United States federal and state legislation regulating AI companions and conversational systems.
Companion AI regulation went from a handful of introduced bills to a working body of law in about seventeen months: — US states now have an enacted chatbot or companion statute, and roughly half of those do not take effect until 2027, so the statute book today and the compliance picture in eighteen months are different documents.
Other trackers record what stage a bill has reached. This one codes the definitional anatomy of each policy: what it actually treats as a companion, what kind of test its definition asks, what device narrows it, and whether the result reaches the tools young people are already using.
Berkman Klein Center for Internet & Society, Harvard University · last verified 9 September 2026 · source & contributions on GitHub
Hover — or tap, or tab to — any dotted term for its definition.
| Legislation ↕ | Jurisdiction ↕ | Status ↕ | Latest action ↕ | Youth focus ↕ | Test ↕ | Narrowing device ↕ | Reaches assistants ↕ | Mechs ↕ |
|---|
Every dated action in the corpus, newest first: when a bill was first introduced, the last thing that actually happened to it, and when its obligations start to bind. The jurisdiction, status, youth and mechanism filters on the Legislation tab apply here too.
Which regulatory mechanisms each piece of legislation carries. Each bar counts the records coded for that mechanism; the coding records presence, not stringency, so a bar says how many instruments impose an obligation of that kind and nothing about how demanding any of them is. A mechanism can sit at zero: the key is kept in the list so that the absence is visible rather than missing.
The sixteen mechanisms are grouped into five clusters, and every mechanism belongs to exactly one. The cluster figure counts the records carrying at least one mechanism in that family, so it is not the sum of the bars beneath it.
Click a cluster to filter the corpus to it, or click any mechanism for what it means as a legal rule, where this dataset draws the line around it, and how each instrument carrying it is worded — marked verbatim text, close paraphrase or not yet transcribed.
Enacted and in-force law, legislation that has moved out of committee, and the key pending vehicles the analysis relies on. Columns run in cluster order, banded across the top. Hover any cell for the legislation and mechanism, or any column heading for the mechanism's definition. Use the Legislation tab for all .
One row per instrument, holding the three definitional fields side by side: the term the text uses, the test that decides what the term covers, and the narrowing device that pulls things back out of it. They are coded separately because they frequently do not line up — a categorical noun can sit on top of a functional test, and a broad test can be narrowed back to a product category by its carve-out.
The final column, reaches assistants, is this tracker's judgement about the statutory language rather than a measurement, and it depends on all three of the columns before it. Where the definitional clause has not been read against the enrolled text it is coded unclear: a gap in the coding, not a finding about the instrument. Hover any coded value for its definition.