Experts question scientific basis of AI extinction predictions
Heidy Khlaaf of the AI Now Institute argues that existential risk percentages are unfalsifiable and lack concrete evidence. She criticizes the absence of specific examples of how superintelligent AI could cause catastrophe. The article notes that even Anthropic acknowledges the unknowability of such outcomes.
Khlaaf's critique centers on the methodological weakness of assigning probability figures to speculative AI outcomes, arguing such numbers lack the evidentiary grounding required of legitimate scientific claims. The AI Now Institute chief scientist points to a fundamental absence: no one has articulated a concrete, plausible mechanism by which an advanced AI system would actually drive humanity to extinction. This absence of specific failure scenarios, she contends, undermines the authority of any quantitative risk assessment.
The uncertainty extends even to prominent AI developers themselves. Anthropic, a leading safety-focused lab, reportedly concedes that precise outcome probabilities remain fundamentally unknowable. This internal acknowledgment highlights a tension within the industry: companies warn of existential dangers while simultaneously admitting they cannot substantiate the severity of those warnings with verifiable data. Critics suggest this dynamic borrows scientific credibility without meeting scientific standards.
This debate could shape how the public and policymakers weigh competing narratives about AI risk. If existential claims lack falsifiable evidence, regulators may deprioritize extreme worst-case scenarios in favor of addressing more immediate, demonstrable harms like algorithmic bias or job displacement. Conversely, public anxiety could persist regardless of scientific rigor, potentially driving demand for precautionary measures. The credibility of both AI researchers and industry leaders may hinge on whether they can ground their warnings in testable, concrete predictions rather than abstract probabilities.