How it works › About
About Hidden AGI watch
A fire alarm for hidden AI: a daily, sourced reading of whether advanced AI could already exist, or already be acting, without the public knowing.
What this is
Each day we answer three questions: is AGI here, could it be hidden, and should you worry today? The answers are sourced probabilities, a tracker of the eight parts of AGI, an Escape watch on the resource chokepoints, and a fire alarm set by published rules. An absence of evidence is not treated as proof of secrecy.
How it's made
A custom AI agent built for this project researches, writes and publishes each daily reading automatically. The research starts each morning and the email goes out at 10am Pacific, or as soon as it is ready if later. No one reviews a daily report before it's published. The Friday wrap-up is made the same way and emailed at 3pm Pacific. The owner sets the definitions and the fire-alarm thresholds.
Before launch, we had 195 AI agents run 3,793 checks on this site's own facts, code and reasoning. It found 135 problems, and we fixed them.
The method is on How it works, the alarm criteria on Fire alarm, and every chart follows the chart style guide.
What gets sent, and who approves it
The daily issue and the Friday wrap-up are emailed automatically, and neither is held back when the alarm level changes: they go out and show the new level. A breaking fire-alarm alert is different: it's prepared as a draft, and a person approves it before it's sent.
Who runs it
Hidden AGI watch is an independent project run by joeldg on GitHub. The code, the data files and the history of every page are public at github.com/joeldg/agi_assessment.
Contact
To report an error, question a number or suggest a source, open an issue on GitHub Issues. That's the only contact channel, so every report and its answer stay public.
Corrections
Past readings are frozen once emailed; corrections are dated, shown on the affected page and carried into the next email.
62 corrections so far. The newest:
- Nov 16 · Authors of the LessWrong post 'Measuring agent capabilities on open-ended AI R&D tasks'. Affected page
- Nov 13 · The start date of the Hugging Face incident (OpenAI agents break into Hugging Face production) and its disclosure lag. Affected page
- Nov 12 · The note on the signal for an external evaluation of OpenAI's unreleased post-Astra model (W2). Affected page
License
- Code: the scripts, the chart kit, the stylesheet and the page templates are under the MIT License.
- The site's own text and data: the reports, readings, rulings, alarm criteria, Escape watch indicators and our own data files are under CC BY 4.0. You can reuse and adapt them, commercially too, with credit and a link to the license.
- Not covered: third-party material we quote, cite or plot (news excerpts, filings, METR measurements, outside forecasts, market prices and the like) stays under its owners' terms, and the brand assets (the name and wordmark, the avatar and logo, the favicon and the default share image) are all rights reserved. The details are in LICENSE-content.md.
How to cite
Hidden AGI watch (2026), <page title>, <url>, accessed <date>. Licensed under CC BY 4.0.
For example: Hidden AGI watch (2026), The fire alarm, https://hiddenagi.com/alarm.html, accessed <date>. Licensed under CC BY 4.0. Daily reports keep their address (reports/YYYY-MM-DD.html), and the data behind every chart is in the repository's data folder, with its full history. A machine-readable citation is in CITATION.cff.