# Demis Hassabis 支持 AI 预飞安全测试

- 来源：Gary Marcus：The Road to AI We Can Trust（RSS）
- 作者：Gary Marcus
- 发布时间：2026-07-14 22:30
- AIHOT 分数：48
- AIHOT 链接：https://aihot.virxact.com/items/cmrkt578k00yibi5qepu5wjl4
- 原文链接：https://garymarcus.substack.com/p/breaking-demis-hassabis-endorses

## AI 摘要

DeepMind 联合创始人兼 CEO Demis Hassabis 公开支持对 AI 系统实施“预飞安全测试”（preflight safety testing），即在部署前进行类似航空业的安全检查。这一立场与当前业界对 AI 安全监管的讨论相呼应，强调在模型发布前通过严格测试来降低潜在风险。Hassabis 的背书为推进 AI 安全标准化提供了重要行业支持。

## 正文

Good news, for once.

Jul 14, 2026

I am genuinely excited.

In my 2023 Senate testimony, in a 2023 essay here, and in my 2024 book Taming Silicon Valley, I strongly advocated for a system of preflight testing for AI models deployed at large scale, ideally mandatory, transparent and independent, modeled perhaps on the FDA’s analysis of costs and benefits for pharmaceuticals. When Senator Kennedy asked for my most urgent suggestion for AI policy, preflight testing was it.

At the time, back in 2023, the Senate seemed receptive. But nothing happened, and things changed markedly when Trump came to office. As recently as a few months ago, getting the US to implement preflight testing seemed hopeless, given a great hostility to AI regulation that was then widespread in Washington.

But the Mythos moment got The White House to consider and in fact implement preflight testing to some degree (though not with the transparency or independence or breadth or mandatoriness that I believe to be critical).

Now, I am thrilled today to report that Google DeepMind’s CEO Sir Demis Hassabis has just come out strongly and publicly in favor a version of preflight testing.

Hassabis adds that

Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release. Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market. Labs would also work with the Standards Body to address any critical post-release vulnerabilities.

I am especially pleased by his reference to independent leading technical experts; it cannot be just the government and big tech companies that make these decisions, particularly because of the potential for conflict of interest in a time in which the US government is contemplating take a stake in AI companies.

You can read Hassabis’s full essay here. The FINRA model that he suggests is excellent, and the world will be a better place if it is implemented, with transparency and independence. I hope that his brave essay will be a turning point.

Discussion about this post

Problem being that transparency is hard. If the protocol is formalized ahead of time, then LLM vendors will train their models to detect when they are being tested, and to behave accordingly. See the Volkswagen Dieselgate scandal, where precisely this happened: https://en.wikipedia.org/wiki/Volkswagen_emissions_scandal. So the tests will necessarily be opaque and adaptive - but one person's "opaque and adaptive" is another person's "unaccountable and corrupt".

Transparent theft is still theft.

Ready for more?