Is AI ready for your science yet?

The bar for each state

• Routine. Independently validated, in production at several organizations, with a tool you can use today.
• Works with care. Validated and usable, but it needs expert checking or only works for part of the field.
• Research only. Shown in papers or by the developer, not yet replicated at scale or not available to use.
• Not yet. No credible demonstration, or claims that failed when others checked.

The four states follow NASA's Technology Readiness Levels, the scale engineers use to judge whether a technology is ready to fly.

How a cell gets its rating

• Each cell names one concrete task, so the rating is about a real job, not a whole field. Inside it, three or four smaller tasks get their own ratings, because one job can be routine in one part and still research in another.
• Each cell cites 7 to 13 sources. Every source is tagged by kind, by whether it backs or cuts against the rating, and by whether it comes from the tool's maker.
• Every source was opened and read before it went in. Where only a summary could be read, the cell notes it. For Routine and Works with care, at least one comes from someone other than the tool's maker.
• Claude, Anthropic's AI model, helped me find and compile the sources. Where a source comes from Anthropic itself, the cell says so.
• Independent tests, trials, regulator decisions and failure reports count for more than launch posts.
• When the evidence is mixed, the cell takes the lower state and says why. Ratings can move down as well as up.