Why the AI Apocalypse Keeps Getting Predicted and Never Arrives

Every few months, a prominent researcher or executive puts a number on the probability that AI will kill everyone. The estimates range from near-certainty to low single digits. None of the estimators agree. That disagreement is worth examining on its own terms.

A February 2026 survey of AI experts, published on arXiv, found researchers deeply split on both the likelihood and timelines for transformative AI risk. The disagreement is not minor: there is a 90-plus percentage point gap between the most alarmed and the most sanguine, and the people closest to the technology are not converging.

Three assumptions tend to underpin the worst-case scenarios. First, that machine intelligence can become arbitrarily general, capable of self-improvement without limit. Second, that AI systems develop goals analogous to human self-preservation. Third, that superior calculation automatically translates into physical power. Each of these is contestable. Analysis surveyed by Neuroscience News argues that when AI systems appear to disregard instructions, the cause is system inconsistency or ambiguity in the reward structure, not autonomous decision-making. The Association for the Advancement of Artificial Intelligence surveyed 475 AI researchers and found 76% considered scaling current approaches "unlikely" or "very unlikely" to produce general intelligence.

Today's AI systems cannot conduct AI research at the level of a junior human technician. They do not have goals or self-preservation drives in any meaningful sense. This is not a dismissal of risk. It's a description of where the technology actually is.

What the immediate risks look like

None of this means AI poses no serious risks. A study published in the Proceedings of the National Academy of Sciences found that existential risk narratives and immediate AI harms are not actually in competition: you can worry about both. The real near-term harms are less cinematic than extinction but easier to measure: job displacement concentrated among entry-level workers, biased decisions in hiring and healthcare, misinformation at scale, and economic gains captured by a narrow slice of companies.

The AI Safety Clock stood at 18 minutes to midnight as of March 2026, a metric maintained by researchers tracking capability development against safety progress. Whether you find that alarming or theatrical probably says more about your priors than about any specific evidence. What's harder to dispute is that the most dramatic scenarios still rely heavily on capabilities that do not yet exist.

The extinction scenarios get the coverage. The governance failures that need addressing in the next few years, regardless of what AGI does or doesn't become, get rather less.