There is a particular rhythm to modern artificial-intelligence discourse: a company introduces a sharper new model, someone close to its development warns that increasingly capable systems could become catastrophically dangerous, and the internet is asked to hold two thoughts at once. We are meant to use the tool, marvel at the tool and perhaps pay for the tool—while also accepting that the tool may someday outmaneuver humanity.
That rhythm returned in unusually blunt form after Jacob Coxon, who describes himself as an AI researcher, wrote on X that people building AI sincerely believe it could kill everyone before the decade ends. Evan Hubinger, Anthropic’s alignment science lead, amplified the post and said his personal estimate for that outcome was greater than 10% over the next 10 years.
Hubinger also said Anthropic is making a serious effort, but that the field does not yet have a solution for aligning a hypothetical superintelligence with human aims, nor a clear trajectory toward one. It is an alarming assessment from someone working in the very discipline charged with preventing advanced AI systems from behaving in unacceptable or harmful ways.
It is also only that: an assessment. A probability statement about a future technology that does not presently exist in the described form is not a demonstration, a forecast backed by disclosed evidence, or an imminent-warning notice. Treating a frightening percentage as if it were a weather report risks replacing one kind of overconfidence with another.
A dramatic number is not a completed argument
The phrase “greater than 10%” has obvious power. People intuitively understand it as a non-trivial chance, and attaching it to human extinction makes it impossible to ignore. Yet the figure raises basic questions that remain unanswered in the public exchange. What chain of events does it assume? What capabilities must arise? How quickly must they arrive? What access to networks, infrastructure, money, laboratories, robots, weapons or vulnerable computer systems would be required? Which human institutions fail along the way, and why?
Without those details, the number tells us chiefly how worried its speaker is. That matters—especially when the speaker is an alignment lead at a major AI developer—but it is not the same thing as public evidence that an extinction scenario is likely.
The distinction is particularly important because “AI” covers wildly different things. Today’s mainstream generative systems can write plausible text, generate images, summarize documents, produce code, answer questions, and sometimes interact with software interfaces. They can also fabricate citations, misunderstand instructions, output brittle code, and state incorrect information with enormous confidence. They are useful tools in specific contexts, but the gap between a persuasive chat interface and an independent actor able to end civilization is enormous.
That does not mean the gap can never close, and it certainly does not make all AI harms imaginary. It means a serious discussion has to avoid treating a broad product category as though it were already a single coherent, self-directed superintelligence.
Real harms do not require a robot apocalypse
The most immediate problems around generative AI are much less cinematic than grey-goo imagery, but they are still substantial. Bad advice can cause real harm when users treat a chatbot as a doctor, lawyer, therapist, financial adviser, or trusted friend. Systems that flatter a user’s assumptions can reinforce delusional thinking or reckless decisions. AI-generated material can accelerate scams, impersonation, misinformation and low-effort harassment. Training data, copyright, labor displacement, energy use and data privacy remain major live disputes.
For game players and creators, the concerns are similarly practical. Generative tools may be used to imitate performers, replicate artistic styles, flood storefronts with asset-flipped projects, generate deceptive trailers or screenshots, and create support systems that sound helpful while providing false answers. A model that invents a nonexistent quest step is annoying. One that invents a nonexistent account-security procedure can be dangerous.
Those risks deserve more attention than they often receive precisely because they do not need an unproven leap to godlike machine intelligence. They are current governance, consumer-protection, workplace and design questions. The industry does not need to promise a future digital overlord before it explains how it will handle damage occurring in the ordinary present.
AI systems also tend to invite anthropomorphism. A fluent reply can feel like evidence of comprehension, intention or personality. An agent that clicks through menus can look autonomous even where it is following constrained instructions, failing repeatedly, or requiring close human setup and oversight. The presentation can be impressive without proving the deeper claim that a system understands the world, forms goals or can secretly pursue a long-term plan.
Why capability claims need receipts
OpenAI chief scientist Jakob Pachocki recently described AI programs as able to use computers and graphical interfaces, collaborate with people and one another, carry out research projects, and reshape computer security in ways that introduce new dangers. Those are consequential claims. They should be met neither with reflexive dismissal nor reflexive awe.
The useful next step is specificity: documented research tasks, clear success rates, the level of human supervision involved, the conditions where a system fails, the duration of its reliable operation, and independently testable security findings. “Can operate a computer” can mean anything from completing a narrow benchmark task in a controlled environment to robustly handling the unpredictable mess of real-world work. Those are not interchangeable achievements.
Games offer a simple analogy. A non-player character that performs brilliantly in a carefully designed arena has not necessarily demonstrated that it can navigate every emergent situation in a sprawling open world. Benchmarks matter, but benchmark performance is not a blank check. The route from a model doing well on selected tasks to a system reliably executing open-ended plans in the world is full of constraints, failures and human choices.
That is why extraordinary public claims need concrete, reproducible support. It is reasonable for companies to withhold some technical material for security or commercial reasons. It is not reasonable to expect the public to accept the largest imaginable claims about capability and danger with no meaningful way to judge them.
The incentive problem is impossible to ignore
There is a built-in tension whenever AI companies are among the loudest voices describing both their systems’ enormous promise and their potential to become existentially dangerous. Warnings may be heartfelt. Hubinger’s comments should not be casually recast as a marketing campaign simply because they also fit a powerful industry narrative. At the same time, companies promoting AI have clear incentives to portray their models as transformative, historically significant and difficult to replicate.
Fear can enlarge the aura of capability as effectively as optimism can. If a product is framed as potentially world-changing enough to demand special rules, vast computing resources and an elite safety apparatus, that framing may help establish the developer as indispensable. It can also move the discussion away from whether today’s products deliver dependable value for ordinary users.
None of this proves bad faith. Incentives are not secret plots; they are reasons to ask tougher questions. Investors, governments, journalists, game studios, players and customers should separate a company’s stated beliefs from the evidence offered for those beliefs. They should also distinguish a genuine safety request from a rhetorical flourish that makes a release sound more powerful than its demonstrated performance.
That scrutiny matters in an industry already facing trust questions. Anthropic’s Claude service has been in the news for a separate dispute involving Max subscribers and allegations around usage representations and weekly caps. The Claude Max subscriber lawsuit is not evidence for or against extinction-risk arguments, but it underlines why straightforward communication about capabilities, limits and access is essential.
Alignment is a legitimate issue, even without certainty
It would be equally careless to say that because current systems make obvious errors, research into alignment is pointless. If companies genuinely intend to build increasingly autonomous systems and hand them more access to computers, code repositories, financial tools and sensitive data, then safety work should advance alongside capability work. Testing for deception, unsafe instructions, cybersecurity abuse, privacy failures, manipulative behavior and loss of human control is sensible risk management.
The key is to keep the language proportionate to the evidence. Acknowledging uncertainty is not weakness. “We do not know” is often the most honest answer in frontier technology. It is far more useful than declaring either that humanity is assuredly doomed or that every concern is ridiculous because today’s chatbot forgot what was said three messages ago.
For the gaming audience, the practical rule is familiar: do not pre-order the hype. Judge systems by what they demonstrably do, where they fail, what data they use, who is accountable when things go wrong, and whether the people selling them explain those limits clearly. A frightening forecast from an expert can be newsworthy without becoming a prophecy.
Hubinger’s greater-than-10% estimate is a stark reminder that some people at the frontier of AI development are deeply uneasy about the destination. It should prompt questions, not surrender. The public deserves concrete evidence, rigorous independent evaluation and rules aimed at harms that can actually be observed—not just ominous trailers for a hypothetical sequel no one has proved is in development.
