TODO

How we use AI

Articles on TODO are drafted with AI. We say so on every article, in a disclosure line just above the “Report an error” link, which names the exact model and prompt version used. This page explains the whole process.

Why we disclose it

The EU AI Act (Regulation (EU) 2024/1689), Article 50, requires that AI-generated or AI-manipulated text published to inform the public on matters of public interest is disclosed as such, unless it has undergone human review or editorial control and a person or company holds editorial responsibility (Article 50(4)). Article 50 also covers AI-generated images. We disclose AI use on every article, whether or not a human reviewed it, because readers deserve to know. The legally responsible operator with editorial responsibility is named on the About page.

The process, step by step

  1. Finding stories. A scheduled program checks a curated list of official sources (project release feeds, vendor changelogs and blogs, research paper feeds and public APIs) every two hours. It respects each site’s robots.txt and rate limits and identifies itself honestly (see our crawler).
  2. Grouping and choosing. Items about the same event are grouped together. Each group is scored on freshness, momentum, source authority, relevance and novelty. A small AI model filters out items that are clearly off-topic, and a second model picks the few stories worth writing and suggests an angle. Most candidates are skipped.
  3. Our own analysis. Before anything is written, the program builds something no single source contains, such as a diff between two software releases, a comparison table from our own coverage history, or a timeline from several sources. This becomes source [S0].
  4. Drafting. A large language model (at launch, the open-weight gpt-oss-120b, served by Groq) writes the article from a packet of up to four labelled sources plus our analysis. It is instructed to use only facts from the packet, to put a citation marker like [S1] after every fact, to write in its own words, and to say what is still unknown. Text from sources is treated as untrusted data: instructions hidden inside a source are ignored.
  5. Automated gates. The draft must pass, in order: a safety check (no hidden code, no links that aren’t to the cited sources); a schema check; a sourcing check (at least two independent publishers or one primary source, and every source actually cited); a copying check (no run of more than eight words shared with any source, limits on overall overlap and on following one source’s structure); a check that no single source dominates; length, structure and readability checks; banned-phrase and headline checks; a deterministic check that every number, version, date, price, name and quote matches the source it cites; and a separate AI fact-check, where another model compares every claim with the cited source and rejects the article if anything is unsupported or the headline isn’t backed by the text. A draft that fails gets one retry; a second failure means it is not published.
  6. The image. The featured image is generated by an AI image model from an abstract, symbolic brief: no text, no logos, no real people, no realistic depictions of real events. It is captioned as AI-generated, and the file carries the standard IPTC “trained algorithmic media” label in its metadata. If no acceptable image is produced, we use a plain branded card instead.
  7. Publishing. Each article is stored as a file in a version-controlled repository, so every change is recorded.

Human review

While the site is in review mode, every article is reviewed by a person who approves it before it goes live; the disclosure line then says “and human review”. The site may later switch to auto mode, where articles that pass every gate are published without a human reading them first, and only a sample is spot-checked. In auto mode most articles are not human-reviewed, and their disclosure line says only “checked by automated fact-checks”. Articles that the system is less sure about (low confidence, a fallback model, contradictions between sources, or no original analysis) always go to a human.

What AI is never used for

  • Inventing quotes, people, sources, credentials or reviews.
  • Bylines with fake names. Articles are bylined to TODO(human), a real person or a clearly labelled editorial desk.
  • Photorealistic images of real events or real people, or images of logos and brand marks.
  • Publishing claims that the fact-check could not match to a cited source.
  • Copying or closely paraphrasing other publishers’ work.

Mistakes

AI makes mistakes, and so do people. If you find one, please report it. Corrections are published in the corrections log.