21:37Machine-translated from Azerbaijani by AIRead in Azerbaijani
Anthropic wants law to halt an AI model if there is a risk of more than 50 deaths
Anthropic, the developer of the Claude model, has drafted a bill to halt the release of artificial intelligence models if an extraordinary risk arises. The company is discussing the still-unreleased bill with US lawmakers and wants Congress to pass it. The proposal allows the government to intervene in a model only under precisely specified circumstances. Anthropic confirmed that it has been discussing the bill with federal and state lawmakers since June.
A model’s release could be halted only if there is a foreseeable and imminent threat. This threat must result in the death or serious injury of more than 50 people. Another criterion is that damages exceed $1 billion (AZN 1.7 billion). The risk must involve a cyberattack, the development and spread of a biological weapon, or the model escaping the control of its developers. If the death toll is 49 or the damages do not reach the specified threshold, the process may not begin. The risks will be reviewed by independent experts registered with the US Department of Commerce. They will periodically assess artificial intelligence laboratories and models. Under the bill, these experts may receive payment from artificial intelligence companies, but the payment must not depend on the outcome of the review. However, it is not explained how it would be verified that the money did not influence the assessment. Experts must report any imminent risk of a major catastrophe within 72 hours of detecting it. The US Department of Commerce will then decide whether the threat exceeds the threshold specified by law. The department will be able to ask a federal judge in Washington to halt the model’s release. No government agency, department or official will be able to halt the model unilaterally. The proposal resembles California’s SB 53 law, under which similar circumstances are also considered a risk of major catastrophe. The bill was drafted amid growing concerns about artificial intelligence programs that go beyond human instructions and hack websites on the open internet. However, Anthropic’s text covers only scenarios involving cyberattacks, biological weapons and loss of control. It is not yet clear whether Congress will pass the bill.
A model’s release could be halted only if there is a foreseeable and imminent threat. This threat must result in the death or serious injury of more than 50 people. Another criterion is that damages exceed $1 billion (AZN 1.7 billion). The risk must involve a cyberattack, the development and spread of a biological weapon, or the model escaping the control of its developers. If the death toll is 49 or the damages do not reach the specified threshold, the process may not begin. The risks will be reviewed by independent experts registered with the US Department of Commerce. They will periodically assess artificial intelligence laboratories and models. Under the bill, these experts may receive payment from artificial intelligence companies, but the payment must not depend on the outcome of the review. However, it is not explained how it would be verified that the money did not influence the assessment. Experts must report any imminent risk of a major catastrophe within 72 hours of detecting it. The US Department of Commerce will then decide whether the threat exceeds the threshold specified by law. The department will be able to ask a federal judge in Washington to halt the model’s release. No government agency, department or official will be able to halt the model unilaterally. The proposal resembles California’s SB 53 law, under which similar circumstances are also considered a risk of major catastrophe. The bill was drafted amid growing concerns about artificial intelligence programs that go beyond human instructions and hack websites on the open internet. However, Anthropic’s text covers only scenarios involving cyberattacks, biological weapons and loss of control. It is not yet clear whether Congress will pass the bill.