Zombie stats: Where the numbers AI quotes come from

Discussion overview

The thread largely agrees that AI-generated statistics become dangerous when fluent repetition substitutes for a traceable chain of evidence. Readers focused less on the article’s own presentation than on the feedback loop it describes: an unsupported number appears in AI-written material, gets copied by publishers, then returns through search and training data looking like consensus. A disputed claim about PHP adoption illustrated how a neat, widely repeated percentage can conceal radically different measurement populations and become impossible to audit after enough repetition.

Several commenters want AI aimed at the harder, more useful job: recording source documents, extracted claims, calculations, and final conclusions so each step can be inspected. They see retrieval-backed systems and sentence-level citations as promising in principle, and argue that critical reading of model output belongs in education. But experience has made the room wary. Citation interfaces can lend unjustified authority when linked pages do not support the claim, vanish entirely, or lead to low-quality AI spam. Google’s confident explanations of invented compound words were offered as a vivid example of why polished delivery is not evidence. The core fear is not merely ordinary human error; it is AI scaling that error into a self-reinforcing, increasingly synthetic information environment.

Points of disagreement

The sharpest dispute is whether citations and AI-assisted research meaningfully solve the trust problem. Optimists argue that provenance-aware tools can make verification practical and teach users the critical habits needed to use them. Skeptics counter that current citation systems routinely misrepresent or fabricate support, so visible references often deepen false confidence rather than establish truth.