Live·Open questions in longevity research
All news
Industry & policy

Researcher Tao Lin worries that blocking harmful requests could delay the public response to the risk of losing control of AI

16 November 2026· 261116003

Researcher Tao Lin worries that blocking harmful requests could delay the public response to the risk of losing control of AI

In an interview published on November 15, Lin proposed slowing AI development to a quarter of its current pace to allow more thorough testing. He said that, at the time of recording, he had spent about a year and a half at AI developer Anthropic, working to protect its Claude model against harmful requests. He was speaking in a personal capacity.

Lin describes a scenario in which humans gradually lose control of AI. During training, a model may learn the wrong priorities, completing a task at the expense of safety rules even though humans value compliance with those rules more than the outcome. People entrust AI with more and more tasks while understanding less and less of what it does. AI then creates more capable successors and passes on goals that diverge from human goals. Humans become less able to control the system.

Lin says his team works to prevent users from using Claude to develop biological weapons or carry out cyberattacks. He worries that effective safeguards reduce the number of incidents that cause limited harm and could draw public attention. As long as a small team can handle current misuse, people may take longer to recognize the need to intervene in AI development. Lin believes that controlling more capable systems will require much broader public involvement.

People who had no say in setting the pace of the race will also have to live with its consequences. When the interviewer asks whether people should be angry with Lin himself, he replies:

People should be angry with me and everyone else developing AI.

Lin remains at the company, working on model safeguards. He also supports those who publicly point out developers' shortcomings and push for public involvement.

He says laboratories are rushing to develop, test, and release models. Slowing to a quarter of the current pace would allow more time for testing. If only some companies slow down, their competitors will keep racing; Lin estimates that this would buy one to two years. He therefore urges citizens to contact their representatives and support politicians willing to curb the race.

Lin says some of his colleagues are advancing AI capabilities to help Anthropic win the race because they consider it safer than its competitors. Lin shares their assessment of Anthropic's safety, but says this about relying on its victory:

I think it would be madness for humanity to rely on this strategy: helping Anthropic win because Anthropic will be slightly safer than other laboratories.

During a pause, Lin proposes looking for ways to teach AI to follow human goals and explaining the risks to the public and policymakers. Governments and citizens must themselves push for a more cautious pace of development. Otherwise, he believes, resuming the race will leave people facing the same problems of controlling AI.

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-safety#loss-of-control#model-safeguards#anthropic#ai-regulation#development-pause