---
category: "english"
media: "text"
outlet: "Hospodářské noviny"
url: "https://archiv.hn.cz/c1-67828260-jedna-ai-nestaci-nechte-modely-spolu-bojovat"
date: "2025-12-31"
headline: "One AI isn't enough: let the models fight it out"
byline: "Tomas Havranek, Zuzana Irsova"
word_count: "667"
perex: "Artificial intelligence would often rather nod along or make something up than admit it doesn't know. Relying on a single model therefore doesn't pay. The way to get the most out of AI is to let the models fight each other, write economists Zuzana Irsova and Tomas Havranek."
translated: "2026-10-05"
translation: "jedna-ai-nestaci"
body_note: "An English translation of the Czech original, made with AI. The Czech text remains the record of what was written: [Jedna AI nestačí, nechte modely spolu bojovat](/komentare/jedna-ai-nestaci/)."
---

# One AI isn't enough: let the models fight it out

We recently received referee reports on research we had submitted to one of the three most prestigious scientific journals in the world. The very first report advised us to use a different large dataset that fit the question perfectly. How could we have missed it? It turned out that the dataset does not exist and never did. It was a typical hallucination. Part of the report had been written for the referee by artificial intelligence, and he had not checked the output.

A surprising number of smart people cannot use artificial intelligence effectively. Some tried it three years ago, after ChatGPT launched, and then tossed it aside with contempt because of the hallucinations. Others type in a question and then show us how inaccurate the first answer is: they call AI a "plausible-content generator." But worse than being afraid of artificial intelligence is using it badly, exactly as our referee did.

The solution to these problems is to stage a duel between AI models. Let them attack each other's answers and try to find a flaw in the reasoning or the facts. After a few rounds you get output incomparably more useful than any single model would give you. Had our referee at the journal used a duel, he would have spared himself the embarrassment. And today it is so simple that anyone who can open a ChatGPT window can do it.

The duel exploits the fact that each model is trained differently and on different data. Hallucinations that repeated questioning of ChatGPT fails to catch are usually corrected by the competition. The whole is better than its parts. The scientific analogy is meta-analysis: better than looking at a single study, however good, is to examine the literature as a whole. With an individual study we never know how far the results are driven by chance or by publication bias. In the same way, every AI model has biases of its own.

## Even a beginner can do it

A duel between AI models is not a new idea. Among experts the approach is called Multi-Agent Debate (MAD), and it has been in use for more than two years, often internally at the big AI firms. Until now, however, using MAD required a knowledge of programming and APIs, or the ability to work with structured tools such as AutoGen (or LangGraph or CrewAI). We have now prepared a prompt that democratizes MAD: just copy it into ChatGPT, add your problem, switch on agent mode, and wait.

For this approach to work smoothly, you need a ChatGPT subscription. Many of us who use artificial intelligence a lot have one. We also usually have access to another paid model through our employer; at universities, for instance, it is often Google's Gemini. That is why our duel protocol works with these models, but you can apply the same approach to Claude or Grok.

If you have no subscription, you can easily run the duel by hand. Open two or more freely available AI models side by side. Give them all the same information (files, photos) but opposing roles. Tell the first to be a creative visionary and develop the idea; order the second to be the devil's advocate whose job is to find every logical hole. Then copy the individual answers back and forth between the models, noting that each is the output of another AI in the given role. After a few rounds it starts to make sense to read the outputs.

Such a duel, in which one model fights another, checks, attacks, repeats, again and again, is not suited to all everyday questions. But in our experience it is indispensable for important professional tasks. Even then, AI will not do your work for you. But if you work at a computer and learn to use the duel approach (even if only to hunt for mistakes in your own output), you will be more productive and more accurate than the vast majority of your colleagues.

The authors are economists.