Breaking

Vinod Khosla Rejects International AI Safety Treaties, Citing Verification Problem

September 21, 2026 at 18:31 EDT

Policy & Regulation
Foundation Models
Security

Vinod Khosla, founder of Khosla Ventures and one of OpenAI's earliest backers, has argued that international treaties are the wrong tool for AI safety because compliance with AI rules cannot be reliably verified. The remarks intensify a live policy dispute over whether frontier AI should be governed mainly through binding diplomacy, technical safety work or national competition.

Sept. 2026 · AI governance

Khosla rejects the treaty path

The debate is not whether AI needs safety work. It is whether international promises can be checked when powerful software can be copied, hidden and run across global infrastructure.

countries joined the Bletchley Declaration

initial members in the AI Safety Institutes network

$11M+

research commitments for synthetic-content risks

ratifications needed for the AI treaty to enter force

Coordination is growing

Each block represents about two participating governments or institutions.

Bletchley countries

AISI initial members

Why supporters agree

AI can be copied and hidden more easily than physical weapons, so broad promises may be weaker than safety engineering, compute controls and domestic capability.

Why critics push back

Verification may be partial and expensive, but data centers, specialized chips and model testing can still give governments observable targets.

The fault line is inspection

Compute monitoringData-center accessFoundation-model testsRisk notifications

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.