiment.