AI DOOM: Insider Drops 10% Extinction Odds

AI technology concept with various industry icons.
AI DOOM ALERT

An Anthropic safety researcher just put a double-digit number on human extinction from AI within ten years—and admitted his lab lacks a clear fix.

Story Snapshot

  • Evan Hubinger says AI has over a 10% chance to kill all humans within a decade.
  • He adds Anthropic lacks a plan to align superintelligence and is not on track.
  • Major institutions and forecasters sharply disagree on the odds and scenarios.
  • The policy question now shifts from “Is doom possible?” to “What guardrails start today?”

A top safety voice names a stark number

Evan Hubinger, a leading safety researcher at Anthropic, wrote that “AI could kill all humans,” putting the risk at greater than ten percent within the next decade.

He said Anthropic is trying, but does not have a plan to align a future superintelligence and is not clearly on track to one. That is not a random internet post. It is a senior insider stating his personal odds and his view of his lab’s current road map gaps.

News outlets amplified the post because it breaks a silence. Many leaders dance around numbers. Hubinger did not. The clarity forces a basic question: if the chance is even close to that, what changes tomorrow morning?

If a bridge engineer said a one-in-ten collapse risk, drivers would not shrug and keep going. They would close the bridge and add supports. The same logic should guide how we scale frontier models now.

What skeptics and forecasters argue back

Analysts at the RAND Corporation argue human extinction by AI is possible but “highly unlikely.” They say no clear scenario shows AI as a conclusive extinction threat. That is a strong counter.

It rests on human resilience, our spread across the planet, and many choke points for failure before extinction-level harm. It suggests caution without panic and favors measured, testable controls over dramatic shutdowns.

The Forecasting Research Institute measured the split. Concerned experts put existential risk from AI by 2100 near twenty to twenty-five percent.

Skeptical forecasters put it near a tenth of a percent. The gap stayed even after structured debate. That tells us the divide is not just facts; it is values, models of agency, and priors about technology curves. Odds this far apart call for humility and hard evidence, not tribal cheerleading.

How big numbers shape policy and markets

Big odds grab headlines, but rules move when they meet operations. If you think risk is ten percent in ten years, you push for strict caps on training runs, audits with veto power, mandatory kill switches, and liability that bites.

If you think risk is near zero, you resist burdens that slow growth, jobs, and national security. The right answer blends both: lock down the tail risks while keeping American leadership in innovation and defense strong.

Common sense says do first things first. Freeze system access to dangerous tools without human approval. Require red-team testing before release, not after. Log training data and model weights with secure oversight.

Make sure the Department of Defense, the Department of Homeland Security, and the Federal Bureau of Investigation have rapid response playbooks for model misuse. Build fail-safes like we did for nuclear plants and air travel, then test them until they break and fix what breaks.

Sorting signal from noise without denial

Hubinger’s claim is a forecast, not a proof. Forecasts can be wrong in either direction. But his second point is the tell: he says his lab lacks a plan to align a superintelligence.

That lines up with what many engineers admit in private. We know how to tune chatbots. We do not know how to guarantee that a far smarter system will keep human goals. When a builder says the seatbelts for the next car are not ready, you do not floor the gas.

Americans should demand three things now. First, transparency from labs on model capabilities and safety gaps, with penalties for hiding material risks.

Second, clear national standards that stop reckless scaling races while backing secure, American-led research.

Third, accountability that puts costs on those who ship unsafe systems, not on families and small businesses who will bear the fallout if things go wrong. Responsibility must beat hype—on both the doom and the boom.

The takeaway that matters this week

One researcher’s odds do not settle the science. They do set the agenda. A double-digit extinction risk claim from inside a leading lab moves the Overton window. Policymakers must not wait for a perfect model of the future.

They need guardrails that scale with capability, red lines on autonomy in critical systems, and a bias for safety over speed when stakes reach civilizational scale. If we get the early moves right, we will not need to bet on ten percent.

Sources:

cbsnews.com, x.com, seekingalpha.com, en.wikipedia.org, forecastingresearch.org

RELATED ARTICLESMORE FROM AUTHOR

Featured on Right is Right

© 2026 Right is Right. All Rights Reserved.