Every artificial intelligence bill introduced this session leans on a definition broad enough to capture a spreadsheet. This is not sloppiness. It is the load-bearing compromise, and it is why none of the bills will work.
The pattern is consistent. A bill defines its subject as a system that makes predictions, recommendations or decisions influencing real or virtual environments — language borrowed, with minor variation, from an international standards document. Then it carves out. The carve-outs run for pages and they are where the actual policy lives.
Drafting this way has an obvious appeal. A broad definition survives contact with technology that has not been invented yet, and the carve-outs can be negotiated with the industries that show up. The cost is that the resulting statute means whatever the negotiation produced, which is to say it does not mean anything in particular.
Regulated parties respond rationally: they hire counsel to determine which carve-out applies, and if none does, they hire counsel to argue for a new one. Enforcement agencies respond rationally too, by declining to bring cases where the definitional question is the whole case.
There is a better approach and Congress has used it before. Regulate the harm, not the technology. Credit discrimination is illegal whether the decision came from an underwriter, a scorecard or a model. Medical devices are regulated by function and risk class, not by whether software is involved.
Sector-specific regulation is slower and less satisfying. It produces no headline, no framework, no single bill that a member can point to. It also produces rules that mean something on the day they take effect, which the current drafts do not.