What this does. Working out which AI rules reach one system means holding a great many criteria in mind at once: what the system decides, who runs it, where its output lands, what it is built into, what it produces, who it talks to, how large it is, and which sector it serves. This tool walks those criteria one at a time and returns the instruments that answer to them. It asks only what bears on the jurisdictions you choose, and every finding states the answers it turned on, so the reasoning can be checked rather than taken on trust.
What the answer is worth. The result is indicative. It applies the rules set out in Chapter 8 of the report and nothing beyond them, so it may include an instrument that turns out not to reach your particular facts, and it is written to err that way rather than to leave something out. Every result needs validating case by case, against the instrument itself and with advice qualified in that market.
If a result looks wrong to you, or a regime you expected is missing, write to contentwire@gmail.com. The rule set is corrected and reissued, and contributions are credited in the issue that carries them.
How to answer the size thresholds, and why they change the answer
Two of the eight questions turn on numbers rather than on judgement, and they are the ones readers most often answer wrongly.
Training compute, written as 10²³, 10²⁵ and 10²⁶ floating-point operations. A floating-point operation is a single arithmetic step, and the figure is the total used to train the model, not to run it. Only the organisation that trained the model knows it. If you bought or licensed a model, you are not the one these thresholds reach, and the developer is; ask them for the figure if you need it for a contract. If you trained something yourself on ordinary business data, you are far below all three.
They change the answer because each threshold engages a different regime. Above 10²³ the European Commission presumes a model is a general-purpose AI model at all, which engages the Chapter V duties. At or above 10²⁵ it presumes systemic risk, which adds the Article 55 duties. Above 10²⁶ the Californian, New York and Illinois frontier statutes engage, as does Korea's high-performance category.
Revenue, users and assets. The other thresholds are ordinary business figures: annual revenue above USD 500 million, which decides whether the US frontier duties fall on you as a large developer; registered Indian users above five million; Californian visitors or users above one million; and bank assets above USD 30 billion. Answer them from your own accounts.
Crossing none of them is the common case, and "None of these" is a complete answer.
Answer for one system at a time. Exposure depends on what a system does and where its output is used, not on where the company is incorporated, so a group answer is usually wrong.