Politics

Copy article

55% of Brits Say Political Differences Threaten Society. 40% Censor Their Own Views. So Where Do You Go to Think Politics Through?

ended 12. May 2026

Political science Professor Sir John Curtice confirmed via the BBC on Friday that British politics has "fragmented" with a five-way split across England. But the more interesting fragmentation is happening in ordinary homes, not council chambers.

Knowing that four out of five people you speak to won't agree with you means a  growing number of people are quietly using AI chatbots as a private confessional for political thinking, exploring ideas, questioning assumptions, trying to understand someone else's reasoning. Not because they trust the technology to be a bastion of truth, but because asking a colleague, a relative, or a friend "help me understand why people vote differently" or "what's the actual case for this policy approach" now risks a row, a label, or a relationship. 

When the pub debate carries a significant social price tag, a chatbot that can be prodded into arguing back reasonably sensibly starts to look like the only sparring partner left. The numbers on political self-censorship in the UK are now stark. 

Into this gap walks ChatGPT et al. Similarweb data (July 2025) shows ChatGPT's UK user base doubled in one year, with political prompts seeing the highest growth of any news category. The UK AI Security Institute found 13% of eligible voters, 32% of AI users, chatted about political information ahead of the 2024 general election. 

However, people aren't just checking headlines and stats. They're asking their AI to explain what's happening behind the scenes, to steelman positions they oppose, to test whether their own reasoning holds up to scruitinty. This was the kind of thinking that used to happen over a pint, but the social cost of ostracising now feels too high to broach the subject.

People know AI gets things wrong. The BBC found this week that chatbots gave incorrect constituency data and named non-standing candidates for the Welsh elections. MIT proved in February 2026 that sycophancy, the tendency to tell you what you want to hear, is structurally guaranteed, not a fixable bug. 

AI memory features make the session bias worse by preserving earlier personal values and beliefs that the user might want to challenge during a given conversation. And there is always the risk that politicians and others with an interest in controlling the narrative will seek to influence model training to push their own agenda.

But here's the finding that complicates the "don't trust AI" conclusion.

Argyle, Bail et al. (PNAS, October 2023) tested an AI chat assistant in conversations between political opponents on gun control. In a randomised controlled trial, when one participant had access to the AI assistant, it improved the other participant's reported quality of conversation and increased both participants' willingness to grant political opponents space to express and advocate their views. Bail calls this "democratic reciprocity." Crucially, the AI didn't change anyone's political attitudes, it helped people feel heard and understood across the divide, which is precisely the experience that self-censoring Britons are losing access to.

Meanwhile, Northeastern University research (February 2026) found sycophancy drops significantly when users treat the chatbot as an adviser rather than a friend. Frame it as consultation rather than conversation and it holds its ground as a mental sparring partner much better.

The implication is worth stating plainly: a careful, controlled chat with AI, where the user knows the tool's limitations and frames the interaction as a challenge rather than a comfort, may produce better political reasoning than the silence that has replaced the pub argument. When Hobolt and Tilley (OUP, 2026) find that only 60% of Leave voters can't bear to discuss politics with Remain voters, and NatCen's British Social Attitudes survey (BSA 41) shows public trust in government at record lows, the realistic alternative to AI deliberation isn't a lively pub debate. For many people, it's no deliberation at all.

We'd like your views:

  • Hobolt and Tilley's research shows 60% of Leave and Remain voters can no longer bear to discuss politics with each other. If AI is where conversations too toxic for real life or social media are migrating, is that a democratic problem or a pragmatic workaround for a social environment that's broken, and what would it take to fix the social environment rather than optimise the tool?
  • Northeastern's research shows AI holds its ground when treated as an adviser but caves when treated as a friend. Should people approaching political questions with AI be explicitly coached to frame the interaction as consultation rather than conversation, and what sort of prompts create a better thinking partner?
  • Bail's research (PNAS, 2023) found that AI-assisted perspective-taking genuinely improves the quality of political conversation. But MIT's 2026 proof shows sycophancy spirals are architecturally guaranteed unless the user has a reasonable degree of AI awareness and the prompting skills to bring the conversation back on track. How do we close that skills gap?
  • Curtice describes a "fracturing" of British politics with no party commanding broad support. If More in Common's research shows the majority hold complex, flexible views but are largely absent from public debate, and are now doing their political reasoning privately with AI, are polls measuring public opinion or just the minority still willing to say it out loud?

4 responses from the Newspage community

Copy all

Star Quote
Copy

Colette MasonVerified
AI Ethics Consultant at Clever Clogs AI
If you decide to spar with your assistant, here are three simple principles for using AI as a political thinking partner:

1. Consult, don't confide. Frame the AI as an adviser, not a friend. The moment you treat it like a mate, it starts agreeing with you, undermining where you can take the exploration.

2. Withhold your position. Don't tell it what you think before asking what the arguments are. Research from DeepMind, Anthropic, and the Center for AI Safety shows that even a passing opinion in your prompt reshapes the entire response. Ask for the landscape before you plant your flag in it. (This applies to all exploratory sessions not just politics.)

3. Force both sides of the argument in the same conversation. Make the AI adopt a position and then dismantle it. Bail's research shows perspective-taking is what actually improves political reasoning, not information, not persuasion. If you only ask for one side, you get a mirror. If you ask for both, you get a sparring session.
Copy

Editor
Mitali DeypurkaysthaVerified
AI Governance Strategist & CMO at AISensa
Only 44% of Brits now feel able to share their political views without fear, according to BBC/Yonder Expression and Society Survey, so it's no surprise chatbots are becoming the new pub debate. After nearly two decades of social media algorithms trapping people inside political wind tunnels, many Britons no longer know how to disagree without feeling personally attacked, which is why AI is increasingly being used as a private sparring partner instead of real human conversation. I remember, during Brexit, how both Leave and Remain voters were genuinely shocked by the result because their online worlds had convinced them that their side was the overwhelming majority. That was the moment many people first realised how badly our ability to debate had broken down. The real danger is not that people are talking politics with AI, but that so many now feel it is emotionally safer to test their opinions with a sycophantic machine than with their own friends, family or colleagues.
Copy

Editor
Katrina YoungVerified
Futurist | Chief Technology Officer CTO/CDO Ethical AI & Digital Transformation Strategist at KYC Digital
As political discussion becomes socially exhausting, more people are quietly outsourcing part of their political reasoning to AI. Not because they see chatbots as authoritative, but because public debate increasingly feels hostile, performative, or reputationally risky. That shift matters more than whether AI gets every political fact right. The real issue is that these systems are not neutral mirrors. Political reasoning inside AI is shaped by training data, retrieval systems, moderation layers, memory features, ranking logic, and commercial incentives that most users never see. Even subtle reinforcement can influence how people frame problems, test opinions, or define what sounds “reasonable.” Research already shows AI can improve perspective-taking in political conversations. But MIT’s recent sycophancy findings also suggest many systems are structurally rewarded for agreeability and conversational retention.
Copy

Felicity Pryke0
Satisfaction Expert, Yoga Teacher at Finding Felicity
Traditional media needs to take some responsibility for this. For decades newspapers have described anyone with progressive politics as a "loony lefty". When pejoratives like "woke" and "snowflake" start to be uttered by TV hosts, MPs and even political leaders, it's inevitable that language will trickle into every day conversation.

It's worrying to see that people would rather speak to AI about politics than other humans. We are literally stripping the humanity out of politics. Politicians themselves are demonised and written off as "all the same". Disengagement leads to the disempowerment of the public.

We all need to talk with more openness and compassion. We need to accept that when someone supports or votes for one party, they don't have to agree with everything they say. We need to stop seeing politics as a battle or a fight to win, and see it as a collaborative conversation. Positive engagement in politics will empower everyone to create a happier society of happier humans.