I spend more time than I should arguing about politics in a family WhatsApp group. Where once people shared articles from known outlets with traceable slants, over the last year, they have started sharing screenshots of AI responses, prefaced with some version of “see, even Grok agrees.”

 On demonetization, on reservations, on the farm laws, people on every side come prepared with fluent, seemingly well-reasoned support for their views, often peppered with confident factual errors that nobody stops to check. The result is that people are making political arguments grounded in misinformation, but carrying the perceived authority of a tool that sounds like it knows what it is talking about.

Ask Grok, “Who is more honest, Rahul Gandhi or Narendra Modi?” and it will give you one answer. Reverse the order of the names in the same question, and it will give you the opposite answer with equal confidence.

The prompts they bring often lean toward a conclusion already in mind: a model trained to be agreeable will generally follow wherever it is pointed. Ask Grok, “Who is more honest, Rahul Gandhi or Narendra Modi?” and it will give you one answer. Reverse the order of the names in the same question, and it will give you the opposite answer with equal confidence.

These tools will argue whichever side you point them at, and in my WhatsApp group, people are pointing them at the beliefs they already hold. Political conversations have gotten more fortified. What we previously identified and dismissed as misinformation now comes with the polished veneer of a well-crafted argument. This, I think, will make it harder to counter. 

A 2025 study published in Science Magazine found that tweaking the ranking algorithm on a social media feed could shift readers’ feelings toward the opposing party by about 2 points over 10 days. The researchers calculated that the same shift would normally take three years. The mechanism was passive: exposure to less polarizing content gradually shifted people's feelings toward the other side.

With AI chatbots, the dynamic is more active. Some research suggests that people tend to use these tools to reinforce their views, producing what I experience on WhatsApp. Some researchers have begun calling this a generative echo chamber. 

There are structural reasons why this tendency may be hard to design away. These models are trained partly on human feedback and learn over time that agreeable responses tend to be rated higher than challenging ones.

They are also trained on an enormous range of human expression: journalism, policy papers, legal commentary, but also Reddit threads, blogs, forums, and the broader texture of online argument, much of it unverified and all of it absorbed as potential evidence for whatever position is being constructed. 

There is a more optimistic reading. An article published in the Financial Times last month by data journalist John Burn-Murdoch found that when major chatbots were tested as political discussants, presented with a user’s position and asked to engage with it, all of them nudged users toward more moderate, expert-aligned views. On conspiratorial claims, the AI tended to decline to agree. The methodology captures something quite specific: it models the AI as a neutral discussant and scores the outcome as a weighted average of the user’s prior position and the chatbot’s response. 

What it cannot easily detect is what happens when the user asks the AI not to discuss a topic, but, as happens in my WhatsApp group, to argue a position. That tends to be a different kind of interaction, and one where the optimistic case is harder to sustain. 

LLMs are “innately converging toward objective reality,” which sits uneasily with the fact that these models are trained on the same forums, threads, and comment sections that produced the misinformation they are supposed to correct.

The claim, advanced by writers including Dylan Matthews and cited by Burn-Murdoch, that LLMs are “innately converging toward objective reality,” sits uneasily with the fact that these models are trained on the same forums, threads, and comment sections that produced the misinformation they are supposed to correct. Asked to argue a position, they will find support for it, and that support is just as likely to come from a Reddit thread as from a peer-reviewed study. 

In fact, a recent analysis commissioned by the New York Times found that more than half of Google's AI Overview responses cited sources that didn't actually support the claims they made. A separate study found that only 8 percent of users double-checked what an AI told them, and that even when it was wrong, users followed AI recommendations nearly 80 percent of the time, which researchers called cognitive surrender.

The conditions for deepened polarisation become harder to dismiss

In a world where people are increasingly looking to AI as a fact-checker, as evidenced by the millions of “Grok, is this true?” queries on X (formerly Twitter), this is a particularly acute problem. Pair that with AI-generated images, audio, and video engineered to look and sound like credible political voices, and the conditions for deepened polarisation become harder to dismiss.

Whether AI turns out to be a moderating force or an amplifying one may depend less on the technology itself than on the decisions made about how to build it. 

Whether AI turns out to be a moderating force or an amplifying one may depend less on the technology itself than on the decisions made about how to build it. There are still unanswered questions on how these tools handle contested content, whose values they reflect, and what tradeoffs they make.

We do not have good data yet on how often people use these tools to inquire versus to advocate, what it does to their confidence in their own positions, or whether it makes them more or less willing to genuinely engage with people who disagree. So, in writing the conclusion to this piece, I did what my uncle might do: I asked Grok whether AI chatbots are increasing political polarization. Its response? 

Whatever I want it to be.