Social media is destroying democracy. AI will make it worse. That, roughly, is the consensus.

It is a story told in a thousand articles, a hundred conferences, and every other panel discussion in Brussels. Social media polarised us, radicalised us, and handed the megaphone to the loudest and angriest voices in the room. And now AI is expected to supercharge the problem: deepfakes, hallucinations, manipulation at scale.

But what if that story is too simple?

A growing body of evidence suggests AI may not be pushing public discourse further toward the extremes in the way many feared. In some important respects, it may be nudging in the opposite direction: back toward evidence, expert consensus, and more moderate views.

That possibility matters to me because it matches something I have noticed in my own behaviour. I use AI to fact-check things. Not occasionally. Routinely. Someone makes a claim in a meeting, I check it. I read a statistic in an article, I check it. I see a hot take on LinkedIn, I check it. I do it with caution. I have seen hallucinations, and I know the systems are not perfect. But most of the time they are already good enough to be genuinely useful. And when I am dealing with something complex or consequential, I check the sources carefully. What has changed is not that I trust AI blindly. It is that checking has become faster, easier, and more instinctive than it used to be.

I used to think this made me slightly obsessive. Then I saw the research.

A recent study analysed nearly 1.7 million fact-checking interactions on X and found that 7.6% of all messages sent to AI bots were fact-checking requests. People tagging @Grok and asking, simply, “Is this true?” Not researchers or policy wonks. Ordinary users, across the political spectrum, voluntarily asking a machine to verify what they had just read.

That, to me, is the underappreciated story.

The counter-narrative

John Burn-Murdoch made this case eloquently in the Financial Times recently, drawing on a growing body of research. His argument: social media is populist and polarising; AI may be the great re-moderator.

The philosopher Dan Williams, writing on his Substack “Conspicuous Cognition,” frames it even more sharply. Social media, he argues, has been a “democratising” technology, shifting power from experts and gatekeepers to the crowd. AI chatbots are a “technocratising” force. They push people back toward expert consensus, evidence-based reasoning, and moderate positions.

Data from the Cooperative Election Study in the US shows the pattern: heavy social media users skew more populist, more anti-expert, more drawn to radical positions. That is a 15-year trend. Meanwhile, a study published in Science found that personalised AI dialogues reduced conspiracy beliefs by around 20%, with effects lasting at least two months and working even on deeply entrenched believers. Another study showed that exposure to AI fact-checks meaningfully shifted users toward more accurate beliefs.

Even Grok, built by Elon Musk with explicitly anti-establishment intentions, tends to align with expert consensus when used for fact-checking. I have tested this myself on politically charged topics, and it was consistently critical of Musk’s own views. That said, Grok has also produced notable errors, including wrongly labelling real photographs as fake during breaking news. The tendency is toward accuracy. But it is a tendency, not a guarantee. Which is exactly the point.

The real opportunity is behavioural

What interests me most is not the fantasy that AI has solved truth. It plainly has not.

This week’s reporting on Google AI Overviews is a useful corrective. Even after upgrades, the system is wrong roughly one time in ten. That sounds manageable until you apply it to five trillion annual searches: millions of incorrect answers every hour. That matters. It should make all of us more cautious.

But that is also why I think the strongest case for AI is behavioural, not technical.

The promise is not that AI always gives the right answer. It does not. The promise is that it may be making verification more normal. It may be training more people to pause, interrogate, and check before they repeat a claim.

That is a very different claim. And a much stronger one.

The underlying structure matters

Social media was not a neutral tool that happened to produce bad outcomes. It was designed around engagement, and engagement tends to reward provocation, emotional intensity, tribal signalling, and conflict. Polarisation is not an accidental by-product of that system. It is one of its outputs.

AI systems are structured differently. They are useful only insofar as they are at least broadly accurate, relevant, and grounded enough to help people get to a better answer. That does not remove hallucinations or bias. But the underlying pull is different. One system rewards emotional escalation. The other, at its best, rewards approximation to truth.

Take vaccines. Social media was an almost ideal environment for anti-vax misinformation because emotionally charged falsehoods could spread faster than evidence-based correction. AI is different. Ask a major chatbot about vaccines and it will not tell you that Bill Gates is trying to microchip the population. It will move back toward the scientific consensus. Imperfectly, yes. With caution required, absolutely. But structurally, that is a very different direction of travel.

And the pattern repeats. We have seen the anti-vax playbook migrate into other movements. The populist anti-EU narratives that now circulate across European social media follow the same logic: emotionally charged claims, thin on evidence, designed to spread faster than anyone can fact-check them. If people develop the instinct to ask “Is this actually true?” before sharing, the dynamic changes.

The invisible layer

Here is where I want to push this further. Why is WhatsApp not equipped with a fact-checking button already?

WhatsApp groups are where some of the most damaging misinformation spreads, precisely because there is no moderation, no context, no friction between receiving a claim and forwarding it. Now imagine a simple button next to any message: “Fact-check this.” One tap. No need to copy text, open another app, type a query. Just an instant layer of context on whatever claim just landed in your group chat.

Imagine that across all your platforms. A fact-checking plugin for every input: your social feeds, your messaging apps, your news aggregator. Not intrusive. Not preachy. Just there when you are curious.

This is where AI follows the trajectory of every transformative technology. Electricity and the internet became invisible. AI will become invisible too. It will pervade the information layer itself, embedded in the platforms where people actually consume and share content.

And when it does, every conversation gains access to something it currently lacks: a reasonable voice in the middle of the screaming match. Not a voice that tells everyone to calm down. Not a voice that picks a side. Just a voice that says, calmly and with evidence: here is what we actually know about this.

Where caution matters

This is where I want to be honest about the risks. None of this means AI is reliably right. None of it means we can relax.

There is the fluency problem. AI is so good at sounding reasonable that the very habit of checking I am describing could be undermined by it. A confident, well-structured answer can look like professional consensus even when it is not. The fluency of the output can lull you into trusting the content. That is a real danger, and it means the critical reading I advocate is not optional. It is essential.

There is the slop problem. AI systems increasingly train on content that AI itself has generated. As low-quality, machine-produced content floods the web, the very sources these tools draw on become contaminated. This is the ouroboros problem: the snake eating its own tail. If AI fact-checks against AI-generated slop, the whole system degrades. Source quality is not a secondary concern. It is existential for the credibility of these tools.

And there is the control problem. The argument I am making assumes that the major AI systems will continue to be broadly aligned with expert consensus. But that alignment is not guaranteed. Sovereign AI projects, state-backed models designed to serve national narratives, could be tuned in the opposite direction. Unfiltered open-source models could be weaponised for radicalisation rather than de-polarisation. The same architecture that makes AI a potential force for reason could, in different hands and with different training, become the most powerful propaganda tool ever built. The guardrails matter. And who controls them matters even more.

But you still have to think

AI can inform your thinking. It cannot replace it.

The fact-checking habit is powerful, but only if you actually read what comes back and judge it critically. You can become hyper-intelligent using AI, but only if you pay attention to the output, interrogate it, and apply your own judgement to what it means. As I argued in my previous article on the unbundling of intelligence and judgement: AI delivers the intelligence. The judgement is still yours.

That is not a limitation. It is the point. AI does not think for you. It raises the floor of what you can think with.

The Brussels lens

After nearly three decades working in and around EU public affairs and communication, I have seen first-hand how much the structure of the information environment shapes what succeeds.

The EU’s instinct has always been to make the rational case. The problem is that social media rarely rewarded rational cases. It rewarded emotional, anti-establishment, low-context narratives instead.

If AI genuinely pushes even a portion of users back toward evidence and context, that could alter the strategic landscape. Not because complexity suddenly becomes exciting. Simply because there may now be a mass-market information layer whose incentives are at least partly aligned with checking and synthesis rather than pure outrage.

That alone would be a significant change.

I have hope

I am a techno-optimist about this. Not a naive one. I know the hallucinations, the sycophancy, the fluency traps, the slop, the risks of capture by powerful interests. But I also know what I see happening in practice.

We have the AI-will-destroy-the-world narrative on one side and the AI-will-solve-everything narrative on the other. Both are lazy. But somewhere between them is a specific, evidence-based possibility that deserves more attention: AI might be a net positive force that takes the edge off the extremism and polarisation that social media amplified.

We spent 15 years watching social media train people to react. AI might train more of us to check.

Maybe, just maybe, that could make the world a little more reasonable.

Have you noticed your own research habits shifting since you started using AI?

Sources and further reading

This article draws on several recent pieces of research and analysis:

John Burn-Murdoch’s Financial Times article “Social media is populist and polarising; AI may be the opposite” (March 2026) first brought much of this evidence together. Dan Williams’ essay “How AI Will Reshape Public Opinion” on Conspicuous Cognition (March 2026) provides the deeper theoretical framework.

The fact-checking data comes from Renault, Mosleh & Rand’s working paper “@Grok Is This True? LLM-Powered Fact-Checking on Social Media” (2026), which analysed nearly 1.7 million interactions on X. The conspiracy belief findings come from Costello, Pennycook & Rand’s peer-reviewed study “Durably Reducing Conspiracy Beliefs Through Dialogues with AI”, published in Science (September 2024). The Google AI Overviews accuracy reporting was covered this week in Popular Science and The New York Times.

The social media polarisation data referenced throughout draws on the Cooperative Election Study at Tufts University.

This article was originally posted by Philip Weiss on Linkedin