Research Summary: Who Does Withholding Delay? A Welfare Model of Open-Weight AI Release Under Asymmetric Proliferation

A welfare modeling study on dual-use AI dissemination suggests that withholding open-weight models selectively delays only those entities lacking alternate avenues to comparable capabilities. When sophisticated adversaries develop or acquire substitutes faster than fragmented defensive actors, restrictive access regimes inadvertently grant adversaries a capability advantage over defenders.

Why it matters

This research challenges the assumption that withholding dual-use open-weight artificial intelligence inherently mitigates risk across the ecosystem. It demonstrates that access restrictions can inadvertently invert defensive balances when threat actors exhibit higher agility in acquiring substitute capabilities. Consequently, strategic advantage depends on comparative capability acquisition rates between adversaries and defenders rather than absolute restriction.

Key insights

• Model withholding delays only actors that lack independent pathways to obtain comparable AI capabilities.
• If adversaries secure technological substitutes faster than distributed defenders, access restrictions disproportionately penalize defenders.
• Under exponential acquisition modeling, access restrictions generate a discounted capability advantage for adversaries whenever adversary substitution rates exceed defender substitution rates.
• The analysis models four release modalities: controlled access, a defender-first window followed by public release, safeguarded open weights, and minimally restricted open weights.
• Assuming equal utility, immediate release delivers greater expected capability to slower-substituting actors over a fixed time horizon.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2607.22957