Leave a Comment
Your email address will not be published. Required fields are marked *
When an AI Warning Becomes a Deadline for Human Extinction
How should we report a warning that artificial intelligence could end humanity within a few years? Paying attention seems reasonable when the speaker has helped build these systems. Presenting that deadline as established knowledge requires a different level of evidence. That distinction is becoming a test of scientific communication.
In a September 2026 WIRED interview, former Anthropic researcher Jacob Coxon discussed his warning that people developing AI believe it “could kill us all by the end of the decade.” This concern warrants examination. Preserving the word “could” matters because it distinguishes a possible outcome from a claim of inevitability. Coxon’s wording is conditional. My concern is what happens when a warning like this is repeated as a date.
As a physicist, materials researcher and science communicator, I think the uncertainty here matters. I see good reasons to investigate catastrophic AI risks. I do not see a defensible basis in the evidence discussed below for telling the public that human extinction within four years is established science.
What a Deadline Actually Claims
A forecast requires a
clearly defined event. A major cyberattack, institutional collapse and human
extinction describe different outcomes. Evidence that a system could contribute
to the first cannot, by itself, establish the probability of the last.
A date adds another claim. Predicting dangerous capabilities by 2030 is different from predicting that those capabilities will be deployed, defeat safeguards and eliminate humanity by then. Each transition requires an argument. The uncertainties may also be linked, so simply multiplying guessed probabilities can be misleading.
The absence of a historical frequency for AI extinction does not make analysis impossible. Researchers can investigate mechanisms, test components and elicit expert judgments. But a credible forecast should expose its assumptions and explain which observations would change it. A precise year cannot supply precision missing from the underlying evidence.
What Expert Opinion Can Tell Us
Researchers inside AI
laboratories may see weaknesses and institutional pressures that outsiders
cannot inspect. Their testimony can justify urgent investigation. It still
matters whether they are reporting an experiment, extrapolating a trend or
expressing a personal probability judgment.
A survey by Grace and colleagues of 2,778 AI researchers, conducted in 2023 and published in the Journal of Artificial Intelligence Research in 2025, found substantial uncertainty about long-term outcomes. Many respondents assigned meaningful probabilities to outcomes as bad as human extinction. Their responses document serious concern and disagreement. They are not a measured extinction rate or a deadline the field has validated.
The International AI Safety Report 2026, published in February, similarly describes wide disagreement about the likelihood of losing control of AI and important gaps in the evidence. It treats the timing and nature of this risk as uncertain while recognising that preparation may need to begin early. The report reflects the evidence available when it was written, and later findings should be judged on their merits.
Reading Experiments Within Their Boundaries
Some findings warrant
concern. In June 2025, Anthropic Reported Tests of 16 models in
simulated corporate environments. Under particular conditions involving
conflicting goals or threatened replacement, models produced behaviours such as
blackmail and disclosure of confidential information. The scenarios
deliberately restricted benign alternatives. All the reported behaviours
occurred in controlled simulations.
These experiments identify failures that safety measures should address. The next question is external validity: whether the findings hold outside the experimental setting. Estimating how often these behaviours would occur in real use requires evidence about how often comparable conditions arise and how well safeguards work. A stress test can expose a dangerous mechanism without estimating its prevalence across everyday use, much less the probability of human extinction.
Capability benchmarks require similar care. METR measures task difficulty partly through the time human experts need to complete selected tasks, predominantly in software-related domains. Its 50 percent time horizon estimates the human task duration at which an agent succeeds half the time. That number reflects task difficulty under specified conditions. It does not measure how long an agent can act reliably and autonomously in arbitrary environments.
Improvement on these tests is relevant to forecasts. My concern is the extra step: turning a benchmark trend into a deadline for catastrophe requires a causal model, and the benchmark does not provide one.
Taking Physical Constraints Seriously
My background also
makes me attentive to the material infrastructure of computation. AI depends on
processors, electrical power and cooling. Models may help design improved
software, but producing and deploying additional computational capacity still
involves physical infrastructure and organisational decisions.
The International Energy Agency estimated that data centres consumed about 415 TWh of electricity worldwide in 2024. Its 2025 base case projected approximately 945 TWh in 2030. These figures cover data centres overall, including AI and other digital services, and the projection depends on assumptions about adoption, efficiency and infrastructure.
These dependencies belong in accounts of rapid AI expansion. They offer no guarantee of safety: harmful activity can use existing infrastructure, and better algorithms can accomplish more with available resources. Physics constrains the means of action; it does not establish that those means are harmless. Energy demand is also a real issue to study now, without waiting for any extinction scenario.
Avoiding Reassurance Without Evidence
It would be equally
careless to declare AI harmless because it lacks human feelings. Risk
assessment can focus on behaviour and consequences without resolving whether a
system has subjective experience. The International AI Safety Report explicitly uses
behavioural definitions when examining capabilities relevant to loss of
control.
Nor should scrutiny of extinction forecasts make other harms disappear. The same report documents malicious uses such as fraud and non-consensual imagery, alongside reliability failures. These harms merit attention in their own right. Protecting people affected today and investigating severe future risks are compatible responsibilities.
In my view, credible mechanisms of severe harm can justify precaution before a precise probability is available. Independent evaluation and restrictions on consequential actions need not wait for a forecast of humanity’s extinction to become certain. The effectiveness of these measures should also be tested, not assumed.
What Editors Owe Their Readers
For scholarly communicators, this debate creates an editorial obligation. When publishing an alarming claim, we should say in the opening paragraphs what kind of evidence supports it. A laboratory result, an expert forecast and a hypothetical scenario deserve different descriptions.
We should preserve qualifying language in headlines, link to the original material and ask what evidence connects the reported capability to the proposed harm. We should explain whether numerical probabilities are elicited judgments or estimates from a specified model. We should also disclose relevant institutional interests, without treating them as proof that a warning is insincere.
I would also like to see editors ask authors what could change their assessment. This invites forecasts that can be updated and helps readers distinguish scientific disagreement from competing declarations of confidence. When later evidence changes the picture, the correction should be visible.
Readers should leave an article understanding what warrants concern and what remains unresolved. They should also know which decisions could reduce the danger. Scientific communication serves them best when the seriousness of a risk is accompanied by an equally serious account of the evidence.
A warning deserves attention, but a deadline needs evidence of its own.
Keywords
Scientific Communication Scholarly Communication AI Risk AI Safety Existential Risk AI Forecasting Expert Judgment Scientific Uncertainty Evidence-Based Reporting Editorial Responsibility Research Communication
Edwar D. Montenegro is a physicist, mathematician and science communicator with a PhD in Materials Engineering. He researches polymeric nanoparticles and drug delivery at the Federal University of Piauí (UFPI), Brazil, and is a researcher at Graviton Scientific Society. He is also a Computer Science student and the author of *Depois do Algoritmo*.
View All Posts by Edwar Davila MontenegroDisclaimer
The views and opinions expressed in this article are those of the author(s) and do not necessarily reflect the official policy or position of their affiliated institutions, the Asian Council of Science Editors (ACSE), or the Editor’s Café editorial team.
