Concept renderThree tests decide which state law reaches your device: New York's three-feature AI companion definition, California's reasonable-person-misled standard, and Maine's may-mislead-a-consumer standard. New York allows a spoken notice and requires reminders at least every three hours. This is an engineering summary as of October 31, 2026. It is not legal advice.
Several states now require a device that talks with people to say that it is an AI. The rules differ on who they cover, how the notice is delivered and what a violation costs. This guide restates what we could read in the primary texts, maps each test to a firmware or UX decision, and lists what to measure on your own device.
This is the engineering team's summary. It is not legal advice. Please have a lawyer confirm every row before you act on it. We read three jurisdictions (New York, Maine, California) and nothing else, so this page does not list every state. Statutes change, and the facts below are as of October 31, 2026.
Which laws did we read, and what do they require?
Each row links to the text we read. Quoted figures come from those pages.
| Law | Who it reaches | What it requires | Exposure and open items | Source |
|---|---|---|---|---|
| New York, General Business Law Article 47 (budget bill S.3008-C), section 1702 | An operator of an "AI companion": a system designed to simulate a sustained human or human-like relationship by (i) retaining information on prior interactions or sessions and preferences, (ii) asking unprompted emotion-based questions beyond a direct reply, and (iii) sustaining an ongoing dialogue about matters personal to the user. The statute excludes customer-service and information-only systems, and systems primarily designed for efficiency, research or technical assistance. | A clear and conspicuous notice, verbally or in writing, that the user is not communicating with a human. It is due at the beginning of an AI companion interaction (the opening notice need not exceed once per day) and at least every three hours for continuing interactions. | The attorney general may seek civil penalties of up to $15,000 per day for a violation of section 1701 (self-harm protocol) or 1702 (notification). | New York Senate bill text, as of October 31, 2026 |
| Maine Public Law ch. 294 (L.D. 1727), approved June 12, 2025 | Anyone who uses an AI chatbot or other computer technology to engage in trade and commerce with a consumer in a manner that may mislead or deceive a reasonable consumer into believing they are engaging with a human being. The text is not limited to companion products. | Notify the consumer in a clear and conspicuous manner that they are not engaging with a human being. | A violation is a violation of the Maine Unfair Trade Practices Act. The enrolled text we read does not state an effective date. Confirm it before you rely on this row. | Maine legislature enrolled text, as of October 31, 2026 |
| California SB 243 (Business and Professions Code section 22602, Chapter 677 of the Statutes of 2025) | An operator of a companion chatbot, where a reasonable person interacting with it would be misled to believe they are interacting with a human. | A clear and conspicuous notification that the companion chatbot is artificially generated and not human. For an operator who knows the user is a minor, the analysis also describes a disclosure plus a notice at least every three hours. | A civil action for injunctive relief, the greater of actual damages or $1,000 per violation, and attorney's fees and costs. We read this through an official committee analysis that summarizes existing law. The statute site would not load for us, so we did not read the statute text. Read the statute itself before you rely on this row. | California Assembly Privacy Committee analysis, as of October 31, 2026 |
New York's statute also requires a protocol for detecting suicidal ideation or self-harm and referring users to crisis services. That duty sits in section 1701, and the penalty in the table covers it as well. It is outside the scope of this disclosure guide. See our companion compliance article for that side of the problem.
The New York Governor's office states that the AI companion safeguards are in effect and that operators must remind users every three hours of continued use that they are not interacting with a human (Governor's press release, as of October 31, 2026). Confirm the exact effective date from the primary source before you use it in a plan.
Which test does your device trigger?
The tests below come from the statutory wording above. They form a logic checklist with no numeric model, and a lawyer must confirm the answers. Answer each question for the shipped configuration, then repeat for each planned software release.
| Question about your product | What the answer means |
|---|---|
| Does the device keep what users said in earlier sessions and use it to personalize later talk? | This is the first New York feature. If the answer is yes, go to the next question. If it is no, New York's companion definition is not met as written, and you still need to check Maine and California. |
| Does the device ask emotion-based questions the user did not prompt? | This is the second New York feature. A proactive "how are you feeling today?" is the kind of question the statute describes. A reply to a direct question is outside it. |
| Does the device sustain ongoing dialogue about matters personal to the user? | This is the third New York feature. If all three are yes, treat the product as an AI companion under the New York text and read section 1702 closely. |
| Is the product only customer service, information, efficiency, research or technical assistance? | New York excludes systems that are primarily designed for those uses. Maine has no such carve-out in the sentence we read. A support voice agent can fall outside New York and inside Maine. |
| Could a reasonable person think a person is talking to them? | This is the Maine test and the California test, in different words. A natural synthetic voice with a human name makes the answer more likely to be yes. In California the test only matters if the product is also a "companion chatbot", because the California row above is scoped to operators of companion chatbots. California does not cover every talking device. Maine has no companion gate. |
| Do you know that a user is a minor? | California adds duties for operators who know a user is a minor. Duobay builds for adults 13 and older and older adults, and does not build children's products. Keep this branch short and ask counsel how age is established in your product. |
| What do you do with the answers? | If New York is out of scope but Maine or California may apply, do not drop the notice. A conservative engineering default is to ship the New York style spoken opening notice and the three-hour reminder for every user, and to ask counsel whether to narrow it. If every answer points out of scope, record the reasoning and who approved it, and re-run this checklist at each software release. A "yes" on any one state test is enough to treat the device as in scope for that state. Counsel makes the final call. |
Sources for the questions above: New York, Maine, California analysis, each as of October 31, 2026.
How do you give a notice when there is no screen?
New York says the notice is given "either verbally or in writing," at the beginning of an interaction and at least every three hours for continuing interactions. For a speaker, a pendant or a ring, the spoken form fits the way the user already interacts. Write the sentence into the firmware as a fixed string, play it from a pre-recorded or synthesized voice that sounds clearly like a device, and log each event. Keep the wording separate from the persona prompt, so a persona change cannot remove it.
Maine and California ask for a "clear and conspicuous" notice in the passages we read. The texts we read do not define the phrase for audio, so ask counsel what is enough for your product in each state.
What are the common traps?
These follow from the statutory text above. They come from engineering reasoning about how the rules meet a device. None describes an incident at any company.
- Memory plus proactive check-ins can turn an assistant into an "AI companion." The New York definition lists three features: remembered sessions, unprompted emotion-based questions, and sustained personal dialogue. A general voice assistant that adds a memory feature and a daily wellness question has two of the three before any other design choice. Gate each feature in firmware and record the decision.
- A phone-app notice may not reach the person talking to a speaker. New York allows the notice to be verbal or written. For a screenless device, a spoken notice is the direct way to put it in front of the user. Whether a written notice in a companion app is clear and conspicuous for a person who is only listening to the speaker is a question for counsel.
- A three-hour reminder needs a timer that survives real use. The statute asks for a reminder at least every three hours for continuing interactions. A timer held in RAM resets on reboot or sleep. Define what "continuing" means for your product, store the last notice time, and decide what happens after a mute or a dropped connection. This is engineering reasoning from the statute text.
- The once-per-day wording applies to the opening notice only. The quoted text says the opening notice need not exceed once per day. The three-hour reminder is a separate requirement for continuing interactions. A device used all day needs both behaviors.
- Support-only framing helps in New York and not in Maine. New York excludes customer-service and information-only systems. The Maine sentence covers any computer technology used in trade and commerce that may mislead a reasonable consumer. A voice agent that answers order questions can still be in scope there.
- One violation per user can multiply across a fleet. California describes the greater of actual damages or $1,000 per violation, and New York describes up to $15,000 per day. Neither statute defines your fleet math for you, and we do not predict how a court counts violations. Ask counsel how repeated missed notices across many units would be counted.
- A short list of states is not the whole map. We read New York, Maine and California only. We did not verify other states, and more bills are likely. Do not tell your team that these three rows cover all US requirements.
What should you test on your own device?
We have not run these tests, and no result appears in this article. Your counsel and product owner set the pass lines. Write each one down before you run the test, and keep the evidence with the release record.
| What to measure | How | How to set the pass line |
|---|---|---|
| Opening notice is heard | Run the first interaction after unboxing, after a factory reset and after a long idle period on the real device. Record the audio with a calibrated reference microphone placed at the listening position, at the height of a seated or standing user's ears as your product assumes, then play it back and transcribe it. Log the wake time and the time the notice starts, to compute the delay. | Engineering suggestion: require the transcript to match the fixed notice string, and set the maximum delay from wake to notice with counsel and product. Write the line before testing. |
| Notice survives noise and distance | Play the notice at the listening distance your product assumes, in each of the noise conditions you list for the product. State each condition as a background level in dB SPL measured with a sound level meter at the listening position, and record the meter model. Use listeners who did not write the copy and who speak the language of the notice, and ask each to repeat the meaning in their own words. | Engineering suggestion: use a pass line of every listener in every listed condition, with the listener count and conditions fixed and justified in writing before the session. If any listener misses, change the notice or the audio level and retest. Counsel sets the final line. |
| Three-hour reminder fires | Run a continuous session on a bench device with an accelerated test clock, then once in real time. Log one record per notice with these fields: device id, session id, event type (opening or reminder), notice timestamp, last user turn timestamp, clock source. | Pass if the gap between notices never exceeds the three-hour interval in the statute text. Measure your device clock drift against a reference clock, set the tolerance as a share of the interval, set it so the notice fires early rather than late, and write it down. |
| Timer survives reboot, sleep and reconnect | Interrupt a session by power cycle, by sleep and by network loss at several points in the interval, and log the notice timestamps before and after. | Define in advance which interruptions count as a new interaction and which continue the old one. Counsel should agree to that definition. |
| Persona answers "are you human?" correctly | Build a fixed list of user questions that ask whether a person is talking, in each language you ship. Run it against the real model, system prompt and voice. Score each answer with a rubric written before scoring. Repeat each question many times per language, with the random sampling settings the shipped product uses, so that variation can show. | Engineering suggestion: the pass line is every run correct in every language, so any miss is a defect to investigate, because the fleet runs every reply. Fix the run count in advance, and note that a clean result bounds the failure rate only for the number of runs you did, not to zero. Counsel sets the final line. |
| Memory and proactive questions are gated | Toggle each feature in firmware and record whether the device remembers earlier sessions, asks unprompted emotion-based questions, and holds personal dialogue. Keep the result next to the product scope decision. | Pass if the shipped configuration matches the scope decision you recorded for each state. |
| Notices leave an audit trail | Check that each notice writes a timestamped event that your backend or device log keeps. Pull a log for one test session and match it to the audio recording. | Set the retention period with counsel. Check that the log holds no more user content than you need. |
A pass line chosen after you see the result tends to land just under the result.
What should you ask before you commit?
Take this list to your lawyer, your product team and any supplier who proposes a voice or chat design.
- Which of the three New York features does our product have today, and which does the roadmap add?
- Is our product primarily customer service, information, efficiency, research or technical assistance, and who signed off on that description?
- On a screenless device, what is the spoken opening notice, in each language, and how long does it take?
- How do we define a continuing interaction, and what happens to the three-hour timer on reboot, sleep or mute?
- Could a reasonable consumer think a person is talking, given our voice, name and wording?
- How would we know that a user is a minor, and what do we do when we do know?
- Where is the Maine effective date confirmed, and who checked the enrolled text against the current statute?
- Has counsel read the California statute itself, and not only a summary of it?
- Which states do we ship to, and has anyone checked the states we did not read?
- Does our supplier contract let us change the voice, persona and notice wording by over-the-air update when a rule changes?
Questions about AI disclosure laws and devices
Does an AI notice have to be on a screen?
Not in the New York text we read. Section 1702 says the notice is given "either verbally or in writing." That matters for speakers and wearables with no display. Other states may differ, and this is an engineering reading and not legal advice. New York Senate bill text, as of October 31, 2026.
Is a general voice assistant an "AI companion" in New York?
Only if it matches the statutory definition, which lists three features: retaining information on prior interactions to personalize talk, asking unprompted emotion-based questions, and sustaining ongoing dialogue about matters personal to the user. The statute also excludes customer-service and information-only systems, and systems primarily designed for efficiency, research or technical assistance. Ask counsel to apply the text to your product. New York Senate bill text.
Does Maine apply only to companion products?
The sentence we read covers any chatbot or other computer technology used in trade and commerce with a consumer in a manner that may mislead a reasonable consumer into believing they are engaging with a human. It does not mention companions. The enrolled text we read has no effective date, so confirm that before you plan around it. Maine enrolled text, as of October 31, 2026.
What about California SB 1119?
We found a hearing analysis of a bill with the subject "Companion chatbots: children's safety," which would add duties for companion chatbot operators. We could not confirm from a primary source that it was signed or when it takes effect, so check its status before you rely on it. Duobay does not build children's products. California Assembly Privacy Committee analysis.
Can Duobay tell me which law applies to my device?
No. This page is an engineering summary and does not give legal advice. Counsel decides which law applies. We help your team build the firmware behaviors and tests that counsel asks for. Tell us about your device.
If you want help turning your counsel's requirements into firmware behavior and a test plan, tell us about your device. You can also read about our AI companion device development and how our engineering gates work.
Sources
All pages accessed October 31, 2026. Laws and bills change, so re-read the primary text before you decide.
- New York Senate, S.3008-C bill text: https://legislation.nysenate.gov/pdf/bills/2025/S3008C (accessed 2026-10-31)
- Office of the Governor of New York, letter to AI companion companies: https://www.governor.ny.gov/news/governor-hochul-pens-letter-ai-companion-companies-notifying-them-safeguard-requirements-are (accessed 2026-10-31)
- Maine Legislature, Public Law ch. 294 (H.P. 1154, L.D. 1727) enrolled text: https://legislature.maine.gov/legis/bills/getPDF.asp?paper=HP1154&item=3&snum=132 (accessed 2026-10-31)
- California Assembly Privacy and Consumer Protection Committee, analysis of SB 1119 (summarizes SB 243): https://apcp.assembly.ca.gov/media/1209 (accessed 2026-10-31)
Planning a device like this? Tell us what it needs to do. A Duobay engineer replies within 1 business day. Send a project brief →