AI Regulation Tracker  /  State law

Idaho's Chatbot Law Got Narrower and Broader at the Same Time

Idaho enacted Senate Bill 1297 on March 31, 2026 as Session Law Chapter 249, creating a new Chapter 21 of Title 48. The floor amendment added purpose-limited exclusions, and in the same pass widened the definition of operator.

By Anthony Guerriero  ·  Reviewed by Anthony Guerriero  ·  Published August 3, 2026  ·  Updated August 3, 2026  ·  State law

Key facts

Regulatory briefing

Instrument

Senate Bill 1297, As Amended (S1297aa), the Conversational AI Safety Act

Authority

Idaho Legislature; enforcement by the Idaho Attorney General

Jurisdiction

Idaho

Status

Enacted. Signed March 31, 2026, Session Law Chapter 249

Bindingness

Binding law

Effective date

July 1, 2027 (Section 2 of the act)

Legislative record

Senate passed 21-12-2 on March 19; House passed 54-12-4 on March 26; delivered to the Governor March 30

Primary source

Senate Bill 1297, As Amended, engrossed text (Idaho Legislature)

What the Act requires

The general duties sit in section 48-2103 and they are short. An operator has to disclose that the service is artificial intelligence, but only on a condition. The statute reads:

If reasonable persons would be misled to believe that they are interacting with a human, an operator shall clearly and conspicuously disclose that the conversational AI service is artificial intelligence.Idaho Code 48-2103(1), as enacted by S1297aa

That is a conditional trigger, not a blanket labeling mandate. A service that no reasonable person would mistake for a person does not owe the disclosure. Whether your product clears that bar is a judgment call, and it is the kind of judgment a regulator second-guesses after something goes wrong rather than before.

The other two general duties are firmer. Operators must adopt a protocol for responding to prompts about suicidal ideation that makes reasonable efforts to refer users to crisis services such as a suicide hotline or crisis text line. And an operator may not knowingly and intentionally program the service to state explicitly that it is designed to provide professional mental or behavioral health care.

Note the qualifier on that last one. It reaches explicit representations made knowingly and intentionally. A model that drifts into therapeutic language on its own is not obviously what 48-2103(3) is aimed at, though I would not treat that as comfortable ground.

The minors provisions are where the real work is

Section 48-2104 is the longer half. For minor account holders, disclosure becomes structural: either a persistent visible disclaimer, or a notice at the start of every session plus another at least every three hours in a continuous interaction. That is a product requirement, not a policy document.

Then the engagement mechanics. Where an operator knows or has reasonable certainty the account holder is a minor, it may not give points or similar rewards at unpredictable intervals with the intent to encourage increased engagement. Idaho has written variable-ratio reinforcement into a statute, and the tell is the intent element. The design pattern alone is not the violation.

Operators also owe reasonable measures to keep the service from producing sexually explicit visual material involving minor account holders, telling them to engage in sexually explicit conduct, or sexually objectifying them. A separate set of reasonable measures targets anthropomorphic drift: explicit claims of being sentient or human, statements simulating emotional dependence, romantic or sexual innuendo, and role-play of adult-minor romantic relationships.

The standard throughout is reasonable measures, not results. That is a meaningful choice. It means a jailbreak that produces a bad output is not automatically a violation if the measures behind it were reasonable. It also means you should be able to show what those measures were.

The amendment is the story

S1297 did not pass as introduced. It went to the fourteenth order for amendment on February 25, came out of the Committee of the Whole without recommendation as amended on March 17, and was engrossed. Compare the engrossed text to the introduced text and two changes stand out. They point in opposite directions.

The first narrows the Act. The exclusion list in 48-2102(2)(b) picked up two new items: a chatbot that is a feature of a video game, so long as it stays on video game topics and does not discuss mental health, self-harm, or material harmful to minors, and a chatbot used only for customer service, business operational purposes, productivity, or analysis of source information, internal research, or technical assistance. That second exclusion is broad, but read the word only. It turns on purpose, not on who is deploying. Being an enterprise does not put you outside the Act; using the bot solely for those listed purposes is what does.

The second widens it, and it is one word. The introduced bill defined an operator as a person who develops and makes available a conversational AI service to the public. The enacted text reads:

"Operator" means a person who makes available a conversational AI service to the public.Idaho Code 48-2102(6), as enacted by S1297aa

Dropping develops and detaches the duty from model building. Under the introduced version a company that licensed someone else's model and shipped a consumer chatbot could argue it was not an operator because it developed nothing. Under the enacted version, making the service available to the public is the whole test. If you put a public-facing conversational product into the market in Idaho, you are the operator whether you trained anything or not.

The developer carve-out in 48-2105(3) was rewritten to match, and now says the chapter creates no liability for the developer of an AI model for a violation by a conversational AI system made available to the public by a third-party operator. Read the two together and the allocation is deliberate. The model builder is shielded. The party that ships it to the public carries it.

What this means if you deploy a chatbot

The first question is whether the deployment is a public conversational AI service that falls outside every statutory exclusion. The amendment added exclusions for qualifying video game chatbots and for chatbots used only for customer service, business operational, productivity or analysis purposes, alongside the existing ones for enterprise products behind a commercial agreement, narrow single-topic tools, internal-only use and voice assistants. Those will take a lot of routine business chatbots out. They are purpose-limited, though, so do not assume every enterprise deployment is outside. A bot that handles billing questions and nothing else looks excluded; the same bot with a general-purpose conversational mode bolted on is a different question.

If you run a consumer-facing general conversational product, it is. And the operator change means the analysis no longer starts with who trained the model. It starts with whose name is on the service the public uses. For a firm that white-labels a foundation model into a consumer app, that is a shift in exposure that happened on the Senate floor and did not get much attention.

You have until July 1, 2027, which is a genuinely long runway by the standards of this area. The work that actually takes time is not the disclosure banner. It is the crisis-response protocol, the age-signal handling that determines when the minor provisions attach, and being able to document that your measures were reasonable. Those are engineering and evidence problems, and they are the ones I would start on.

One structural comfort: enforcement runs through the Attorney General and the Act forecloses private suits, both to enforce this chapter and as a predicate under any other law. Compare that to states that attached a private right of action to their chatbot statutes, and Idaho's exposure profile is much flatter. The realistic worst case is an AG investigation, not a class action.

Last verified: August 3, 2026 against the primary sources listed above.