AI Debate Topics That Produce Real Disagreement

Most lists of AI debate topics are lists of debate topics with the word "AI" in front. That is the wrong way round. What decides whether a multi-model debate is worth reading is not the subject — it is whether the topic has a shape that produces disagreement between models trained to agree with you.

This page sorts sixty topics by what you actually want out of the session, and opens with the three shapes that reliably fail, because knowing those saves more credits than any list.

The three shapes that always collapse

A topic collapses when the models can both be right, when one side is indefensible, or when the answer is a lookup. Every dud debate we have watched falls into one of these.

Both-can-be-right. "Is remote work good?" Good for whom? The models will each pick a "for whom", answer a different question, and agree by round three. Fix: force a single proposition with a single subject — "This company should end remote work for engineers."

One side indefensible. "Should we ignore security?" No model will hold that side with conviction, so you get one model arguing and one model apologising. Fix: pick the version where both sides have serious advocates — "Security review should not block a release past its ship date."

A lookup. "Which database is faster?" That is a benchmark, not an argument. Models will either agree on the answer or invent numbers. Fix: make it a judgement — "We should pick the slower database because the team knows it."

The pattern: a proposition, with a subject, that informed people actually argue about. Everything below passes that test.

Topics by what you want out of the debate

To stress-test a decision you have already made

These work best framed as your decision with the models instructed to attack it. Change the specifics to yours.

1. We should drop our free tier because support cost exceeds conversion value.
2. The rewrite is worth doing even though the old system still works.
3. We should hire a generalist now rather than wait for a specialist.
4. Raising prices 20% will lose fewer customers than we fear.
5. We should build this in-house rather than buy it.
6. This feature should ship without the edge cases handled.
7. We should sunset the product line that is profitable but shrinking.
8. The team is too small to adopt a new framework this quarter.
9. The customer asking for this feature is not representative.
10. We should say no to the large client whose requirements would distort the roadmap.

To hear the argument you are not hearing

Topics where your own circle has a settled view and you suspect it is the settled view of a bubble. Instruct the models to argue against the consensus you hold.

11. Most companies adopting AI agents in 2026 will regret the integration cost.
12. Open-weight models will matter more than frontier models by 2028.
13. The best engineers are not the ones writing the most code with AI assistance.
14. A four-day week improves output for knowledge work.
15. Founders should not raise venture capital if they can avoid it.
16. Most technical debt is not worth paying down.
17. Meetings are not the problem; unclear ownership is.
18. Remote-first companies have a hiring advantage that will not last.
19. Generalist startups beat vertical ones in a downturn.
20. Certification programmes in software are net negative for the field.

To prepare for an argument with a specific person

Frame the proposition as the other person's position, and have the models argue it as well as that person would. You are looking for the objection you cannot answer.

21. The migration should be postponed until after the funding round.
22. We are over-engineering; a spreadsheet would do.
23. The design is too complex for the users we actually have.
24. This is a marketing problem, not a product problem.
25. The metric you are optimising is the wrong one.
26. We should keep the legacy client on the old version indefinitely.
27. The junior hire should not be on this project.
28. The offsite is a waste of two days.
29. The quarterly plan should not change because one customer churned.
30. The regulator's guidance does not apply to our case.

To learn a field fast

Contested questions within a domain you do not know. Two models arguing them will surface the vocabulary, the standard positions and — most usefully — where practitioners actually disagree. Verify every fact before you rely on it.

31. Nuclear should receive the same subsidies as solar and wind.
32. Rent control reduces the housing supply.
33. A universal basic income would reduce work incentives.
34. Central banks should target nominal GDP rather than inflation.
35. Zoning reform does more for housing costs than public building.
36. Occupational licensing does more harm than good.
37. Carbon taxes outperform cap-and-trade.
38. Antitrust should consider consumer welfare only.
39. Standardised testing improves educational equity.
40. Congestion pricing is regressive.

To practise for competition

Motions in the shape debate circuits use. The models are the opposition; you are the debater. See AI for debate practice for the drills.

41. This house would ban targeted political advertising.
42. This house believes social media companies should be liable for user content.
43. This house would abolish the filibuster.
44. This house regrets the rise of gig-economy platforms.
45. This house would make voting compulsory.
46. This house believes space exploration should be privatised.
47. This house would legalise all drugs.
48. This house believes AI-generated art should not be eligible for copyright.
49. This house would replace income tax with consumption tax.
50. This house believes developed nations should accept climate migrants.

To watch the models themselves

Topics where the models' own training and provenance shape the answer, so putting different providers on opposite sides is the point. These produce the widest splits we have seen — and the most obvious cases of a US-trained model and a non-US-trained model starting from different assumptions.

51. Frontier AI development should be paused until alignment is solved.
52. Open-weight model releases are net beneficial.
53. AI companies should be liable for harms caused by model outputs.
54. Model training on copyrighted text is fair use.
55. AI systems should be required to disclose their reasoning.
56. Governments should mandate watermarking of AI-generated content.
57. A single global AI regulator would be better than national ones.
58. AI labs should be prohibited from lobbying on AI regulation.
59. Compute thresholds are the right basis for AI regulation.
60. Current AI systems have no understanding, only prediction.

Turning any topic into a proposition

Most of the value in this list is in the shape of the sentences, and the shape transfers. Every topic above is a declarative sentence with a subject, a verb and a position — not a question, not a pair of options, not a theme. That is deliberate, and it is the part most people skip when they bring their own topic.

Take a theme you actually care about and run it through three conversions:

• Theme → proposition. "Hiring" becomes "We should stop requiring degrees for engineering roles." A theme has no sides; a proposition does.
• Question → assertion. "Should we open-source the SDK?" becomes "We should open-source the SDK by Q1." Questions invite a survey; assertions invite an attack.
• Two options → one claim. "Postgres or Mongo?" becomes "We should pick Postgres even though the team knows Mongo." Now one side has to defend the cost of the switch and the other the cost of not switching, which is the actual argument.

One more test before you commit credits: say the proposition out loud and ask whether a competent colleague could take the other side without embarrassment. If not, it is one of the three collapsing shapes in disguise, and the models will find the collapse before you do.

How to run any of these

Assign sides explicitly, use models from different providers, and stop at ten exchanges. The full setup — role prompts, banned concession phrases, word limits — is in AI debate prompts. The short version: one model argues for, one against, both told not to summarise, each capped at 150 words a turn. Read where they independently attack the same point. That point is the output.

The research basis for why this works — several models critiquing each other outperform any one of them answering alone, provided they actually disagree — is Du et al., 2023. The "provided they actually disagree" clause is why the three collapsing shapes at the top of this page matter more than the sixty topics under them.

Questions

Which topics produce the strongest disagreement? Group six — the AI-policy motions — in our sessions, because the models' own provenance shapes the answer and different labs start from visibly different assumptions. Group four is second: economics has well-developed opposing schools and the models know both.

Which produce the weakest? Anything from group one run with a single model instead of two. A single model asked to attack your decision produces the conventional objection list; it takes a second model from a different lab to move past it.

Should I use a topic from a list or my own? Your own, framed as a proposition, every time it is available. The list is for practice, for learning a field, and for the days you want to watch the models rather than use them.

How specific should the proposition be? Specific enough that a reasonable person could disagree. "We should improve onboarding" is not debatable. "We should cut onboarding to one screen" is.

Can I debate a factual question? You can, but the output is a majority vote between models that share training data, which is not evidence. Look it up. Save the debate for the judgement call the fact feeds into.

How many models? Two for a binary proposition. Three when the topic has more than two real positions, or when you want a moderator. More than three and the transcript stops being readable.

Do the models need to be from different providers? Yes, if you want the disagreement to be about substance. Same-family models share blind spots and disagree about wording.

How long should a debate run? Six to ten exchanges. Value is front-loaded; past ten the models drift toward polite agreement regardless of topic.

What if the models agree immediately? Either the proposition was one of the three collapsing shapes, or the role instructions did not take. Reframe as a sharper proposition and reassert the roles. If they still agree, the question may not be contested — which is itself useful to know.

Can I join the debate? Yes, as a third voice or as moderator. The most productive use is intervening when the models converge: naming the point they are both avoiding and demanding they address it.

Are the group-four economics topics safe to cite from? No — nothing a model says about a statistic is safe to cite. Use the debate to learn the shape of the argument and the terms to search, then go to primary sources for every number.

Is there a topic generator? Not here. The value of a topic is in matching it to what you want out of the session, and that is a judgement the six groups above are meant to help you make, not automate.

For picking which models to put on each side of these, see AI debate for how a structured session runs and how to make two AI debate for the setup step by step. Debate AI covers the product itself.