Is it February 2020 yet?
Of the top news stories, what fraction are about AI? Measured with the same instrument as Covid in early 2020, so the two curves can be laid on one axis and read against the political milestones of that spring.
AI's trajectory laid over Covid's spring
Full history
Rolling-window log odds (or share) with its own window (28 days by default, so the long view stays smooth), from 2019 by default; edit the dates or click “all data” for the full 2017– record. Ticks along the bottom mark AI milestones (hover to read them). Dashed lines are the doubling-time fits from the section below (computed with the fit windows set there).
What level of coverage went with which action?
| Date | US Covid milestone | Covid share then | AI now ÷ that |
|---|
Coverage does not cause policy, but it is a decent proxy for the salience that makes policy possible. The right-hand column is how many times louder AI coverage would have to get to reach each milestone's level.
How fast is it growing? Doubling time of the odds
A share is capped at 100%, so its growth must saturate; the odds (topic items : all other items) are not. On the log-odds scale exponential growth is a straight line, and the slope of a least-squares fit is a doubling time. Fits below are drawn as dashed lines on the history chart when the log-odds scale is selected.
| Source | Series | Fit range | Window | Days | Odds change | r² |
|---|
Google News “Top stories” snapshots
Recent items the classifier tagged as AI
Data table
Last 26 weeks, weekly (current source)
| Week ending | Items | AI | AI share | Covid share |
|---|
Method and caveats
- Wikipedia Current Events (default source): the leaf bullet items on each day's
Portal:Current eventspage, an editorially curated list of the day's important news (~15–40 items/day). Share = AI items ÷ all items, pooled over the window. Global, event-shaped, volunteer-edited; it under-counts slow-burn debate relative to US front pages, so its Covid curve is steep and its AI level low. Compare shapes and multiples as well as levels. - GDELT DOC: percent of all US English online articles matching the query, a volume measure. Both appear in the source menu only when the daily job has managed to fetch them; GDELT throttles aggressively.
- Classifier: one regular expression per topic, applied to the item text plus its parent bullets (the topic often sits in the parent, e.g. “COVID-19 pandemic › …”). Tuned for precision:
AIcounts only as a capitalized token; named labs, models, people and concepts are listed explicitly. Every tagged item is indata/wiki_items.jsonlfor audit. The LLM option uses Claude labels against a fixed rubric that counts only frontier, general-purpose AI (LLMs, foundation models, frontier labs, AI policy and risk, AI-generated media, and speculative pieces about general AI), not narrow machine-learning applications, self-driving cars, robots, chips or big-tech news in general. Bulk labels come from a small model; every chunk with an AI positive is re-adjudicated by a stronger one. - Matching: the first day in the Covid window whose windowed share is ≥ AI's current share. “Multiple of own baseline” divides AI by its 2017 mean and Covid by its Jan–Nov 2019 mean first, since AI had a nonzero baseline and Covid had none.
- Structural difference: Covid was one event with an exponential ramp over ~8 weeks and direct harm to voters; AI is a slow burn with spikes. Level-matching says how loud AI is relative to a known moment, not that politics will follow the same clock.
Source code and data: github.com/Eccentr1city/ai-news-share.