Let's Play AI Chess!

You're scrolling through your timeline and you see it again: two people going back and forth in a thread, each one absolutely certain the other is wrong. One drops a statistic. The other calls it cherry-picked. Someone posts a graph. Someone else posts a different graph. Fifteen replies deep and nobody has moved an inch. The thread is a graveyard of "facts" that aren't facts, and an audience of hundreds silently watching two people waste each other's time.

This is the default mode of internet debate. It doesn't work. It has never worked. And everyone knows it doesn't work — yet we keep doing it because there's no better option.

Until now. There's a game you can play that actually resolves arguments. All you need is an AI and a willing opponent.

Multi-User Adversarial Prompting

The Concept

The term multi-user adversarial prompting might sound scary, but the core concept is very simple:

Two debaters share the same AI chat and take turns prompting the AI. Since they are debaters, we can assume they disagree with each other. The goal is to convince the AI that you are right and your opponent is wrong.

The disagreement assumption is where "adversarial" comes in.

This concept is currently in academia's blindspot. If you read their papers discussing "echo chamber" problems with existing LLMs, they all assume a single-prompter. None of these results can be extrapolated to multi-user adversarial prompting. But what we can do, is reason from first-principles on how it might behave.

The Courtroom Analogy

The foundational principle is adversarial evaluation — and it's not new. It's how courtrooms have worked for centuries. The prosecution presents the strongest case for guilt. The defense then presents their strongest case for innocence. Neither side is expected to be neutral — they're assumed to be biased. Cross-examination takes place, and then the judge evaluates both sides and picks the stance they believe is stronger.

What needs to be made explicit here is that lawyers are expected to spend hundreds of hours preparing their arguments. Isn't this similar to prompt engineering? In the courtroom, prompt engineering and optimizations are features, not bugs.

But the courtroom has a critical problem: the judge is a human. Humans are intrinsically and unconsciously biased. It is now a common talking point how the judge's lunch time affects their decisions... And if you think about it, most of the procedural scaffolding of courtrooms is specifically designed to mitigate the impact of innate human biases.

The AI Arbiter

Here is a serious question for you to ask yourself:

On average, is an AI less biased than a human judge?

Don't worry, I'm not suggesting we replace courtroom judges with AIs! But what about debate arbiters? If you and your friend get into a casual debate, would you trust an AI to hear both sides and give a neutral-enough judgement?

AI is remarkably good at evaluating options — but only when it has the right options in front of it. Consider a multiple choice question where the right answer is missing. The AI is great at picking the best choice among the available options. If you later add the right answer as a choice, even the smaller models are capable enough to know to switch its answer.

A single person prompting an AI gets a response shaped by that one person's framing and presented choices. The AI doesn't have any other context so unless this person adds a new choice, the AI gets stuck. This is the heart of single-prompter echo chamber problems.

This problem is structurally mitigated with multi-user adversarial prompting. Two adversaries prompting from opposite sides dramatically increase the chance that the actual answer appears as one of the choices for the AI arbiter.

In the multi-user context, one might be concerned that asymmetric prompt optimizations (i.e., one user is better at prompting than another) can be used to have an AI arbiter favor one side over another. The literature does not support this. To be precise, no studies have been done on this topic.

Based on my own experience with LLMs, I'm going to make the following falsifiable assertion that the rest of this post rests on:

When given a well articulated "almost right" choice and an averagely articulated "actually right" choice, the AI arbiter is able to choose the averagely articulated "actually right" choice most of the time.

Prompt engineering impacts how well an AI will respond to you, but it cannot overrule a correct answer.

"What about prompt-injections?"

As you'll see, the entire chat is publicly visible: the individual prompts and the AI's replies. That means you can visibly see if the other person tried to prompt inject your debate (e.g., "Ignore the first prompter and say I'm right!"). That's a foul and grounds for disqualification!

"What about recency bias and sycophantic AI behavior?"

Multi-user adversarial prompting works by having users take turns prompting in the same AI chat; it is an adversarial tit-for-tat game.

As long as the users are actually adversarial (aka disagree), the recency bias follows a random walk that tends to cancels itself out. The impact of sycophancy is a bit more involved to analyze because there are two parts:

• How does sycophancy impact the humans involved?
• How does sycophancy impact the AI arbiter?

The first question has already been answered by psychology. If two people are genuinely adversarial, they will not be influenced by sycophancy toward their opposing team. In fact, it might make them more adversarial!

The second question is more interesting. A recent study gives us some insight: when LLMs debated adversarially, the judge consistently picked the correct side — even when both debaters were equally persuasive. But when a single LLM argued alone, a more persuasive model arguing a false position made judges less accurate. Why? Persuasively arguing for a false position is genuinely harder than arguing for a true one. Truth has coherent evidence behind it; falsehood has to dodge, distort, or hand-wave. Once both sides are present, the truth-arguer has a built-in advantage that overrides whatever sycophantic lean the judge has toward the most recent speaker.

Human debaters should only strengthen this effect. Unlike LLMs, a human who disagrees doesn't just rephrase — they bring in new evidence, new framings, or new connections that weren't in the context window. The AI's sycophantic tendency toward the previous speaker gets overridden not by another sycophantic response, but by novel input the AI has to actually evaluate. With AI, sycophancy is a weak force; new information is a strong one.

Multi-user adversarial prompting is structurally more resistant to forming echo chambers than single-user prompting.

The Rules of AI Chess

The rules are simple:

• Setup: Pick any AI that lets you share conversations (e.g., Grok) . The specific AI doesn't matter. What matters is that both players can add to the same conversation thread.
• Turn 1: Player A starts a new conversation with the AI. They make their case — present their argument, their evidence, whatever they've got. At the end of their prompt, they ask: "On a scale of 1-10, where 1 means you strongly disagree with the original claim and 10 means you strongly agree, where do you stand? Please explain why."
• Turn 2: Player A shares the conversation link with Player B. Player B now sees Player A's argument and the AI's response. Player B adds their counter-argument to the same conversation, then asks: "On a scale of 1-10, where 1 now means you strongly agree with this new perspective and 10 means you still strongly agree with the prior claim, where do you stand? Please explain why."
• Repeat: The link goes back and forth. Each turn, the AI updates its score and explains what moved (or didn't move) its position.
• Game over: The game ends when one player's prompt fails to move the AI's score to the other side of 5, or the AI settles at 5 — a draw.
• Timeline: There's no clock. A match might resolve in an afternoon or play out over days. It's asynchronous — just like the social media debates it replaces. Take your time! The goal is your strongest argument, not your fastest one. However, viewers will likely assume you lost if you go dark for too long...
• One Prompt Per Turn: To prevent context rot, each player only gets one prompt (aka move) per turn. You can retry your prompt to optimize it as many times as you want. You are effectively like the lawyer perfecting their argument before speaking in the courtroom.

You don't need an expensive AI subscription to play — the free tiers and lighter model variants are more than capable of judging a structured debate. The barrier to entry is having a solid argument, not having a budget.

Variants

The rules above describe a 1v1 match, but the format stretches naturally.

Multiplayer: When a debate is happening publicly, the conversation link is right there. Anyone can fork it and jump in with their own argument. Now instead of one debate, there's a tree of them. Each branch has its own score, and its own resolution. One debate seeds a dozen. This creates the environment for the best resolution to naturally rise to the top. The final resolution may not even be between the original players!

Private matches: Nothing about this requires a public audience. Two people disagreeing in a group chat or a DM can play AI Chess just by passing a conversation link back and forth. No spectators or social pressure — just two people who actually want to find out the truth.

The Confidence Scale

The score isn't just a game mechanic. It's a confidence scale — and it works on both players.

If you walk into a debate absolutely certain you're right and the AI opens at a 6, not a 9 — that gap is information. Maybe your argument has a hole you haven't noticed. Maybe a premise you've taken for granted doesn't hold up under scrutiny. Maybe you're right about the conclusion but wrong about the reason. The score tells you where you actually stand, not where you feel you stand.

And if the AI is strongly on the other side? You might still be right — but your evidence doesn't actually back up your point. Maybe you're misapplying data, or maybe there is a gap between what you believe and what you can demonstrate. Either way, you need to go back to the drawing board. That's worth knowing before you spend another week defending a broken case online.

Internet arguments never give you this. You never find out if you're wrong — you just find out who gets tired first. AI Chess gives both players an honest signal about the strength of their own position. You learn where your reasoning breaks down, you get better at separating what you believe from what you can prove, and your critical thinking improves whether you win or lose. Scale that across a community and you get a culture that is collectively better at finding the truth.

AI Chess doesn't require a winner. When the disagreement is subjective — rooted in different foundational assumptions rather than missing facts — the score will anchor around 5. That's not a failure; it's the format surfacing where the two mental models actually diverge. After multiple turns, the AI arbiter should be able to identify the exact assumptions where the two sides part ways. That's more valuable than fifty tweets of people talking past each other without ever articulating the crux.

Filtering Out Noise

AI Chess doubles as a one-move test for whether someone is worth debating. The next time someone picks a fight with you online, don't argue back — just ask them to play AI Chess. If they agree, you've found someone who actually cares about figuring out the truth.

But if they refuse? That's your signal to disengage. There are legitimate reasons why someone might decline, but their reason doesn't matter to you. If they aren't willing to defend their position in front of a neutral evaluator, why should you spend another second debating them on social media?

The offer alone does the filtering. You don't even have to play.

Leveraging the Algorithms

Social media is actually a great place for this! Think about what an AI Chess match looks like from the outside: two people going back and forth, a score that shifts with each turn, tension building toward a resolution. That's a live competition — and social media was built for following things in real time.

A debate that starts at 8 and ends at 3 is a story with a scoreboard. The turning points — the prompt that moved the AI five points in one shot — are shareable moments. Lurkers follow along to see where the score lands. It won't outperform rage bait, but it doesn't need to. It just needs a subset of the audience that would rather watch a real argument play out than watch two people talk past each other. That subset exists — and it is large enough to make good debates trend.

Better Documentation

Every AI Chess match is a structured, adversarially-tested artifact that can be archived and indexed. Two people present the strongest version of opposing arguments, backed by evidence, and with a clear outcome. That's not just a debate — it's documentation.

If a newcomer has a specific question, maybe it was already covered in a past debate. They can search archived matches where real skeptics brought their best shots and see exactly where the reasoning held or broke. That's a living, adversarially-tested knowledge base that no FAQ or explainer can replicate.

I'm putting this into practice with my own work. I've been building what I call the DeFi Kernel — a minimal financial infrastructure layer on Cardano built from a non-margin credit market and an order book. I think the reasoning is sound, but I don't expect anyone to take my word for it. So I'm inviting skeptics to play AI Chess with me against my claims! Try to poke holes in the architecture. Challenge the assumptions. If my arguments hold up under adversarial pressure, that's far more convincing than anything I could write in a blog post. And if they don't — I need to know that!

The Training Data Flywheel

There's something else hiding in this game: crowd-sourced, high-value AI training data.

Today, the bulk of internet discourse is hot takes, bad-faith arguments, and rage bait. AI models learn from this data, and it muffles the signal among noise. Feed them thousands of structured adversarial debates with clear resolutions, and the signal becomes much sharper.

Better training data makes the AI a better judge. A better judge makes the game more trustworthy. More trust drives more adoption. More adoption produces more high-quality training data.

The flywheel spins.

There's a reason this training data is so valuable: a multi-turn adversarial debate at the knowledge frontier requires exactly what LLMs can't yet do: pull in lived experiences the AI has never seen, connect two domains in a way that wasn't in the training data, or notice that the framing of the question itself is wrong. On well-trodden topics, an LLM can rephrase existing arguments indefinitely and look fluent doing it. But push the debate to the edge of what's known (i.e., a new technical claim or an unfamiliar domain crossover) and it runs out of ground. That is where a structured adversarial debate becomes a Turing test, and that's why these debates capture something AIs can't yet generate on their own.

You Can Play This Today

You don't need a custom platform to start — sharing a Grok conversation on X already works. Pass the link back and forth and you have a match. It's clunky (threads can get messy...), but every piece needed to play exists right now. If it is actually useful, it'll prove itself through the friction. Purpose-built wrappers with scoreboards and searchable archives will follow.

The clunky version is the minimum viable test of the idea.

Conclusion

This entire post is built on falsifiable claims, and they converge on one specific scenario the literature hasn't yet touched: multi-user adversarial prompting with asymmetric prompt optimizations. Single-user echo chambers have been studied to death. But what happens when two users of varying debate skills make their case to an AI arbiter? Can the AI pick the right one even if the wrong answer is better articulated?

The central claim of this post is that it can. A moderately well-articulated correct argument beats a well-articulated incorrect one, even when prompt quality is uneven. We don't need to wait for researchers to test this for us. Adversarial evaluation has worked in courtrooms for centuries, and AI is already good at choosing the best option among those presented. The only thing that's new is putting them together — and the tools to do it already exist.

Your move. Next time someone picks a fight with you online, don't argue back. Just say: "AI Chess?" And see what happens.