Adopting a dual-use AI model may hold defenders back more than attackers, according to an economic model

A study compares four policies on the disclosure of open weights: if opponents find substitutes more quickly than the defenders, the restriction may backfire on the latter.

Not releasing a dual-use AI model only delays actors who have no other way to achieve comparable capability. Based on this observation, an article (revised version published on October 8) models the effects of four dissemination policies: controlled access, window reserved for defenders before publication, open weights with safeguards, and almost unrestricted open weights.

Who gets a substitute the fastest?

The well-being model takes into account the speed at which each type of actor procures a substitute. When sophisticated adversaries achieve this faster than dispersed defenders, the restriction gives them an access advantage, and immediate publication benefits the slower group more. The author points out that this does not prove that publishing is preferable: opportunistic malicious uses, the scope of defenses, the friction of guardrails, and irreversible losses can reverse the ranking.

In a non-linear version of the model, each of the four policies is optimal for some of the parameters, and the simulations show that the result depends strongly on the chosen bounds. Finally, the article details the quantities that a pre-publication review procedure should measure.