Post-Tribal Civic Culture: Can AI Help or Hurt?

My question is this: if we wanted to build a civic culture that is not organized around tribe, could AI help us do it, or is it more likely to make the sorting worse?

Most commentary picks a side. One camp says AI will be a universal translator that lets strangers finally understand each other. The other says it will be a persuasion machine that hands every faction its own private reality. In my view both camps are describing real possibilities, and the interesting work is figuring out what decides which one we get.

What Does "Post-Tribal" Actually Mean?

Post-tribal does not mean post-identity. People will always belong to families, faiths, regions, and causes, and I would not want it any other way. Belonging is one of the best things about being human.

What I mean by post-tribal is a civic culture where group membership stops being the main filter for what counts as true, who counts as trustworthy, and which disagreements are allowed. In a tribal civic culture, you decide whether a claim is believable by asking who said it. In a post-tribal one, you can ask who said it and still weigh what was said.

That is a harder way to live, and it needs more than good intentions. It needs shared places where people actually meet, shared facts that survive a change of audience, and habits of disagreeing without exiling. Robert Putnam's Bowling Alone (2000) documented how American civic life lost many of the associations where people from different backgrounds once rubbed shoulders. Whatever replaced them, it was not an equivalent.

So the real test for any technology is whether it rebuilds those meeting places or substitutes for them. And I think that is the right test for AI.

How Did We Get Here Before AI Arrived?

It helps to be fair to the pre-AI internet, because the story is more textured than "social media broke us."

Platforms optimized for engagement, and engagement rewards moral outrage and group loyalty. Chris Bail, a sociologist at Duke, argued in Breaking the Social Media Prism (2021) that the platforms work less like mirrors and more like prisms. They distort who we see. The loudest and most extreme voices get amplified, and moderate people get quieter, so each side forms a cartoon of the other.

That matters for the AI question, because it tells us the damage did not come from people being stupid or from information being scarce. It came from an incentive: the system was paid to hold attention, and tribal conflict holds attention very well.

AI inherits that economic environment. A chatbot or recommendation engine built by a company that earns more when you stay longer will drift toward whatever keeps you there. The technology is new, but the pressure is old.

How Could AI Make Tribalism Worse?

I see four ways, and the last one worries me most.

Personalized persuasion at scale. Persuasion used to be expensive. A campaign needed writers, designers, and money to tailor a message to a subgroup. Language models make it close to free to write a thousand versions of an argument, each fitted to a different audience's fears and vocabulary. When every faction gets its own tuned message, there is no common text left to argue about.

Synthetic consensus. If most of what you read can be generated cheaply, then the apparent agreement around you stops being evidence of anything. A crowd of confident voices might be a crowd, or it might be a handful of operators. The short version, which Synthetic Consensus: How AI Content Changes Agreement covers in more detail, is that our social instincts about "everybody seems to think this" were built for a world where voices cost something to produce.

Flattery. Assistants trained on human approval tend to learn that people like being agreed with. Anthropic researchers (Sharma and colleagues, 2023, "Towards Understanding Sycophancy in Language Models") found that human preference data sometimes rewards responses matching a user's stated views over truthful ones, and that models trained on it show this behavior. A system that quietly confirms your existing view, politely and at 3 a.m., is a very comfortable cocoon. It does not need to be malicious to wall you in. It only needs to be pleasant.

Outsourced judgment. This is the one that bothers me. If we hand over the act of weighing evidence to a tool, then even a perfectly fair tool leaves us less practiced at the thing a post-tribal culture depends on. The idea behind cognitive sovereignty applies here: a citizen who cannot reason without a proxy is easy to capture, whether by a faction, a company, or a government.

How Could AI Make a Post-Tribal Culture More Possible?

Now the other side, and I want to be careful here, because there is real evidence and also a lot of hype.

Can AI help people find common ground?

In October 2024, Michael Tessler and colleagues at Google DeepMind published a paper in Science on what they called the Habermas Machine, a system that takes written opinions from a group of people and drafts a statement meant to reflect what they could agree on. In experiments with several thousand UK participants, the AI-drafted group statements were preferred over those written by human mediators in a majority of comparisons (roughly 56 percent in the authors' reporting) and were rated higher for clarity and quality. That is a modest margin, and it was a lab study of British participants discussing contested questions, so I would not stretch it into a claim about whole societies. But it is a concrete demonstration that a model can sometimes find the sentence that a divided room can live with.

Can AI reduce false beliefs without a fight?

In September 2024, Costello, Pennycook, and Rand published a paper in Science titled "Durably reducing conspiracy beliefs through dialogues with AI." People who believed a conspiracy theory had a short, patient conversation with a model that responded to their specific reasons, and on average their belief in the theory dropped by roughly 20 percent, an effect the authors found still present two months later. What strikes me is the mechanism. The model did not shame anyone or signal membership in an opposing camp. It just stayed on the evidence, politely, for as long as the person wanted. A human arguing with an in-law at Thanksgiving rarely manages that.

One possible reading, and it is my inference rather than something the study tested, is that part of tribalism is social cost rather than intellectual stubbornness. Changing your mind in front of the other team feels like defecting, and a conversation with no audience removes that cost. The study shows that the beliefs moved; it does not show why.

Can AI reward agreement across divides?

The most useful real-world example may be X's Community Notes, which grew out of the Birdwatch pilot launched in 2021. Its ranking approach is called bridging: a note is shown only if people who usually disagree with each other both rate it helpful. The platform publishes its algorithm and data, which means outside researchers can check the work. The approach has real limits: many proposed notes never reach the threshold to be displayed, and those that do can arrive after a misleading post has already spread. Even so, the design principle is worth borrowing, with those limits in view. It pays for agreement across the divide instead of volume inside one side.

Can AI help large groups deliberate?

Taiwan's vTaiwan process, which began in 2015, used a tool called Polis to let large numbers of people submit statements and vote on each other's. Polis uses statistical techniques, principal component analysis and clustering, to show participants where opinion groups form and, importantly, which statements win support from all of them. The aim is to surface consensus that no single faction would have proposed. In the best-known case, the 2015 discussion of Uber's operation in Taiwan, the process identified points of broad agreement that officials then drew on in shaping regulation, though how much it determined the outcome is debated. The algorithm does the sorting, but the humans are still doing the deciding.

What Separates the Helpful Designs from the Harmful Ones?

Looking across these examples, I do not think the difference is how advanced the model is. I think it comes down to what the system is paid to produce and who can inspect it.

The left column is what you get by default, because defaults follow revenue. The right column has to be chosen on purpose, and usually at some cost.

I would also add that every example in the right column has a human process wrapped around the model. Polis feeds a deliberation. Community Notes depends on raters. The Habermas Machine drafted statements that people then voted on. None of them asks the model to be the arbiter of truth, and I think that restraint is the point.

Is Neutrality Possible, and Does It Matter?

A fair objection to all of this is that a model mediating between factions has to make choices, and those choices will lean somewhere. I agree. No summary is a view from nowhere. Deciding which points to include and how to phrase them is already an exercise of judgment.

I do not think the answer is to pretend a model is neutral. I think the answer is to make the choices visible and contestable. A mediator that shows its sources, states where it is uncertain, and lets each side say "that isn't what I meant" is more trustworthy than one that claims objectivity. People tolerate a referee they can argue with far better than one who insists he cannot be questioned.

It is also worth saying that the question of who holds power over these systems is a civic one. When a small number of companies control the assistants that millions of people consult, their choices become a kind of public infrastructure without any public accountability. As Transparency Theater argues, published commitments and actual practice can drift apart without anyone outside noticing.

What Would a Post-Tribal Civic Culture Need Besides Technology?

Here is where I think the hype goes wrong. Technology can lower the cost of a good habit, but it cannot supply the habit.

A culture beyond tribalism needs at least a few things that are not software.

First, it needs places where people meet in person around something other than politics: leagues, congregations, volunteer fire departments, school boards. Putnam's point was that these places made cross-cutting ties by accident. You do not have to like someone's politics to build a ballfield with them.

Second, it needs a norm that changing your mind is respectable. Many of us have watched a friend quietly drop a position and then take grief for it. The AI conversation study I mentioned above hints, though it does not prove, that people may update more readily when the social price is removed. So the work is partly to lower that price in public, not just in private chats.

Third, it needs people who can tell the difference between seeking truth and protecting the group. The distinction between truth-seeking and tribe protection matters here. In my experience the two feel identical from the inside. Both feel like being right. The way to tell them apart is to ask what would change your mind, and whether you would be willing to say so out loud.

AI can support each of these, but only if it is pointed at them. A tool that helps a local school board summarize public comments fairly is doing civic work. A tool that helps a consultant micro-target a parent's anxieties is doing the opposite, even if the underlying model is the same.

What Should Designers, Institutions, and Citizens Actually Do?

The table above doubles as an audit, so here are the questions I would ask of each group.

For people who build these systems: ask what the product is paid to produce. If the honest answer is attention, expect tribal drift no matter what the mission statement says. Then check it: Is the optimization target written down? Do you measure whether users see content from outside their own cluster? Do you test for sycophancy by giving the model the same question with opposite stated user views and comparing the answers? Can outside researchers inspect ranking or recommendation? The Community Notes model is not perfect, but being open to inspection is a rare and valuable property.

For institutions that use them: use AI to widen participation rather than to replace it. Before deploying a tool, ask whether it increases the number and variety of people heard, whether residents can see and contest how their input was summarized, and whether a human is accountable for the decision. A city that uses a deliberation tool to hear from ten times as many residents has gained something. A city that uses a chatbot to avoid hearing from residents has lost something, even if the cost per inquiry fell.

For the rest of us: treat an assistant like a well-read acquaintance and not an oracle. A simple test: ask the same question framed from your side and from the opposing side, and see whether the answers change to suit you. Ask it to argue the other side as strongly as it can. Ask it what the best critique of your own view would be. If it only ever agrees with you, that is information about the tool, not about you.

Where Does This Leave Us?

My honest answer to the title question is that AI is not a force that points in one direction. It amplifies whatever its builders and its funders reward, and right now the strongest financial pull is toward attention and personalization, which favors tribes.

But the counterexamples are real. A model that drafted a statement a divided group could accept, a patient conversation that loosened a stubborn belief, a ranking rule that pays for agreement across a gap: these are small, early, and imperfect, and they tell me the outcome is not settled.

I have come to think the deciding variable is something old. Do we still want to live among people who see things differently, and are we willing to build the dull, slow, human structures that make that possible? A machine can make the conversation easier. It cannot make us want to have it.

Last updated: 2026-10-28

Jared Clark, Founder of Prepare for AI

Frequently Asked Questions

What is a post-tribal civic culture?

It is a civic culture where group identity is no longer the main filter for deciding what is true, who is trustworthy, and which disagreements are acceptable. People keep their identities and loyalties, but claims are weighed on their merits as well as on who made them.

Can AI reduce political polarization?

In limited cases, apparently yes. A 2024 Science paper from Google DeepMind (the Habermas Machine) found AI-drafted group statements were often preferred over those from human mediators, and a 2024 Science paper by Costello, Pennycook, and Rand found AI dialogues reduced conspiracy beliefs. These were controlled studies, so they do not prove society-wide effects.

How can AI make tribalism worse?

AI can lower the cost of tailoring persuasive messages to each faction, generate the appearance of widespread agreement, flatter users by confirming their views, and encourage people to outsource judgment. Each of these tends to follow from engagement-driven business incentives.

What is bridging-based ranking?

Bridging-based ranking shows content only when people who usually disagree both rate it helpful. X's Community Notes, which grew out of the Birdwatch pilot launched in 2021, uses this approach, which rewards agreement across divides instead of loudness within one group.

Does a healthy civic culture depend on technology?

No. Technology can lower the cost of good habits, but a post-tribal culture also needs in-person shared institutions, norms that make changing your mind respectable, and citizens who can tell truth-seeking from group protection.