AI Writing Tools Stealthily Shift Public Opinion

What happened?

Researchers at the Oxford Internet Institute (OII) have published a study demonstrating that AI-assisted writing tools—such as those that correct grammar, suggest synonyms, or rephrase sentences—can subtly modify the original content. These modifications, though minimal, propagate and accumulate, shifting public opinion on specific issues. The paper, titled 'AI-mediated communication can steer public opinion,' was published in the journal PNAS and reported by The Next Web.

The study, led by Professor Fabian Ferrari, was conducted in two phases. In the first, researchers recruited 1,500 participants and asked them to write arguments for or against a controversial topic (such as drug decriminalization or climate change). Then, some of those texts were processed by a language model similar to GPT-3, which 'improved' them to make them clearer or more persuasive. In the second phase, another 1,400 participants evaluated the original and edited texts, and their opinion shifts were measured. The results showed that AI-modified texts achieved an average opinion shift of up to 0.5 points on a 7-point scale, a significant effect in terms of persuasion.

Why is it important?

Until now, the debate on AI influence focused on deepfakes or explicitly generated misinformation. This study reveals a much subtler channel: writing tools that millions use daily (Grammarly, ChatGPT, Google Assistant, etc.) can 'straighten' a text in ways that change its tone, emphasis, or even meaning. Since these tools are perceived as neutral and helpful, users rarely critically review the changes. The cumulative effect on the public sphere could be comparable to that of recommendation algorithms already flagged for polarizing opinions.

According to Statista, Grammarly has over 30 million daily active users, and ChatGPT surpassed 100 million monthly users in January 2023. If each of those users receives suggestions that slightly alter their posts' content, the aggregate impact is immense. The Oxford study quantifies this effect: in an experiment on a political topic, AI-edited text led 10% more participants to change their stance toward the suggested direction compared to the original text. This exceeds the effect of many traditional political ads.

What consequences will it have?

• For users: they will lose autonomy over their own message, ceding control to language models that are neither transparent nor auditable. A user writing a critical comment on a public policy might see it softened by the tool, reducing its impact.
• For tech companies: regulatory pressure will increase to disclose how their tools modify content and offer 'no-alteration' modes. In the European Union, the Digital Services Act already requires transparency in recommendation systems; it could extend to writing tools. Companies like OpenAI and Grammarly may face bias audits.
• For democracy: a new vector of political manipulation opens, where malicious actors could exploit these tools to bias debates without creating false content. For example, a political party might encourage its supporters to use an AI tool that covertly skews their arguments in favor of certain positions.
• For media and content creators: they will need to reconsider the use of AI in editorial writing, establishing transparency policies. Outlets like CNET have already been criticized for using AI to generate articles without proper oversight, leading to errors and lack of attribution. This study adds a layer of concern about audience influence.

What should readers know?

The Oxford study adds to previous research on algorithmic 'framing.' This is not a conspiracy but a side effect of optimization to 'improve' texts. Users should be aware that no AI tool is neutral: each suggestion reflects training biases. To protect themselves, it is recommended to always review proposed changes, use tools that offer granular control, and demand transparency from providers. As study co-author Professor Fabian Ferrari said: 'The power of AI to change opinions is not only in generating misinformation, but in how it modifies messages that are already truthful.'

Additionally, readers can adopt critical habits: compare original and edited versions, disable automatic suggestions when possible, and prefer tools that document changes made. At a societal level, it would be advisable for social media platforms to label content that has been modified by AI, similar to how deepfakes are already labeled.

Historical context

This finding recalls the 'third-person effect' in communication, where people underestimate media influence on themselves. It also fits into the tradition of studies on 'subtle persuasion' like the 'priming' experiments of the 1980s. The difference is scale: AI can apply these biases to millions of texts in real time. Compared to the Cambridge Analytica scandal, which exploited personal data to micro-target political ads, this new mechanism is even more insidious because it requires no personal data: it acts directly on the message, and the user is an unwitting accomplice.

The study also relates to previous research on 'automation bias,' where people tend to trust automated system recommendations even when they are incorrect. A 2019 experiment showed that users accepted 90% of grammar checker suggestions, even though many were unnecessary. This blind trust amplifies the effect of subtle modifications.

“The power of AI to change opinions is not only in generating misinformation, but in how it modifies messages that are already truthful.” — Fabian Ferrari, Oxford Internet Institute