发生了什么
According to the BBC, Anthropic co-founder Jack Clark said companies may eventually need to be required to maintain AI “kill switches” that can be checked by independent third parties. Clark said most AI labs, including Anthropic, already have ways to shut down systems, but he did not describe their technical design or provide evidence of independent verification.
The BBC reports that Jack Clark, one of Anthropic’s seven founders, said society may eventually want rules requiring AI companies to have a way to shut off software completely if it becomes too dangerous. He framed the issue as part of a broader policy discussion rather than announcing a new Anthropic product or policy.
Clark told the BBC that “most labs have different ways of being able to pull the plug,” including Anthropic. He specifically questioned whether companies should be required to have a kill switch and whether that switch should be verifiable by a third party. The report gives no technical description of Anthropic’s controls, no independent assessment, and no proposed verification standard.
The comments came amid renewed public debate about catastrophic AI risks and after Anthropic chief executive Dario Amodei called for the pace of AI development to slow and for development to be more closely monitored. The BBC also notes that US lawmakers have proposed a Kill Switch Act that would require ways to shut down problematic AI tools and give certain government agencies authority to demand that tools be turned off or limited.
The report says US President Donald Trump has rejected attempts to slow AI development. It also notes that some industry figures believe fears about human extinction from AI are overstated or may be intended to generate hype. Those reactions provide context, but the BBC report does not resolve the underlying technical or policy questions.
为什么这很重要
A mandatory, independently verified shutdown mechanism would turn a largely internal safety practice into a possible regulatory requirement. The proposal addresses a practical question about control over increasingly capable AI systems, but the report does not establish whether existing kill switches would work against every failure mode, how quickly they could operate, or who would have authority to activate them.
The proposal is consequential because it focuses on an enforceable control rather than a general appeal for safer AI. If adopted, independent verification could require companies to demonstrate that shutdown mechanisms exist, are accessible to authorized parties, and cannot be quietly disabled. The source does not say whether any current system meets those conditions.
A kill switch would not by itself address every AI risk. The report does not establish whether a shutdown could stop copied model weights, systems operating across multiple providers, locally deployed models, or agents that have already taken external actions. It also does not specify how regulators would balance emergency intervention against misuse or unauthorized shutdowns.
The policy question is gaining attention alongside broader arguments over whether frontier AI development should slow. Clark opposed leaving AI as a “totally unregulated industry,” but the BBC reports no agreed industry position, government commitment, implementation timetable, cost estimate, or evidence that legislation will advance.
互动机制：它实际上是如何运作的
以交互方式探索这一发展背后的基础技术。
crm_get_transaction(id='4092').Why can ethical evaluation not be reduced to one model score?
接下来看什么
The next significant developments would be concrete legislative language, technical standards for verification, and evidence about how shutdown controls perform in realistic tests. It is also unclear whether policymakers would apply such requirements to model developers, deployers, cloud providers, or all three.
Watch for the text and progress of the proposed US Kill Switch Act, including which systems it would cover and which agencies could order a shutdown. The source does not establish that the bill has passed or that comparable rules exist elsewhere.
Technical standards will matter more than the label “kill switch.” Useful details would include independent testing, authentication of shutdown authority, protections against tampering, response times, audit records, and how controls operate when systems are distributed across cloud providers or connected tools.
Anthropic’s own disclosures could clarify what controls it currently uses, what threats they address, and whether outside evaluators have tested them. No such documentation or test result is provided in the BBC report.
The report links Clark’s remarks to a wider safety debate, but it does not independently confirm claims that AI could kill all humans or establish the probabilities cited by Anthropic researchers. Those risk estimates remain contested and should not be treated as findings of this report.