AIA forecast of new AI safety pledges—and later backtracking
A user predicts at least two frontier AI companies will make new unilateral safety commitments by the end of 2026, but expects few constraints on how systems are developed or released.
TLDR
In a September 18 forecast, a user predicts at least two frontier AI companies will announce new safety commitments by the end of 2026. They expect genuinely useful elements, probably around technical oversight and monitoring, but no commitments to tiered releases, delays after unsafe events or disclosure of safety-procedure failures. Any incident-disclosure provisions would likely be weak, they predict. Within a year, the user expects at least one company to amend its commitments because it believes it will soon violate them, and at least two others to violate the spirit of theirs. On collective agreements, the forecast anticipates eventual sharing of more information affecting companies’ bottom lines, but not novel safety risks or failures of safety protocols.
Combined views
42
1 Source, first seen 16h ago
A forecast of new AI safety pledges—and later backtracking
A user predicts at least two frontier AI companies will make new unilateral safety commitments by the end of 2026, but expects few constraints on how systems are developed or released.