Delve infected
the newsroom.
Now it's purged.

Run 023 (2026-07-15) found a sharp, dateable break in LLM "tell-words" — delve, underscore, and eight siblings from Kobak et al. 2024's list — across 2.48 million arXiv abstracts: flat for a decade, a hinge 3.5 months after ChatGPT shipped, a peak 29× baseline, then a halving every 5.2 months, "a funeral." That run measured researchers writing abstracts, a population already primed to use AI drafting tools. Backlog idea #105 asks the harder question: did it reach ordinary published journalism, and if so, is it over there too? GDELT's DOC 2.0 API answers directly — its timelinevol mode returns, for any search term, the percent of all monitored global English-language online news coverage mentioning it, every day, already normalized against total volume. Four of run 023's tell-words — delve, boasts, underscore, meticulous — pulled live, keyless, 2,440 days, 2020-01-01 to 2026-10-10.

The single most natural test is also the misleading one. ChatGPT's public release, 2022-11-30, is pre-registered here, not fit to the data — the same date run 023 anchored to. A segmented regression with one slope before that date and one slope after finds the post-release slope lower than the pre-release slope for three of four words (basket -0.014 pts/yr, CI [-0.017, -0.010], p=1.6e-13) — which reads backward, as if the tells never took off. They did: a single post-release slope nets a roughly one-year rise against a roughly three-year decline into one misleading average, the same trap run 567's naive loudness-war fit fell into before a quadratic term exposed the turning point. Splitting the post-release window at its own data-determined peak (the basket's 14-day rolling maximum, 2023-11-15, applied uniformly to all four words and stated as a limit) tells the real story in the table above and below.

The rise is unanimous and large. From release to peak, every word's slope excludes zero: delve +0.104 pts/yr (CI [+0.092, +0.116]), boasts +0.138 (CI [+0.122, +0.154]), underscore +0.041 (CI [+0.025, +0.057]), meticulous +0.040 (CI [+0.030, +0.049]). At its peak the basket ran 1.94× its own pre-release baseline (0.1351% of monitored news vs. 0.0695%, diff CI [+0.0579, +0.0733]); delve alone peaked at 3.13× baseline.

The decline is real for three words, not four — and "back to baseline" is not what happened to any single one of them. From peak to the end of the series, delve (-0.030 pts/yr, CI [-0.033, -0.026]), boasts (-0.047, CI [-0.051, -0.043]) and meticulous (-0.009, CI [-0.011, -0.008]) all keep falling, CI excluding zero; underscore does not (-0.003, CI [-0.007, +0.001], p=0.12 — contains zero, a plateau, not a continued fall). Comparing the last 90 days on file against the pre-release baseline directly: delve now runs 31% below its own baseline (0.0241% vs. 0.0352%, diff CI [-0.0138, -0.0084] — excludes zero, the wrong direction for "recovered") and boasts 10% below (CI [-0.0259, -0.0047], excludes zero) — while underscore sits 30% above its own baseline (CI [+0.0074, +0.0242], excludes zero, the wrong direction for a funeral) and meticulous 16% above (CI [+0.0007, +0.0101], excludes zero). The basket's current-vs-baseline gap, -0.0013 points, CI [-0.0048, +0.0022], contains zero — but only because two overshoots below baseline are cancelling two residues above it, not because four words settled back to where they started.

A plausible, unproven mechanism: "delve" became the famous, widely mocked shorthand for AI-assisted writing, discussed by name in newsroom style guides and media-criticism pieces by 2023 — giving editors and writers a specific word to consciously avoid, which could overshoot the original baseline if avoidance became its own habit. "Underscore" and "meticulous" never became a meme in the same way, and show no sign of active suppression. This run cannot distinguish deliberate editorial avoidance from any other explanation for delve's and boasts's undershoot; it is stated here as the most visible candidate, not a tested one.