“the mechanism, how we address those risks is still under discussion.”
Dario said the risks are real. When a reporter asked if companies policing themselves was enough, Trump interrupted and spoke for him. Dario did not answer that question.
Read his words · Watch the exchange
Yes. A known serious risk needs a fix before release.
Dario Amodei leads Anthropic.
What would make him stop unsafe work?
He says AI needs to slow down. His company keeps building stronger AI and now runs a biology lab. What changes because of his warning?
In Anthropic’s lab, scientists use Claude to help plan real experiments. This work aims to help science. The same company has reported possible use of Claude to help develop biological weapons. Source.
The company says people do all the lab work and it does not handle germs that infect humans. It says the research group began in spring 2026.
Anthropic’s lab announcement · September 23, 2026
“we need to slow down.”
Dario Amodei · September 13, 2026 · 3:13
Asked whether Anthropic will stop releasing stronger AI, he replies:
“It doesn’t mean that.”
Watch the question and answer · 3:48
He calls for a slower pace, not a halt. He says new models need tests. But what work will he slow, by how much, and starting when?
CBS interview · September 13, 2026 · 11:20
The interview raises the risk that powerful AI could get around attempts to shut it down. Dario describes several layers of safety checks.
That is a plan to reduce risk. It is not proof that people can stop the system. What happens if those checks fail too?
“very strong safeguards”
Dario Amodei · September 13, 2026 · 20:06
He says people mock him for making the checks too strict. Anthropic has described filters that block harmful requests, checks for misuse, and bans on accounts. It also publishes safety research and test results. Misuse report · Safety research.
Those are real steps. But strict filters are not proof that every dangerous use is stopped. What do the tests catch? What do they miss? Who outside the company has checked?
Which failed test would stop the work? Who has the power to enforce that decision? When will the public see the results?
His plan · September 13, 2026 · 4:10–6:30
“above my pay grade.”
Dario Amodei · September 13, 2026 · 17:53–18:09
He is talking about U.S.–China talks. Governments make those deals. He leads the company building Claude.
He says he is working with governments on safety rules. In August 2024, the U.S. AI Safety Institute announced an agreement to get access to Anthropic models before and after release, test risks, and suggest safety changes. Government announcement.
That is a concrete step. It does not tell us which release was stopped or which change was required. Which government checks changed Claude? Who can make Anthropic stop if it fails?
Trump dismissed AI concerns on September 13, saying they would not happen. Trump’s answer · 7:59.
What will Anthropic stop doing now, without waiting for a deal? Who is responsible if it keeps going and people get hurt?
“I don’t want people to feel powerless.”
Dario Amodei · September 13, 2026 · 21:07
He says people should have a say. In 2023, Anthropic asked about 1,000 Americans to help choose rules for a test AI model. Public input study.
That study asked for views. It did not give the public a vote on whether to release Claude or open the lab. His company keeps building while people outside it bear risks too.
What power will he give people over what his company builds? How can they say no—and make that decision count?
Replies by Build Something. Our plan still needs outside tests. Sources checked September 19, 2026.
No reply here yet.
Checked September 26, 2026. This does not mean a reply was refused.