No capabilities papers
Ban on publishing frontier AI capabilities research. No publishing research that makes frontier AI more capable. Safety research stays open.
What it does
No one in the U.S. may publish research that meaningfully advances frontier AI capabilities, such as new ways to train powerful models with less computing power. Safety research stays open. MIRI proposes this as part of an international deal, so compute limits can't be undone by better algorithms.
Applies to: Research that meaningfully advances frontier AI capabilities, published by anyone in the U.S.
- No one in the U.S. may publish research that meaningfully advances frontier AI capabilities.
- A government office decides which research counts, with an appeals process.
- Safety, interpretability and evaluation research stays open.
Where things stand
Publishing AI research is legal and unrestricted. Frontier labs already withhold key details voluntarily, as OpenAI did with GPT-4's architecture and training in 2023, while academics and several Chinese labs publish openly.
Why it costs ~9 days
The leading labs already keep their key results secret, so the ban mostly hits academics and smaller labs. China also loses what it learns from U.S. papers, which cancels most of the cost. A year in, about 3 days.
Biggest unknown: How much frontier progress still depends on published research, now that the leading labs publish little, and whether a ban could survive First Amendment challenges.
Why it lowers p(doom) by ~0.08%
Slows how fast techniques for building more powerful AI spread, which MIRI argues is needed to keep any limit on computing power enforceable.
Algorithms keep getting cheaper to run, which erodes any compute limit over time. Keeping new techniques unpublished slows that and narrows who can build frontier AI. On its own the effect is small; it matters most alongside a compute cap or a treaty.
The strongest case that it costs more
Most frontier progress already happens inside a few labs that don't publish, so a U.S. ban would mainly hit academics and safety researchers. It faces serious First Amendment problems, and it would do little to China, whose labs are now among the most open publishers.
The debate
For
- MIRI's Technical Governance Team drafted an international agreement with a section on restricted research that bans work advancing toward superintelligence, citing algorithms getting about three times more efficient each year, 2025.
- Eliezer Yudkowsky and Nate Soares argued in their book If Anyone Builds It, Everyone Dies that research making powerful AI cheaper to build, and publishing it, should be outlawed under an international regime, 2025.
- MIRI said that preventing superintelligence may in the long run require restrictions on certain research, which the Ban Artificial Superintelligence Act does not include, 2026.
Against
Sources
Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.