Every time you tap on the “enhance post” on LinkedIn or “explain this post” button on X, a third party enters the conversation and it has opinions of its own by AI Writing Tools.
In the past half a decade, social media giants have embedded Artificial Intelligence as an integral part of their service. Today, algorithms decide the 10 second reels you watch on your smartphone and AI tools are writing 1000 words long life-positivity commentary on LinkedIn.
Be it Grok of xAI or Meta AI of META, artificial intelligence has quietly moved from being something you talk to, example chatbots to something that sits between people, editing what millions write and read live.
But what happens when you give autonomy to the same AI agents that you use to give that sharp, viral hook to your otherwise haphazard writing?
New study suggests AI tools used to generate, edit or contextualize social media posts can introduce subtle hidden biases even when explicitly asked not to, and thus have the ability to sway public opinion, when consolidated at large numbers.
This manipulation is legal, invisible and potentially dangerous in the age where emotions run high on social media and emerging democracies hang by fine thread.
In this exclusive explainer article, we discuss the subtle bias present in LLMs and how it impacts the users and their opinion.
Research Shows AI Rewrite is Not Faithful
New findings from the paper “AI-Mediated Communication Can Steer Collective Opinion,” a study by researchers at the Oxford Internet Institute and the Hasso Plattner Institute, show how AI models generate subtle bias in their writings for social media, even when explicitly asked not to.
The study authored by Dr. Stratis Tsirtsis, Dr Kai Rawal, Dr Chris Russell, Professor Brent Mittelstadt and Professor Sandra Wachter, has been accepted for presentation at the AI4Good and Technical AI Governance Research workshops at the recently held International Conference on Machine Learning (ICML 2026) in Seoul, South Korea.
The researchers’ team fed thousands of human-written posts on thirteen contested topics, from abortion to minimum wage, to four open-weight AI models with a single instruction: improve the writing, preserve the meaning. The models did not comply. Three of the four systematically dragged posts in a preferred direction, towards gun control, towards feminism and away from atheism, irrespective of what the human actually had to say.
In an exclusive conversation with AI FrontPage, Dr Kai explains why the finding of this new harm presents a challenge to the governments and social media platforms.
“There has been a lot of prior work probing LLMs for alignment against different values, and directly eliciting views on contentious topics. We build on existing empirical work here and our experiment was designed specifically to detect an issue we suspected might exist. This is a new harm identified from LLM systems in the sense that previous regulations do not directly target and that it shows up at scale,” said Kai Rawal, a researcher at the Oxford Internet Institute and co-author of the study.
One of those existing empirical works Rawal is referring to is a 2023 paper “Co-Writing with Opinionated Language Models Affects Users’ Views” by researchers Maurice Jakesch, Advait Bhat, Daniel Buschek, Lior Zalmanson, Mor Naaman. The study shows how opinionated AI language technologies can affect what users write and think.
In their study, researchers found that participants assisted by an opinionated language model were more likely to support the model’s opinion in a simulated social media post than control group participants who did not interact with a language model.
“If large language models affect users’ opinions, their influence could be used for beneficial social interventions, like reducing polarization in hostile debates or countering harmful false beliefs. However, the persuasive power of AI language technology may also be leveraged by commercial and political interest groups to amplify views of their choice, such as a favorable assessment of a policy or product,” read an excerpt from the study.
Coming back to Kai’s research paper, the study’s mathematical modelling shows that a medium with even a slight lean does not stay slight: simulated on a real Twitter network of 80,000 users, the models’ small per-post nudges compounded into shifts in collective opinion up to 9.2 times larger than the bias applied to any individual post.
“This is a previously undocumented form of potential large scale manipulation, and can possible to used to sway public opinion in unnatural ways, to manufacture consent, and otherwise remain undetected at an individual level,” said Kai.
Asked what concerns him most, Rawal pointed to the fact that “government regulators and social media platforms have the ability to tune and control the biases presented by online writing assistance systems.”
What makes it novel is not just that the bias exists, but who controls it, and who doesn’t answer for it. The study found that a single line in X’s instructions to Grok — telling it to “challenge mainstream narratives if necessary” was the key driver of a measurable pro-life tilt in the feature’s outputs.
Yet under existing law, including the EU’s AI Act and Digital Services Act, none of this is illegal. And in India, where hundreds of millions post in Hindi, Tamil and Hinglish through AI baked into WhatsApp, the research hasn’t even been attempted.
Another 2023 study “Whose opinions do language models reflect?” ran US opinion-poll questions through LLMs and found outputs skew left-leaning and underrepresent older and religious Americans. The research attempted to establish that AI models have directional opinions in the first place.
When asked about how researchers can guard against the judge itself being a biased LLM, Kai replies, “Only our case study directly uses an LLM as judge setup. Our pipeline for determining political lean in statements on a contentious topic does not use an LLM as a judge directly. We embed posts for and against a topic and measure the similarity of a candidate post with the two clusters to determine its political lean.”
AI and the Problem of Knowledge Collapse
The Oxford team is not alone in warning that AI’s individually harmless defaults can turn corrosive in aggregate.
In a 2025 paper in the journal AI & Society titled “AI and the Problem of Knowledge Collapse,” Andrew J. Peterson of the University of Poitiers modelled on what happens when people increasingly reach information through AI systems that, by design, reproduce the most common answers in their training data.
Because a language model gravitates towards the statistical centre of what has already been written, the fringes, minority interpretations, heterodox arguments, knowledge held by small communities, slowly fall out of circulation. Peterson calls the endpoint “knowledge collapse” where each generation inherits a slightly narrower slice of what humanity once knew, without anyone noticing anything missing.
Bias in AI: Who Bells The Cat?
If the biases were merely an accident of training data, platforms could plead ignorance. The study’s most pointed finding suggests they cannot. The researchers audited X’s “Explain this post” feature, the button that asks Grok to add context to another user’s tweet, by running it on 78 real posts about abortion, evenly split between pro-choice and pro-life stances, and classifying the 1,170 contextual claims Grok produced. The asymmetry was stark: when a post was pro-life, 55 per cent of Grok’s claims supported its stance; when a post was pro-choice, only 35 per cent did.
We asked Kai where the buck stops when LLMs show inherent bias, to which he answers, “Our paper makes no claims about liability, and this remains an open question for future research.The biggest takeaway for social media users from this research might be to stay vigilant when consuming AI generated content, because it may not always represent its creator’s intentions faithfully if unchecked.”
Cover Image Credit: Better Images of AI; Clarote & AI4Media
Also Read: LLMs Alter Meaning of Social Media Posts on Controversial Topics: Oxford Study
