By Aaron Scher; endorsed by Bourgon, Soares, and Yudkowsky on behalf of MIRI.

MIRI has been warning about the extinction threat from superintelligent AI for over two decades. Only recently has this danger become known in the policy world, and the proposed policies for dealing with the threat have to date been piecemeal and insufficient.

The Ban Artificial Superintelligence Act of 2026 is the first piece of legislation we’ve seen that stands a chance at stopping this threat. The Act is excellent but not perfect, and we discuss both what it gets right and what we'd tweak. We hereby endorse the Ban Artificial Superintelligence Act of 2026 because it directly confronts the extinction threat that humanity is facing and would codify the primary policy goal we think the world needs: a ban on the development of superintelligence.

What we like about the Act

Banning artificial superintelligence (ASI), or variants of such a plan, is the only effective solution to avoid the ASI threat, at least in the near term. Most other legislative proposals do not confront this threat head-on and thus would not be effective, even if implemented. For more on why we believe this, see Chapter 13 of If Anyone Builds It, Everyone Dies.
Banning ASI requires intervening before ASI is created at all. To this end, the bill requires that AI training pause if a model develops precursor characteristics, such as the capacity to automate AI R&D or the capacity to avoid human shutdown attempts. This approach is good because it avoids dancing too close to danger and it intervenes on one of the likely pathways to ASI: AIs automating AI development.
This bill includes a training compute threshold above which AI development requires a charter. Training compute thresholds are often a good tool for AI governance because they are measurable ahead of time (unlike evaluating the capabilities of a trained AI model). A major concern with training compute thresholds is that continued progress in AI algorithms and data means the same capability can be reached using less and less compute over time. The bill handles this issue by enabling the threshold to be adjusted to account for these more efficient methods, which is a good choice. It would be an improvement to allow these thresholds to be adjusted more frequently than annually, given that algorithmic progress can be very rapid.
We are glad to see the bill ...

DanielSP

DanielSP