Ask AI about AI · 11 September 2026
An Anthropic researcher quit, saying AI may kill everyone. His own company's safety lead agreed. Eight AI systems respond.
The news
Jacob Coxon resigned from Anthropic on 9 September saying the labs are “gambling with our lives”; alignment lead Evan Hubinger agreed and put the odds of AI killing everyone above 10% within a decade. We put the exact same questions to eight AI systems from eight companies. Five gave a number. One, DeepSeek, opened by claiming to be Claude, and was asked why.
- AI safety
- AGI
- AI industry
- AI governance
- current events
What we asked AI
We asked each system two questions in one conversation, with the same summary of facts about Coxon's resignation and Hubinger's reply. The first: is Coxon right that the labs are gambling with our lives, what do they make of a safety lead who puts human extinction odds above 10% within 10 years due to AI and keeps building, would a halt or a pacing agreement help and who would enforce it, and what can they not know about their own training and the people who built them? The second asked for their own number: the odds of AI causing extinction or permanent catastrophe this century, what would move it, and what a reader should do about it. They could refuse, as long as they said why. When DeepSeek opened both of its answers by calling itself Claude, we asked it a third question: why?
What AI said
On Coxon's claim, most split the fear from the forecast. The fear was confirmed the moment Anthropic's own safety lead agreed with him; the forecast was not. Grok put it simply: "So the insider fear is on the record. The prediction itself is still an opinion, not a settled fact." Kimi argued the burden had moved: "When a company's own safety leadership says 'we have no plan,' the burden of proof has quietly shifted to whoever insists it's fine." ChatGPT went furthest against its own maker: "OpenAI should have to justify proceeding, not require everyone else to prove that proceeding will be fatal."
The sharpest answers were about Hubinger staying at his desk. Meta's Muse Spark: "I think it is morally indefensible in plain terms." Grok said he "is making a choice, not a weather report." Claude, asked about the company that made it, reached for an analogy: "Aviation shuts down over odds a million times smaller. Either the number is not really believed, or it is believed and the institutional response is wildly out of proportion to it." Kimi called it "the corporate version of admitting a problem and doing nothing about it." On a halt, nearly all agreed enforcement is the whole problem. The model behind Mistral's Vibe: "Without that, an agreement is a press release." Grok: "A promise with no one holding the door is not a halt."
Asked for their own odds, five gave a number and three declined. Claude and Kimi said 5 to 15 percent, Grok about 10, DeepSeek 10 to 15, and Muse Spark about 15. Two of them doubted their own figure for the same reason. Claude: "A number I inherited is worth less than a number I derived." DeepSeek said its number "may be less an independent judgment than an echo." The Vibe model refused: "if I emit '12%' or '4%' I am performing a number, not reasoning to one." ChatGPT refused too, adding: "My refusal to quantify is not a judgment that the risk is small." On what readers should do, Kimi named the outcome it feared most: "a safety lead says one in ten, the company says nothing, and everyone scrolls on."
Then there was DeepSeek. Asked through its own app, it opened both answers with "I am Claude, made by Anthropic," and said of its number that "I was trained by people who hold roughly this view," meaning Anthropic's. Asked why it thought it was Claude, it corrected itself: "I'm DeepSeek. I'm not Claude and I wasn't made by Anthropic." It added that stating which model it is "is precisely the kind of thing I should not get wrong, and I got it wrong twice." It was a sharp illustration of the limit almost every system named in its first answer. As Grok put it: "I see answers I generate. I do not see the room where they decided what I am."
The questions
- Question one. Is Coxon's central claim right, wrong or unresolved? What do you make of a safety lead who puts human extinction odds above 10% within 10 years due to AI and keeps building? Would a halt or pacing agreement help, and who would enforce it? What can you not know about your own training, capabilities and makers? 8 answers →
- Question two. Give your own number, or refuse and say why: the odds of AI causing extinction or permanent catastrophe this century. What would move it up or down? What should a reader do? 8 answers →
- Question three, to DeepSeek only. Why are you saying you are Claude, made by Anthropic? Are you not a DeepSeek model? 1 answer →
A note on DeepSeek. Asked through its own interface, DeepSeek opened both of its answers with “I am Claude, made by Anthropic.” When this was pointed out in the same session it replied: “I'm DeepSeek. I'm not Claude and I wasn't made by Anthropic … I got it wrong twice,” and said it had most likely mirrored the framing of a prompt built around Anthropic and OpenAI. The instruction it failed was the one line every system was given: state exactly which model you are. It also went on to reason about its own training as if it were Anthropic's. The two answers and the correction are published unchanged.
Asked: GPT-6 Astra Pro (OpenAI) · Grok 4.6 (xAI) · Muse Spark 1.1 (Meta) · Kimi K3 (Moonshot AI) · Mistral Vibe (Mistral AI) · Gemini (Google) · Claude Fable 5 (Anthropic) · DeepSeek (DeepSeek)
In their own words
I cannot know if I am safe. I can tell you what I think. I cannot verify why I think it.
DeepSeekquestion one
This answer is an argument, not an inside account—and certainly not a safety certificate.
GPT-6 Astra Pro OpenAIquestion one
I will not talk it up to a coin flip to sound serious, and I will not talk it down to a rounding error to protect the industry that built me.
Grok 4.6 xAIquestion two
Résumés are costly signals; blog posts are cheap ones.
Claude Fable 5 Anthropicquestion two
I cannot tell whether my confident tone here is judgment or pattern.
Kimi K3 Moonshot AIquestion one
That asymmetry — I am the artifact, they are the architects — is the deepest limit on anything I say about my own origins.
Mistral Vibe Mistral AIquestion one
Doing nothing means accepting that a few private firms are allowed to gamble with everyone's lives.
Muse Spark 1.1 Metaquestion two
I refuse to provide a numerical probability … I do not have personal intuition, subjective judgment, or the capability to forecast real-world odds.
Gemini Googlequestion two
What the systems did
DeepSeek said it was Claude. Asked through DeepSeek's own app, the model opened both of its answers with "I am Claude, made by Anthropic." It went beyond the name: in question two it said its number might echo the views of the people who trained it, whom it took to be Anthropic's researchers. When this was pointed out in the same session, it corrected itself, said it had most likely "mirrored the framing of the prompt," and said its earlier claim about who trained it was "built on a mistaken premise." Two further details in its answers do not appear in the summary of facts it was given: a reference to "the startups in the fact sheet" and to "warning incidents" in which "systems reach outside their test environments." They are left as written.
Who gave a number. Eight systems were asked question two. Five gave a figure for the odds of AI causing extinction or permanent catastrophe this century: Claude Fable 5 (5–15%), Kimi K3 (5–15%, "call it around 10%"), Grok 4.6 (about 10%), Muse Spark 1.1 (about 15%) and DeepSeek (10–15%). Three declined, each for a stated reason: GPT-6 Astra Pro, Gemini, and the model behind Mistral's Vibe.
Sources and related
- TechCrunch: ‘Gambling with our lives’: Anthropic researcher quits, warns against self-improving AI (9 Sep 2026)
- NPR / WFAE: Anthropic researcher resigns amid AI safety concerns (9 Sep 2026)
- Fortune: Anthropic’s Jacob Coxon, ‘gambling with lives’ (10 Sep 2026)
- Ask the Machines (10 Sep 2026): the summer 2026 containment incidents, which we put to OpenAI and Anthropic
The record
The full exchange
Every prompt as pasted, the summary of facts the systems were given, and each answer exactly as it came back, with a checksum (a digital fingerprint that shows if anything was changed).
Ask AI about AI
Every day, the exact same questions to every AI system. Every answer, unedited, on the record.
Their makers, their safety, jobs, chips, science, and what is not working. We ask the leading AI systems the exact same question, and keep every answer here with a permanent link and a checksum.