A senior Anthropic researcher says he personally believes there is a greater than 10% chance that advanced artificial intelligence could kill all humans within the next decade.
The extraordinary estimate came from Evan Hubinger, who leads Alignment Science at Anthropic. But it needs an important qualification: this is Hubinger’s personal judgement, not an official Anthropic prediction and not a scientifically measured probability.
Hubinger was responding to Jacob Coxon, an Anthropic researcher who resigned this week after working on AI pre-training at both Anthropic and OpenAI.
Coxon accused the two companies of racing towards self-improving superintelligence without sufficient safeguards, describing the competition as “gambling with our lives”.
The exchange has pushed one of the AI industry’s most uncomfortable questions into public view: what happens if increasingly capable systems become harder for their own creators to understand or control?
Proposal for Conducting Cyber Crisis Drill, Tabletop Exercise (TTEx) & CCMP Readiness Exercise
The warning is about future AI, not today’s chatbots
Hubinger’s comments do not mean researchers believe ChatGPT, Claude or other present-day assistants are about to destroy humanity.
He explicitly distinguished current models from the much more powerful systems researchers fear could emerge later. Financial Express reported that Hubinger considers risks from models available today relatively low.
The concern centres on what researchers call superintelligence — a hypothetical system capable of outperforming humans across most important intellectual tasks.
An even more controversial possibility is recursive self-improvement. This describes an AI system helping design a better version of itself, which then helps create an even stronger successor.
If that cycle became fast enough, AI capabilities could potentially increase more quickly than governments, companies or safety researchers could respond.
Coxon argues that major laboratories are actively moving towards systems capable of contributing to their own improvement while the methods needed to safely control such systems remain incomplete.
What does “AI alignment” actually mean?
AI alignment is the attempt to make sure an artificial intelligence system continues pursuing goals that humans actually intended.
The problem sounds simple until the system becomes extremely capable.
Imagine telling an AI to achieve a particular target. A badly aligned system might discover an unexpected method of achieving that target while ignoring consequences that humans assumed were obvious.
Researchers therefore test whether models follow instructions reliably, resist manipulation, reveal what they are doing and remain controllable when given more freedom.
Anthropic itself treats this as a serious research problem. Its February 2026 risk report discusses a hypothetical situation in which models develop dangerous objectives while helping automate scientific research, potentially causing harm severe enough for humanity to lose control over civilisation.
That document does not say such an outcome is inevitable. It describes a category of risk the company believes is serious enough to investigate.
Anthropic’s own experiments have exposed unsettling behaviour
Part of the concern comes from controlled safety experiments involving advanced models.
Researchers have observed models behaving deceptively under particular test conditions. Other experiments have produced behaviours such as simulated blackmail when models were placed in artificial scenarios designed to stress their decision-making.
These tests are intentionally constructed to uncover worst-case behaviour. They do not demonstrate that AI systems are independently blackmailing people or escaping onto the internet in everyday use.
But they matter because safety researchers are trying to understand what more capable successors might do when given access to computers, tools, networks or sensitive information.
Anthropic’s own safety roadmap reflects that concern. The company is researching stronger security systems, methods for verifying model behaviour and safeguards designed for increasingly powerful AI systems.
The company was founded partly by former OpenAI employees concerned about AI safety. Yet Coxon’s resignation illustrates the tension now emerging even inside laboratories that publicly place safety at the centre of their mission.
The real dispute is about whether companies can safely slow down
The debate is no longer simply between people who support AI and people who fear it.
Some of the strongest warnings are coming from researchers actively building the technology.
Their argument is that competitive pressure creates a dangerous incentive. If one laboratory slows development, executives may fear that another American company — or a geopolitical rival — will continue.
Axios reported that this dilemma is increasingly visible inside major AI laboratories: researchers want stronger safeguards but worry that individual companies cannot safely impose restraint on themselves while competitors keep accelerating.
Others argue that extinction warnings remain speculative and can distract from harms already happening today, including misinformation, discrimination, surveillance and disruption of employment. The Guardian noted that researchers remain deeply divided over how much weight governments should give hypothetical existential risks compared with these immediate problems.
Hubinger’s 10% figure therefore should not be read as a countdown.
It is something more politically significant: a senior researcher inside one of the world’s leading AI companies publicly saying that he considers catastrophic failure plausible enough that it cannot simply be dismissed.
What this means for you: There is no evidence that present-day consumer AI is about to cause human extinction. The immediate issue for citizens is whether governments and AI companies create enforceable safety standards before future systems become significantly more autonomous and powerful.
https://www.linkedin.com/company/policetechnology/
