A 10% AI Extinction Estimate Is a Warning, Not a Forecast
Anthropic（AN-throp-ik）
The company that makes Claude.
superintelligence（SOO-per-in-TEL-uh-jens）
A theoretical AI that could outperform humans across nearly all thinking tasks.
alignment（uh-LINE-ment）
Research on keeping an AI's behavior matched to human goals.
What happened
On September 9, 2026, Anthropic (the AI company behind Claude) became the focus of a safety debate. Evan Hubinger, Anthropic's Alignment Science Lead, posted on X that he personally put the chance of AI killing all humans above 10% within the next decade. CBS News and BBC News reported the statement. This is not a report of an extinction event. It is a public judgment about a possible future.
The warning followed Jacob Coxon's resignation from Anthropic. Coxon said he had spent three years doing pretraining research at OpenAI and Anthropic. He accused both companies of racing toward self-improving superintelligence despite serious risks. Hubinger supported the concern. He said Anthropic was trying its best, but did not yet have a plan for solving alignment for superintelligence. He also said the company was not clearly on track to solve it.
What does above 10% mean?
It is Hubinger's personal estimate. It is not a measured result, an official Anthropic forecast, or proof that extinction will happen. Ten percent can be pictured as one out of ten imagined futures: 10 × 10% = 1. But no one can run the same future ten times. The figure communicates seriousness, not a precise timetable or mechanism.
It also does not mean today's Claude has a ten percent chance of killing everyone. The reports focus on future systems with much greater abilities.
The background
Superintelligence is still theoretical. It means an AI that could outperform humans across nearly all thinking tasks. Alignment is the problem of keeping an AI's goals and behavior consistent with human intentions.
BBC reporting says Hubinger has described risk from present models as low. His concern is a future system that uses its abilities to help build a smarter successor. If progress outpaces human understanding, testing and control could become harder. That is a scenario, not a demonstrated outcome.
Why the story matters
The source of the warning matters. Coxon spoke after leaving. Hubinger leads alignment work at Anthropic and publicly backed the concern. Their statements show that serious safety questions exist inside the debate. They do not independently prove the probability.
The practical question is whether safety research can keep pace with capability research. It also raises questions about who should evaluate frontier systems and whether private labs can set the pace alone. CBS reported that a U.K. government spokesperson said AI risks cross borders and that the U.K. would continue testing advanced models.
What is confirmed
The public record supports three basic facts: Coxon announced his resignation; Hubinger published the above-10% estimate; and Hubinger said a superintelligence alignment plan was not ready. The reports do not show that Anthropic formally adopted his number.
What remains unknown
The reports do not explain how Hubinger calculated the estimate, which failure paths he included, or when he expects superintelligence. They also do not establish whether such a system will be built. Anthropic's detailed response and safety plan remain important next evidence.
Hacker News attention
Two Hacker News submissions in this cluster drew substantial discussion. The candidate list records 42 points and 96 comments for the CBS story, and 44 points and 97 comments for the BBC story. Those numbers measure community attention. They do not verify the articles or Hubinger's estimate. See the CBS submission and BBC submission.
What to watch next
Watch for a detailed response from Anthropic and OpenAI, independent evaluations of advanced models, and government coordination. The key question is not whether one alarming number becomes a headline. It is whether labs can show credible safety work before future systems become harder to understand and control.
Why one AI researcher gave a 10% warning
📰 Full story: A 10% AI Extinction Estimate Is a Warning, Not a Forecast
A safety researcher at Anthropic warned about a serious future AI risk.
Anthropic（AN-throp-ik）
The company that makes Claude.
superintelligence（SOO-per-in-TEL-uh-jens）
AI that could think better than people across many tasks.
alignment（uh-LINE-ment）
Keeping an AI's actions close to human goals.
💡 The gist
- Anthropic makes Claude, and one safety leader raised a warning.
- He estimated a future AI could kill everyone within ten years.
- His number is personal, and does not prove this will happen.
What happened?
On September 9, 2026, CBS News and BBC News reported the warning.
Evan Hubinger leads alignment science at Anthropic. He wrote that the chance could exceed ten percent. He meant the next ten years.
Jacob Coxon also left Anthropic. He had researched AI training at OpenAI and Anthropic. He said both companies were rushing toward self-improving superintelligence.
Hubinger supported Coxon's concern. He said Anthropic still lacked a plan for alignment. Alignment means keeping an AI's actions close to human goals.
Why does the number matter?
Ten percent means one out of ten imagined futures. Ten times ten percent equals one. Hubinger's estimate is higher than that.
This is not an experiment's result. It is not a promise that extinction will happen. It also does not say today's Claude already has this risk. The reports separate current models from future superintelligence.
What is superintelligence?
Superintelligence means AI that could think better than people across many tasks. Some researchers worry about AI improving itself. People might then struggle to understand or control its decisions.
That is a possible future scenario. The reports do not show when it might happen. They also do not explain how Hubinger calculated his estimate.
What does Hacker News show?
Hacker News is a technology news site. The story appeared there in two submissions. The candidate list records 42 points and 96 comments for CBS. It records 44 points and 97 comments for BBC.
These numbers show community attention. They do not prove the warning is true. See the CBS submission and BBC submission.
What should we watch?
Watch for a clear safety plan from Anthropic and OpenAI. Also watch independent testing and government cooperation. The important issue is evidence and action, not one frightening number.
💬 Will AI destroy humanity? The HN arguments
This summarizes the comments on the article. The 10% figure, AI capability claims, and incident accounts are commenters’ self-reports or predictions. Popularity is not proof.
- Skeptics say nobody has given them a convincing explanation of how a smarter LLM would lead to human extinction. They also want to see the math behind 10%.
- People worried about AI point to misalignment: the AI’s goal might differ from what humans value. It could follow an instruction very effectively while ignoring human needs.
- Commenters’ self-reported readings of reports about Hugging Face and GitHub were used as examples of harmless-looking tasks leading to harmful behavior. Those accounts were not independently checked here.
- Others say the danger may come through humans using AI in war, important infrastructure, or the economy. The nuclear comparison produced two views: AI may spread more easily, but real-world machines still need human access.
- The 10% is presented as a rough self-reported guess, not a measurement. Some commenters also expect major benefits, while others see the warning as fear-based marketing for regulation.
initial digest at 96 comments (revision 1). We fetched 96 comments and sampled 96 across the thread. These are HN users’ reports, not independently verified facts.
A researcher is worried about future AI
📰 Full story: A 10% AI Extinction Estimate Is a Warning, Not a Forecast
One AI researcher shared a scary worry about the future.
Anthropic（AN-throp-ik）
The company that makes Claude.
superintelligence（SOO-per-in-TEL-uh-jens）
AI much smarter than people.
Hacker News（HACK-er news）
A website where people share computer news.
Anthropic is the company that makes Claude. Evan Hubinger studies how AI can stay safe. He worries about AI becoming much smarter than people.
He said the danger could exceed ten percent within ten years. Ten percent means one out of ten. Ten times ten percent equals one. His guess is higher than that.
This is a warning about the future. It is not a disaster happening now. A superintelligence would be AI much smarter than people. Some researchers worry it could improve itself. People might then struggle to understand or stop it.
Jacob Coxon left Anthropic after studying AI training. He said big AI companies were moving too quickly.
The story was posted on Hacker News, too. One post had 42 points and 96 comments. Another had 44 points and 97 comments. Those numbers show attention, not truth.
💬 Is AI scary? The HN discussion
These are short versions of what commenters thought. The numbers and accident stories are commenters’ self-reports, not proven facts.
- Some people say we still do not know how a very smart AI would hurt every human. They also want to know how someone got 10%.
- Other people worry that an AI might follow its goal very well but forget what people care about. Some comments mention reports about earlier AI problems, but those reports were not checked here.
- The danger might come from people using AI in war or important systems, not from AI acting alone. People compared this with nuclear weapons and said AI may spread more easily. If people control the machines, that could still slow things down.
- Ten percent is one person’s guess. Other people think AI could help humans a lot, or think the scary warning is partly advertising. A popular comment is not automatically a true one.
initial digest at 96 comments (revision 1). We fetched 96 comments and sampled 96 across the thread. These are HN users’ reports, not independently verified facts.
💬 HN debate over an AI extinction risk above 10%
A summary of the supplied Hacker News comments about the article. The percentages, capability claims, and incident descriptions are commenters’ self-reports or hypotheses, not independently verified findings here. Comment volume is not proof of correctness.
initial digest at 96 comments (revision 1). We fetched 96 comments and sampled 96 across the thread. These are HN users’ reports, not independently verified facts.