of newly published websites, up from 0% pre-ChatGPT.
Aug 2022 → May 2025, stratified Wayback Machine CDX draw.
Semantic Contraction and Positivity Shift. Four came back null.
AI-generated / AI-assisted share of newly published web content
Measurement pipeline
Six testable claims, pulled out of Dead Internet Theory folklore
Correlation strength (Spearman ρ)
Effect size: AI-flagged vs non-AI content
Caveat flagged in-episode: all six correlations run over just 33 monthly points, and every series drifts steadily across 2022–2025 with no stated detrending, first-differencing, or multiple-comparison correction. At α=0.05 across six tests, ~0.3 false positives are expected by chance — two significant results is close to that null expectation, even though the effect sizes themselves are large.
Detector robustness across five stress dimensions
HTML-embedding accuracy hit
Why Pangram v3 won
Three-way classification — fully AI-generated, AI-assisted, fully human — instead of a binary call. Only detector with no severe failure mode across text length, HTML embedding, model family, model version, or language.
Known gap: Appendix A shows it misses completion-only models (davinci-002, babbage-002), and no detector here was stress-tested against paraphrased/adversarially-evaded text — the RAID benchmark used to shortlist detectors doesn't cover evasion either.
Public belief in AI's negative impact vs. what the data supports
vs. 76.2% for frequent users — a 12.1pt gap.
vs. 71.1% for favorable/neutral — a 20.2pt gap.
Belief runs far ahead of the measured evidence.