Findings

Analysis of 2026-06-11, on weekly data 2025-06-02 → 2026-05-25 (52 weeks). Methods and caveats at the bottom; the Explore charts show the live data behind each claim.

A year of watching people argue about whether images and video are AI-generated, reduced to what actually changed. The short version: the classic tells are not fading evenly — some are collapsing, some are holding, and the discussion itself is moving from images to video.

  1. Hands are still the #1 tell — and they aren't fading. The most-cited indicator in the corpus (2,637 comments), and its share of weekly citations is statistically flat across the year. share trend p = .30 (no trend); first → last week: 5% → 23% of citations Stable salience doesn't mean hands still betray AI — it may mean "check the hands" survives as a cultural script people recite about hard cases.
  2. The "yellow filter" tell collapsed. The strongest trend in the data: the piss-yellow colour cast — a signature of 2024–25 image generators — fell sharply as a cited tell. −0.27 percentage points/week, p < .001, robust to the media-mix control Its shape is a spike then a fall, so a straight trend line understates the peak. Consistent with generators fixing the artifact — a tell with a half-life.
  3. Voice became a tell. "AI voice" rose from essentially zero to ~4% of weekly citations — and the rise survives controlling for the growing share of video posts, so it is not just "more video now". +0.09 pts/week, p = .005 raw and controlled
  4. Lighting citations are rising. +0.04 pts/week, p ≤ .006 raw and controlled
  5. The vibe tell is fading. "AI style" — it just looks AI — declined through the year. As specific artifacts get fixed you might expect more vibe-reasoning, not less; the data shows less. −0.06 pts/week, p ≈ .01–.02 raw and controlled
  6. "Too old to be AI" is (probably) rising. Citing a post's age as proof it predates capable generators grew through the year. +0.11 pts/week, p = .015–.018 Moderate confidence: this verdict is sensitive to exactly how the video share is measured, unlike the claims above.
  7. The argument moved to video. Video posts' share of tell-citing comments grew from 2.5% to 16% in a year. Every per-tell trend has to be read against this shift — motion and audio tells ride it. see the submission-type chart for the live view
  8. Most comments never cite a tell at all. Of 777,779 substantive comments, 16,749 (2.1%) cite a genuine, curated visual indicator. Most are verdicts — "obviously AI" — without a stated reason. The atlas catalogues the minority who say why. 29,479 flagged by the pipeline; 16,749 survive curation

Method, briefly

Each indicator's weekly share of indicator-citing comments (shares, so corpus growth cancels out), fitted with weighted least squares — weeks weighted by sample size — over 2025-06-02 → 2026-05-25 (52 weeks). A trend is called robust only if it is significant both raw and with the week's video share as a covariate (video = Reddit-hosted or external video link), so a tell can't look like it's trending just because the media mix shifted underneath it.

Caveats that apply to everything above: comments cluster in threads (one viral post can carry many comments), so significance is approximate; the corpus records what people cite, not whether they were right — there is no ground truth on what actually was AI; people post precisely when they cannot tell, so these are the tells that survive into hard cases; and coverage depends on the curated seed list — counts have shifted under curation before and can again. The full caveat list is on the About page.