Why AI extinction probabilities are meaningless numbers
Experts' estimates of AI wiping out humanity range from 0 to over 95 percent, but these figures lack scientific grounding. The probability of an AI apocalypse is not a well-defined scientific question, making such numbers unreliable. Comparisons to everyday risks like lottery odds or traffic fatalities illustrate the difficulty of assigning meaningful probabilities to such scenarios.
The article notes that expert assessments of an AI-induced extinction event, often called p(doom), vary from zero to above 95 percent. This full-spectrum disagreement supports the author's claim that the question is not scientifically answerable, as it lacks a testable definition.
For contrast, the author cites the UK National Lottery, where winning any prize has a 20.4 percent chance, leaving a 79.6 percent loss rate. Road traffic deaths, about 1 million annually, translate to a global average risk near 0.01 percent, though over 90 percent occur in lower-income nations. Climate change, a more complex risk, uses models showing a 100 percent likelihood of exceeding 1.5°C, 92 percent for 2°C, and 20 percent for 3°C under current policies.
The prevalence of ungrounded p(doom) numbers could skew public perception and policy decisions. It may foster either excessive fear or unwarranted indifference toward AI development. Since these figures lack scientific rigor, they could erode confidence in more credible risk assessments. Society might be better served by concentrating on measurable, incremental hazards and establishing robust oversight mechanisms rather than speculative doomsday probabilities.