CNN and Fox word gap, 5 to 12 October 2026: counts, intervals and tests

Companion to the editorial. Every figure is the desk’s own count of words on the CNN and Fox News SiriusXM audio feeds, per 10,000 words, with the method that produced it. The tape is a receipt: no transcript line is quoted here or in the editorial.

Design, fixed before the counts

Window: utterances from 2026-10-05T00:00Z to 2026-10-13T00:00Z (eight UTC days). Channels: CNN and Fox News only. Tokens: lowercased runs of letters and digits, apostrophes kept inside a word, a trailing ’s removed; the denominator is every such token. Consecutive identical utterances on one channel were collapsed to one (CNN 444, Fox 419). Primary condition: only utterances inside blocks labelled program, in clock hours where both channels had at least 1,000 such tokens (124 of 192 hours; CNN 792,795 words, Fox 805,879). Eleven words were fixed in advance; the scan list is the 25 most CNN-heavy and 25 most Fox-heavy of the 1,214 alphabetic words of three or more letters, outside a fixed stopword and filler list, that appeared at least 100 times in the paired hours. Tests: label-swap permutation inside each paired hour (10,000 swaps), Holm correction across the eleven, a family-wise p for the scan (share of swaps whose largest gap anywhere in the vocabulary reached the word’s gap), a 5,000-resample bootstrap over paired hours, same-sign counts over the eight days, and a split by Eastern clock (day 06:00 to 17:59, 84 hours; night, 40 hours). Seed 20261012.

The eleven pre-registered words (program only)

Stated sensitivity forms (not part of the eleven):

The same eleven under three conditions

Per 10,000 words; gap is CNN minus Fox. Raw tape: 136 paired hours, CNN 1,381,231 words, Fox 1,384,865. Ad blocks removed: 136 hours, CNN 1,054,503, Fox 1,062,908. The same seven words exceed noise in all three; the same four do not.

Block-swap tests (added after the primary results; not pre-registered)

Neighbouring hours are not independent, so the label-swap test was repeated with whole blocks of consecutive paired hours swapped as units. Program only, eleven words, Holm across the eleven. With 6-hour blocks (25 blocks) all seven words that exceed noise on the hour-level test still do. With 12-hour blocks (16 blocks) six do; socialism does not (Holm p 0.086).

6-hour blocks

12-hour blocks

Daypart halves, 6-hour blocks, uncorrected

Eastern daytime hours (84 hours, 16 blocks):

Eastern evening and overnight hours (40 hours, 9 blocks):

The scan: 25 words each way (program only)

Family-wise p is conservative for rarer words, because the statistic is an absolute rate difference. With labels shuffled inside paired hours, the largest gap on any word had a median of 2.9 per 10,000 words and a 95th percentile of 6.0. Names of the networks and of hosts are left in the lists on purpose.

CNN-heavy

Fox-heavy

What the commercials did

Share of each channel’s words by block label over the whole window: CNN program 57.7%, ad 23.7%, promo 1.8%, unlabelled 16.8%; Fox program 58.4%, ad 23.2%, promo 1.5%, unlabelled 16.9%.

The word trump across the whole window: CNN program 2,876, unlabelled 470, ad 22, promo 12; Fox program 1,319, unlabelled 269, ad 489, promo 26. Of Fox’s 489 uses in ad blocks, 467 sit in 89 airings whose ad record names an advertiser with Trump in its name, and 340 of those are in 55 airings of Trump Watches. On CNN, 7 uses sit in 2 airings of that kind, none of them Trump Watches. Advertiser attribution is the Observatory’s ad ledger, which dedups creatives by transcript similarity; airing counts are blocks.

Transcription spellings of Mamdani

Within two edits of “mamdani” (program tape, exact spelling excluded): CNN 7 uses, Fox 53 uses, the commonest “mondani” (Fox 29, CNN 3). With the variants added, Fox has 199 uses and CNN 27, against 146 and 20 for the exact spelling.

Reproducibility

The pre-registration text and the analysis script were hashed before the primary run: see the sha256 values in the desk’s source receipt for the editorial. The script reads the Observatory database read-only. Capture is one SiriusXM login, a faster-whisper small model for transcripts, and Claude Haiku 4.5 for block labels. No transcript line is reproduced; only counts.