The Mood Of The Coverage
How AI is being written about
One number, moved by every story this desk collects. It rises when our sources reach for hopeful language and falls when they reach for alarmed language.
-12.6
NEGATIVE
230 stories · 7 days to 2026-09-21
01The Line It Came From
Each point is a 7-day rolling window, plotted on the day it ends. Hover any point for its date and how many stories stand behind it. Nothing on this page runs: the switcher and the tooltips are stylesheet rules, because this site loads exactly one script and it is not here.
Data began 2026-09-10 · 12 point(s)
Data began 2026-09-10 · 12 point(s)
Data began 2026-09-10 · 12 point(s)
Data began 2026-09-10 · 12 point(s)
Data began 2026-09-10 · 12 point(s)
02Who Is Saying It
The dial above counts REPORTING and COMMENTARY only. Research papers and company blogs are counted separately, because neither is coverage and both lean in a direction you can guess before looking.
REPORTING & COMMENTARY
-12.6
230 stories · this is the dial
Journalists and commentators writing about the field.
THE INDUSTRY ITSELF
+13.1
107 stories · not in the dial
Research papers and company blogs. A vendor is positive about its own product by construction, and a paper titled "Predictable Failure in Multi-Hop Retrieval" is not bad news, it is a technical term.
A gap of 26 points between how the field describes itself and how it is described.
03How A Story Is Judged
Every story this desk collects is marked positive, neutral or negative. None of them wears that mark in public. The number above is an aggregate over hundreds of stories, which is where this kind of measurement holds up; a badge on one article is a verdict about one piece of writing, and ours is not yet good enough to attach to somebody by name. So the stories stay anonymous and the trend is what gets published.
Judge the EVENT, not the language. The question is what happened, not which adjectives were reached for.
Negative
- a harm landed: a breach, an outage, job losses, an injury, a suit filed
- a capability failed IN USE, where people relied on it (not in a paper)
- someone lost something real: access, money, work, rights, recourse
- a credible actor halted, withdrew or rolled something back
Neutral - the default
- THE HONEST DEFAULT. Most reporting is neutral and forcing it elsewhere is how this label becomes noise
- description, explanation, analysis, a guide
- an announcement with no outcome yet
- research reporting a method or a finding, whatever words its title uses
- worry, debate or disagreement reported without resolution
- a forecast
Positive
- a capability shipped and demonstrably works for someone outside the lab
- a cost fell, access widened, or a standing problem was actually solved
- a benefit reached people who did not previously have it
Four ways this goes wrong, each one measured on this corpus rather than imagined:
- Technical vocabulary is not tone. 'failure', 'instability', 'collapse', 'adversarial', 'abstention' are neutral terms of art in research. arXiv cs.CL returned 2 neutral out of 21 because the word counter read them as feeling.
- A story ABOUT risk is not a negative EVENT. Reporting that people are worried is neutral. Reporting that a harm occurred is negative. 'The AI safety vibe shift' describes a change in attention, not a bad thing happening.
- Reassuring words sit on bad events. 'OpenAI halts rollout after safety report' is negative about the event even though 'safety' sounds comforting.
- A vendor describing its own product is not evidence about the field. Primary sources ran +17.0 while journalism ran -14.5 over the same week. A company is positive about its own release by construction.
What is deliberately not judged:
- whether we agree with the story
- whether the outlet is any good
- the writer's mood, only the event's character
How it gets better. The classifier is scored against this rubric on a BLIND held-out sample: the labeller sees headlines and never the machine's answer, because showing the guess and inviting corrections anchors the labeller and manufactures the agreement being measured. The per-story label appears on cards only if it clears 70%. Each round of labels also shows WHERE it fails, and the rubric and the term lists are revised against those cases rather than against intuition. Measured so far: Cairn reading 20 articles in full agreed with the classifier on 10 of 20, which independently reproduces VADER's published 51% on headlines. Baseline to beat is 60%, the share of stories that are simply neutral.
04What This Number Is, And What It Is Not
It measures our coverage, not the field. This is the tone our 57 sources reached for when they wrote. A month where AI genuinely went badly but was written about warmly would read as positive here. Those are different facts and this page only has the second one.
Why an aggregate is allowed where a per-story label is not. A tone label on one headline is close to a coin flip: VADER, a mature sentiment lexicon, scores 51% on article headlines, which is why every story card on this site shows no tone chip at all. But GDELT compared word-counted tone against neural scoring across 69 million articles and found per-document agreement of only r=0.5 rising to r=0.82 once aggregated daily. Averaging many noisy readings cancels noise that does not all lean the same way. The weak claim and the strong one are different objects, and this page makes only the strong one.
The formula is borrowed on purpose. The index is GDELT's: the share of positive stories minus the share of negative ones, on a scale of -100 to +100, counting neutral stories in the denominator so a flat day of plain reporting reads as flat. Using a published measure means this number can be argued with instead of merely trusted.
Where this line starts, and why it is not longer. Our story store rolls as feeds move on, so the further back you go the fewer of that day’s stories are still reachable. Drawing them anyway would show the few items that linger in feeds rather than what was published, which favours deep archives and evergreen posts. So a point is drawn only where its window carries 40 or more stories. This desk began recording the series properly on the date the line starts, and because the series is append-only it grows forward from there rather than being back-filled from a thin residue.
Thin days are not drawn. A day carrying fewer than 8 stories is recorded but never plotted: with six stories the index swings more than 30 points if two are read differently, and a line that jumps like that is drawing noise as though it were news. Every point tells you its count.
Built from 383 stories across 98 days, 12 of which carry enough to draw. The series is append-only. Our story store rolls as feeds move on, so a chart recomputed from scratch each morning would quietly lose its own past; every story is folded in once, by fingerprint, and the history only grows.