The Danger Of Relying On Abstractions-replace verbatim "the internet" with "AI" (medium.com)

A writer discovered that a widely circulated learning graphic—the “Learning Pyramid” that claims exact retention percentages for different study methods—has no traceable scientific source. The numbers are unusually rounded and consistent across many sites, Edgar Dale’s original “Cone of Experience” included no such figures, and investigators (including Will Thalheimer) failed to obtain the underlying evidence. The author used the pyramid after seeing it repeated across sources, then realized this was likely a collective copying error: a telephone-game effect where secondary sources copy each other’s unverified claims until the myth becomes accepted fact. For the AI/ML community this is a cautionary tale about relying on abstractions like AI and aggregated summaries without checking primary research. Unverified numeric claims, metrics, labels or “canonical” charts can propagate through datasets, papers and models, causing dataset contamination, benchmark myths, and model hallucinations that amplify falsehoods. Practitioners should trace provenance, demand reproducible methodology, verify statistical claims against original studies, and treat rounded or consistently repeated figures with skepticism. When designing datasets, benchmarks, or automated ingestion pipelines, build provenance checks and conservative uncertainty estimates to avoid codifying myths into models.