Summary

For companies in the US and Europe about to put AI in front of customers: which rules require you to say it is AI, for which uses, from which date, and what the system has to do.

What you'll learn

  • Which EU duties already apply to chatbots, phone agents and agent-written email, and which got more time
  • When a company that has an agent built carries the duty itself, and when its vendor does
  • When AI counts as obvious and when a notice is still required
  • What a first message, a call greeting and an email label have to do
  • When AI-written text and AI images must be labelled, and when they need not be
  • Which US state laws apply to a customer-service bot, and which do not

An AI disclosure rule decides what a customer is told before they talk to, read or look at something an AI produced. For a company putting an agent in front of customers, it is a build requirement: the notice sits in the first message, the call greeting, the email header or the file's metadata, and it has to be there on the day the agent goes live.

The answer differs between the EU and the US, so both are set out here: the EU AI Act's transparency rules after the 2026 amendments, the Commission's guidelines on applying them, Germany as one example of national enforcement, and the US federal and state laws we read at their source. Each rule carries its date and a numbered source in the Notes. Nothing here is legal advice for a specific case.

Do you have to tell people it is AI?

In the EU, yes, for most customer-facing uses: from August 2, 2026 a chatbot, voice agent or email-writing agent must identify itself as AI no later than its first exchange with a person, unless that is plain from the context.[1][2] Realistic generated images, audio and video (deepfakes) carry a disclosure of their own. In the US, as of November 2026, no general federal rule exists; the duty depends on the state and the use.

The table gives the short answer per use. Each row is explained in the sections below.

UseEU rule and dateUS rules we readWhat the system should do (our summary)
Chat on your website or in your appArt. 50(1): tell users it is AI, since August 2, 2026[1]Maine: clear notice where a consumer could be misled;[16] California: no misleading bots online in sales;[11] Utah: answer honestly when asked[17]Notice in or before the first message; honest answer when asked
Phone agent, inbound or outboundArt. 50(1): notice, since August 2, 2026; the guidelines' example is a spoken statement at the start[1][5]Maine covers voice;[16] Utah when asked;[17] FCC: AI voices count as artificial voices for outbound calls[19]Spoken notice at the start; reminders in long calls
Email sent by an agentArt. 50(1): notice; the guidelines' example is a label at the top[1][5]No US rule read is written for email; Maine may reach an agent that converses with a consumer by email (our reading)[16]AI label at the top of each agent-sent email
Email you draft with AI and send yourselfNo Art. 50(1) duty: a person sends it (our reading)[5]No rule read that requires a labelNone from these rules
Published text: blog, product copy, adsArt. 50(4) covers only public-interest text, and not if a person reviewed it[1][5]No general rule readLabel only public-interest text without human review
Realistic images, audio, video (deepfakes)Art. 50(4): disclose, since August 2, 2026[1]California AI Transparency Act: generator providers embed a hidden disclosure, since August 2, 2026[12]Visible label on deepfakes; machine-readable mark from the provider

Two dates are often confused. The duty to say "this is AI" in a chat, a call or an agent's email has applied since August 2, 2026, with no transition. The extra time to December 2, 2026 covers only the machine-readable marking of generated content, for systems placed on the market before August 2.[4][5]

Who has to tell them: you or your vendor?

Under Article 50(1) of the EU AI Act, the provider of the AI system has to tell people they are talking to AI, and the provider can be the company that runs the agent under its own name, not only its software vendor.[1][5] The Commission's guidelines of July 20, 2026 read "provider" broadly: anyone who develops an AI system "or has it developed" and puts it into service under its own name or trademark. Their own example is a company that builds a chatbot in-house and puts it into service for its own use under its name.[5]

The definition also covers having a system developed, so a company that commissions an agent and runs it under its own brand falls within its wording, even though the guidelines' worked example is the in-house case. Such a company is then likely the provider of that agent and carries the Article 50(1) duty itself (our reading). The guidelines add a second case: a company that takes an existing generative system, modifies it, for example with new training data, and puts it into service under its own name becomes the provider of the new system.[5]

Most customer-facing agents sit on a model or voice platform from another company. Our reading of the guidelines is that a business presenting such an agent to customers under its own brand is likely the provider for the agent's notice, while the platform vendor remains the provider of its own model. The guidelines do not settle every platform set-up, and the Commission notes that only the Court of Justice can give a binding reading of the Act.[10] If you buy a finished product that customers use under the vendor's name, the vendor is the provider and you are a deployer.

Deployers have separate duties under Article 50(3) and (4); see the EU AI Act deployer guide.[1][5]

The practical consequence is the same either way. Whoever builds the agent should build the notice into it, because a provider cannot hand the chat duty to a line in the customer's terms.

When is it obvious that it is AI?

The EU AI Act drops the notice only where it is obvious to "a reasonably well-informed, observant and circumspect" person that they are dealing with AI.[1] The Commission's guidelines of July 20, 2026 say this exception "should be interpreted restrictively" and limit it to cases where almost no doubt is left, judged from the point of view of the intended and foreseeable audience.[5]

The guidelines give examples on both sides:

SituationHow the Commission guidelines treat it
Coding assistant used by professional developersListed example of obvious AI
Internal assistant for trained, AI-literate staffListed example of obvious AI
Chatbot in a website helpdesk or online platformListed example of AI that is not obvious
Phone agent with a realistic human-sounding voiceA factor that may weigh against obviousness
Chat with a human name or profile pictureA factor that may weigh against obviousness

Source for all rows: the Commission's Article 50 guidelines.[5]

A customer-service chatbot is not obvious AI under the Commission's guidelines. A human name, a photo or a natural voice can weigh further against obviousness, so they make the notice more necessary, not less.

The guidelines also list systems outside the chatbot duty altogether, because people do not interact with them directly in that sense: recommender systems, spam filters, translation and transcription tools, search without generated answers, autocompletion, back-end decision support and product visualisation such as virtual try-on.[5] An email agent, a phone agent and a website chat are not on that list.

How do you disclose in chat, on the phone and by email?

Each channel gets its own form of notice (table below), at the latest at the first interaction under Article 50(5); a line in the terms is not enough.[1][5]

ChannelExample the Commission guidelines giveOur example wording
Website or app chatA written notice at the start, such as "You are interacting with an AI system""You are chatting with an AI agent. Ask for a person at any time."
Phone callAn explicit spoken statement at the beginning, with reminders in longer interactions"This is an AI agent answering for [company]."
Email sent by an agentAn AI label at the top of the email"Written and sent by an AI agent for [company]."
Any channel, when askedDisclosing its nature when asked"Yes, I am an AI agent."

Source: the Commission's guidelines;[5] the example wording is ours, not a legal formula.

The guidelines name what fails: wording only in the terms, metadata or watermarks as the only notice, ambiguous words such as "assistant", and generic notices such as "Services on this website use AI".[5] The word "assistant" matters for naming: a bot called "your assistant" has not told anyone it is AI.

One notice usually suffices. The guidelines say a single, prominent notice before the first interaction "is likely to suffice in most instances". They expect reminders in riskier contexts: financial advice, insurance, legal and health matters, complaints handling, vulnerable people, and AI agents acting on someone's behalf.[5] Where children or people with disabilities are likely users, the notice must be accessible to them.[1][5]

The system also has to answer honestly when asked about its nature.[5] Utah's law makes the same point for consumer transactions in the US.[17]

Consumer-contract directives may also require saying, before signing, that a service is AI-run.[5]

Call recording and data notices are in what an AI phone assistant has to tell callers.

Must AI-written text, images and voices be labelled?

Since August 2, 2026 deployers label deepfakes, and AI text only if it informs the public on public-interest matters.[1]

Ads and product pages sit outside the text duty unless they claim things on health, consumer safety or sustainability.[5]

Which AI images, audio and video are deepfakes

A deepfake, in the AI Act, is generated or manipulated image, audio or video that resembles existing persons, objects, places, entities or events and would falsely appear authentic. Both conditions must be met. The guidelines exclude clearly unrealistic content and include depictions of professional or consumer services, and a cloned voice of a real person is within scope.[5] A realistic AI photo of "your team at work", a generated video testimonial or a cloned voice of your managing director therefore needs a visible disclosure. Evidently artistic, satirical or fictional work gets a lighter disclosure that does not spoil it.[1]

Which AI-written text has to be labelled

The text duty applies only when three conditions meet: the text is published to a broad audience, it is intended to inform the public, and it concerns a matter of public interest.[5] The guidelines' in-scope examples include an AI summary of a town-council decision in a newspaper and listed companies' investor reports. A chatbot answer seen only by the person who asked is out of scope.[5]

A named editor who checks the facts removes the text duty; a spell check does not.[5][6] In our reading, ordinary marketing and blog text sits outside it. Google's view of AI content is a separate question: what Google says about AI-written content.

Nothing generated before August 2, 2026 needs a retroactive label. In-scope text generated earlier but published on or after that date does.[5]

What changes on December 2, 2026?

On December 2, 2026, providers of generative AI systems placed on the market before August 2, 2026 must have machine-readable marking in place for the audio, images, video and text those systems generate. The date comes from a new Article 111(4) of the AI Act, added by Regulation (EU) 2026/1744 (the "Digital Omnibus on AI"), published in the Official Journal on July 24, 2026 and in force since July 27.[4]

The deadline covers one duty: Article 50(2), which requires generated output to be marked "in a machine-readable format and detectable as artificially generated or manipulated".[1] Systems placed on the market from August 2, 2026 have had to comply from that date. The Omnibus calls the delay "a transitional period of four months".[4] It changed no other part of Article 50's duties.[4]

The December deadline is for watermarks in generated files. The "this is AI" notice in a chat or call has applied since August 2.

The Commission's guidelines spell out the consequence for agents that both chat and generate content: the transition covers only the marking, "while compliance with the disclosure obligation for AI systems directly interacting with natural persons must be ensured as of 2 August 2026".[5]

Marking is mainly the model vendor's job, because it is done in the system that generates the file. A company running an agent that generates images or audio under its own name should confirm with the vendor that the output carries the marking, since the Commission's Q&A says deployers cannot simply rely on the provider's marking for their own deepfake disclosure.[6]

The Code of Practice and the EU icons

The final Code of Practice on Transparency of AI-Generated Content (June 10, 2026) covers marking and the deepfake and public-interest text labels, with the form and timing rules of Article 50(5). Signing is voluntary; non-signatories must show compliance another way.[7]

The Commission also offers icons for labelling AI-generated content. Its page, last updated October 13, 2026, is explicit: "The use of these EU icons is optional, but the labelling requirements under Article 50 AI Act are not."[8]

What are the fines, and who enforces them?

Up to €15 million or 3% of turnover, whichever is higher; SMEs, start-ups and small mid-caps get the lower.[3]

No published Article 50 fine turned up by November 2026. Penalties across EU laws: AI fines in Europe.

Each member state's market-surveillance authority enforces the duties; the German set-up is below.[9]

The guidelines are not binding: only the Court of Justice can interpret the Act finally, the Bundesnetzagentur says, but supervisors start from them.[10]

What do US rules say about chatbots and voice agents?

As of November 2026, no US federal law requires businesses in general to tell people they are dealing with AI. Deception rules apply as usual. The duties come from state laws, which differ by use: California's bot law since 2019, Maine's chatbot rule since 2025, Utah's disclosure rule since May 2025, and newer companion-chatbot laws, of which the three we read exclude customer service.[11][13][14][15][16][17]

California: misleading bots online in sales

California's bot law (Business and Professions Code § 17941, operative since July 1, 2019) bars a bot from misleading a person in California online about its artificial identity to drive a sale or purchase or to influence a vote. A clear and conspicuous bot disclosure removes liability.[11] "Online" means public-facing websites and apps; our reading is that the law does not reach phone calls.

California AI Transparency Act: hidden marks in generated media

The California AI Transparency Act, operative since August 2, 2026, applies to "covered providers": companies that create or produce a generative AI system publicly accessible in California. They must embed a hidden ("latent") disclosure in generated images, video and audio, where technically feasible, and offer a free tool to detect it. Text is not covered. Penalties are $5,000 per violation, enforced by public prosecutors.[12]

The current code, as amended by SB 1000 with effect from September 30, 2026, contains no user threshold.[12] Large online platforms (over 2 million monthly users) have duties from January 1, 2027, and camera and recorder makers from 2028.[12] Our reading: a business that uses a vendor's image or voice generator is not the covered provider; the vendor is.

Maine and Utah: rules that reach ordinary businesses

Maine's law (10 M.R.S. § 1500-DD, enacted 2025) is the broadest rule we read. A business may not use an AI chatbot "or any other computer technology" in trade with a consumer in a way that may make a reasonable consumer believe they are dealing with a human, unless it tells them clearly. It covers text and voice, and a breach counts as an unfair trade practice.[16]

Utah requires a business using generative AI with a consumer to disclose that it is AI if the consumer clearly asks. Licensed professions must disclose at the start of high-risk interactions. A notice at the outset and throughout gives a safe harbour; fines reach $2,500 per violation.[17]

Companion-chatbot laws exclude customer service

California's SB 243 (since January 1, 2026), Washington's ESHB 2225 (from January 1, 2027) and New York's General Business Law Article 47 cover companion chatbots and exclude bots used only for customer service.[13][14][15] Oregon (January 1, 2027), Idaho and Nebraska (both July 1, 2027) passed similar laws; we have not read their enacted texts, so we do not say whether they reach a customer-service bot.[20][21] Colorado's 2026 law concerns automated decisions, not chatbot notices.[18]

Phone calls: the FCC's AI voice ruling

For outbound AI-voice calls the FCC's February 2, 2024 ruling applies; see the rules for cold outreach.[19]

What does Germany add to the EU rule?

Germany's KI-MIG (in force July 29, 2026) adds supervision only, with no separate disclosure rule. The Bundesnetzagentur is the authority; BaFin and other sector supervisors keep their fields. Its national fines concern other articles, so for Article 50 Article 99 applies directly.[9] No German statute separately requires AI voice agents to disclose themselves, as of November 2026.

Germany is one example of how member states enforce the same duty. The Bundesnetzagentur's guidance pages restate the Commission's line: chatbots must say they are AI clearly and understandably before the interaction, wording in terms of use alone is not enough, and deepfakes must be clearly labelled.[10] Data-protection authorities are involved only where their remit is affected.[9]

Two German rules not written for AI apply alongside: § 5a(4) UWG (hidden commercial purpose), § 7 UWG (consent for advertising calls).[22]

Whether an AI sales-call agent is an "automatic calling machine" under § 7 is unsettled in our sources.

What should the system do and record?

An agent facing EU customers needs the notice from its first interaction, as the table of channels above shows.[1][5] The list below is our recommendation for building it into the system.

1

Notice in the first message

The chat opens with a line that says it is AI, visible before the customer types. Not only in the terms, not only "assistant".
2

Spoken notice at the start of each call

The phone agent says it is an AI in its greeting, and repeats it in long calls.
3

Label on every agent-sent email

Emails the agent sends on its own carry an AI label at the top, set in the template so no draft goes out without it.
4

Honest answer when asked

The agent is instructed never to claim to be human and to say it is AI when asked. A handover to a person on request is good practice, not a requirement of these rules.
5

Labels and marks on generated media

Realistic images, audio or video the agent creates are labelled as AI-generated, and the generator's machine-readable mark is kept intact.
6

A record of the notice

The system logs which notice text was shown, in which version and when.

No source we read requires a log of disclosures. Keeping one is still the simplest way to show a supervisor, or a customer who complains, that the notice was there on a given day. Building the notice into the system rather than relying on staff to add it is the same principle as building approvals into an agent: rules that matter are enforced in code.

Which dates come next?

The next AI disclosure date in the EU is December 2, 2026, when machine-readable marking becomes mandatory for generative systems placed on the market before August 2, 2026.[4] In the US, several state laws take effect on January 1, 2027 and July 1, 2027.[12][14][20][21]

DateWhat appliesWhere
July 1, 2019California bot law in forceCalifornia[11]
January 1, 2026Companion-chatbot law (customer service excluded)California[13]
July 27, 2026Digital Omnibus on AI in forceEU[4]
July 29, 2026KI-MIG in force; Bundesnetzagentur supervisesGermany[9]
August 2, 2026Art. 50 applies: chatbots, voice, deepfakes, public-interest text; new generative systems markedEU[2]
August 2, 2026AI Transparency Act operativeCalifornia[12]
December 2, 2026Marking for generative systems placed on the market before August 2, 2026EU[4]
January 1, 2027Platform provenance dutiesCalifornia[12]
January 1, 2027Companion-chatbot lawsWashington,[14] Oregon[20]
July 1, 2027Conversational AI lawsNebraska, Idaho[21]
January 1, 2028Hidden disclosure in new cameras and recordersCalifornia[12]

These dates move. California amended its AI Transparency Act in 2025 and again in 2026, and the Commission updates its guidance pages, so check the cited source before relying on a date.

Frequently asked questions

Do I have to tell customers that my chatbot is AI? In the EU, yes. Since August 2, 2026, Article 50(1) of the EU AI Act requires AI systems that interact directly with people to tell them they are interacting with AI, at the latest at the first interaction, unless that is obvious. The Commission's guidelines of July 20, 2026 say helpdesk chatbots are not obvious. In the US, Maine requires a clear notice where a chatbot could mislead a consumer, California prohibits misleading bots online in sales, and Utah requires an honest answer when a consumer asks.

Does the December 2, 2026 AI Act deadline apply to chatbots? No. Regulation (EU) 2026/1744 gave providers of generative AI systems placed on the market before August 2, 2026 until December 2, 2026 to add machine-readable marking to generated audio, images, video and text. The duty to tell people they are talking to an AI has applied since August 2, 2026 with no transition. The Commission's guidelines state this explicitly for systems that both chat and generate content.

Is there a US law that requires AI disclosure? There is no general federal law, as of November 2026. State laws apply by use: California prohibits bots that mislead people online about being a bot to drive a sale (since 2019), Maine requires a clear notice where an AI chatbot could make a consumer believe they are dealing with a person, and Utah requires disclosure when a consumer clearly asks. California, Washington and New York's companion-chatbot laws exclude bots used only for customer service.

Do I have to label AI-written blog posts and product texts? Usually not under the EU AI Act. Article 50(4) covers AI-generated text published to inform the public on matters of public interest, and even then not if a person reviewed it and someone holds editorial responsibility. The Commission's guidelines name advertisements and product descriptions as outside that duty, unless they make claims about health, consumer safety or sustainability. Realistic AI images, audio and video of real people, places or events are a different case: they must be disclosed as deepfakes.

Notes

  1. Regulation (EU) 2024/1689 (AI Act), Article 50, consolidated text as at July 27, 2026. European Commission, AI Act Service Desk, ai-act-service-desk.ec.europa.eu/en/ai-act/article-50. Checked November 17, 2026.
  2. Regulation (EU) 2024/1689 (AI Act), Article 113, entry into force and application. ai-act-service-desk.ec.europa.eu/en/ai-act/article-113. Checked November 17, 2026.
  3. Regulation (EU) 2024/1689 (AI Act), Article 99 (4)(g), (6) and (6a), penalties. ai-act-service-desk.ec.europa.eu/en/ai-act/article-99. Checked November 17, 2026.
  4. Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 (Digital Omnibus on AI), Official Journal L, 2026/1744, July 24, 2026: recital 38; Article 1, points 20 and 39(b) (new Article 111(4)); Article 4. eur-lex.europa.eu/eli/reg/2026/1744/oj/eng. Checked November 17, 2026.
  5. European Commission, "Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689", C(2026) 5054 final, July 20, 2026: paragraphs 10–12, 31, 33–45, 50, 153–154, sections on deep fakes and 6.2. ai-act-service-desk.ec.europa.eu (PDF); library page digital-strategy.ec.europa.eu, last updated November 5, 2026. Checked November 17, 2026.
  6. European Commission, "Transparency obligations under Article 50 AI Act" (questions and answers), last updated July 24, 2026. digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act. Checked November 17, 2026.
  7. European Commission, "Code of Practice on Transparency of AI-Generated Content", last updated October 19, 2026. digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content. Checked November 17, 2026.
  8. European Commission, "EU icons for labelling AI-generated content", last updated October 13, 2026. digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content. Checked November 17, 2026.
  9. Gesetz zur Marktüberwachung und Innovationsförderung von künstlicher Intelligenz (KI-MIG) of July 22, 2026, BGBl. 2026 I No. 223, §§ 2, 4, 6, 8, 9, 15, 16. gesetze-im-internet.de/ki-mig. Checked November 17, 2026.
  10. Bundesnetzagentur, "Transparenzpflichten" and "Leitlinien zu den Transparenzpflichten". bundesnetzagentur.de, Leitlinien. Checked November 17, 2026.
  11. California Business and Professions Code §§ 17940–17943 (SB 1001, Stats. 2018, ch. 892). leginfo.legislature.ca.gov. Checked November 17, 2026.
  12. California Business and Professions Code §§ 22757–22757.6, California AI Transparency Act (SB 942, Stats. 2024, ch. 291; AB 853, Stats. 2025, ch. 674; SB 1000, Stats. 2026, ch. 861). leginfo.legislature.ca.gov. Checked November 17, 2026.
  13. California Business and Professions Code §§ 22601–22606, companion chatbots (SB 243, Stats. 2025, ch. 677). leginfo.legislature.ca.gov. Checked November 17, 2026.
  14. Washington State Legislature, Engrossed Substitute House Bill 2225, Chapter 168, Laws of 2026, sections 2 and 3. lawfilesext.leg.wa.gov (PDF). Checked November 17, 2026.
  15. New York General Business Law §§ 1700 and 1702 (Article 47, AI companion models). nysenate.gov/legislation/laws/GBS/1700, /GBS/1702. Checked November 17, 2026.
  16. Maine Revised Statutes, Title 10, § 1500-DD (P.L. 2025, c. 294; renumbered by R.R. 2025, c. 1, Pt. A, § 16). legislature.maine.gov. Checked November 19, 2026.
  17. Utah Legislature, S.B. 226 (2025), Artificial Intelligence Consumer Protection Amendments, enrolled copy, §§ 13-75-103 to 13-75-105. le.utah.gov (PDF). Checked November 17, 2026.
  18. Colorado General Assembly, SB 26-189, Automated Decision-Making Technology, signed May 14, 2026. leg.colorado.gov/bills/sb26-189. Checked November 17, 2026.
  19. Federal Communications Commission, Declaratory Ruling FCC 24-17, CG Docket No. 23-362, adopted February 2, 2024. docs.fcc.gov (PDF). Checked November 17, 2026.
  20. Oregon SB 1546 (2026 Oregon Laws, ch. 85), AI companions, as summarised by Miller Nash, "Oregon's new AI companion law: what you need to know". millernash.com. Enacted text not read. Checked November 17, 2026.
  21. Idaho Legislature, S1297 (2026), bill page legislature.idaho.gov; Nebraska LB 525 (Conversational AI Safety Act), as summarised by Hooper Lundy, hooperlundy.com. Enacted texts not read. Checked November 17, 2026.
  22. German Act against Unfair Competition (UWG), § 5a and § 7. gesetze-im-internet.de/uwg_2004/__5a.html, __7.html. Checked November 17, 2026.