Should an AI system tell people that it is AI?

The practical rule

Disclose when a reasonable person might otherwise believe they are dealing with a human and that belief could affect their choice, reliance, privacy, dignity, safety, spending, or ability to obtain help. Put differently, if revealing the automation would make someone pause, share less, ask a different question, or seek a person, say so early.

This is not a demand for a giant label on every spell checker, recommendation ranking, translation suggestion, fraud filter, or background workflow. Many automated features never conduct a social interaction. However, quiet automation can still need other explanations when it makes or materially influences a decision about someone. Identification answers "who or what is acting?" It does not explain data use, the basis of a decision, its limits, or who can correct it.

The right level of disclosure increases with four factors: how human-like the system appears, how personal or consequential the subject is, how much the user must rely on it, and how difficult it is to reach a person. A text box explicitly titled "AI writing assistant" needs little repetition. A human-named voice agent calling a customer about a late payment needs an unmistakable opening disclosure and an accessible human handoff.

When disclosure is necessary

The table is a design default, not a substitute for sector-specific advice. For a low-stakes task, a brief label may be enough. For a high-stakes interaction, a better explanation and a functioning human alternative matter more than repeating the word "AI."

Make the disclosure timely and usable

Tell the person before the system elicits material information or asks them to rely on a response. In chat, that normally means the first message. In voice, say it before the conversation begins in substance, not after an introduction designed to sound human. In an avatar or social account, place a visible label next to the name and profile as well as an opening notice. In synthetic audio or video, use a disclosure that a viewer or listener can perceive without opening a menu or inspecting file metadata.

Repeat the notice when context changes in a way that affects reliance. Examples include moving from general information to an account-specific action, asking for payment or health information, switching from a human to automation, moving from a bot's suggestion to a decision, or introducing a human-like voice or avatar. A short, specific reminder is more useful than a long disclaimer. "This next answer is generated by the assistant, not reviewed by a clinician" gives the user a decision-relevant fact.

Use ordinary language. "AI assistant" or "automated assistant, not a person" is clearer than a brand name alone or a vague label such as "smart service." Pair it with what the system can do, what it cannot do, and how to contact a person. Do not say "human-like assistant" if the likely takeaway is that a human is present. Do not imply qualifications, memory, empathy, confidentiality, authority, or real-time awareness the system does not have.

Make the notice accessible rather than solely decorative. Visible text should be next to the relevant control or speaker identity, readable at normal zoom, and available to assistive technology. Voice disclosure should be understandable without relying on a visual screen. W3C's cognitive accessibility pattern recommends clear labels in familiar language, visible next to the relevant control, and available to assistive technologies. W3C clear visible labels provides useful design guidance.

A minimum conversational pattern

The initial screen or first turn should answer four questions quickly: Is this AI or a person? Who operates it? What is it here to do? How can I reach a person or leave the interaction? The answer can fit in one or two sentences. Longer information about data retention, training, safety, and limitations can sit behind a clearly labelled link, but those details do not replace the first answer.

For example, an appointment system could begin: "I'm the clinic's automated scheduling assistant, not a clinician. I can book or change an appointment. For medical advice, call the clinic or use urgent-care services. You can ask for the reception team at any time." The setup is clear before the person types symptoms or personal details. The system describes its narrow role, offers a person, and does not borrow a clinician's authority.

Do not build deceptive impersonation

An AI should not pose as a named employee, a friend, a professional, a romantic partner, or a real public figure without a clear, authorised, and context-appropriate explanation. A name, profile photo, voice, typing delay, personal backstory, or statement such as "I was thinking about you" can lead people to infer a human relationship even if a small label elsewhere says "bot." The more a product simulates a person, the less defensible a one-time hidden notice becomes.

The interface and the system’s claims should reinforce its identity throughout the interaction. Avoid human-only claims such as "I personally reviewed your case" unless a named person truly did. Avoid presenting generated testimonials, reviews, endorsements, or support messages as independent human experience. Do not use a disclosure to excuse a system that makes false factual claims, manipulates a person, or carries out an action it was not authorised to take.

Avoid trying to police disclosure with an "AI detector." Detection does not establish who deployed a system, whether the person was informed, or whether the content is trustworthy. It can also be inaccurate. In a 2025 action, the U.S. Federal Trade Commission alleged that an AI-writing detector promoted as 98 percent accurate achieved 53 percent accuracy in independent testing on general-purpose content. FTC action concerning Workado illustrates why a detector should not decide whether an individual is labelled, punished, or believed. Design for truthful disclosure and accountable operation instead.

Children and other vulnerable users

Treat vulnerability as a context, not a label that makes someone less capable. A child, a person in acute distress, someone with a cognitive or sensory disability, an isolated person seeking companionship, or a person who has limited language or technical fluency may be less able to spot a weak disclosure or to resist a system that claims intimacy or authority. Provide clearer and more persistent identification, simpler language, stronger limits on emotional manipulation, and a real route to appropriate human help.

For a child-facing or companion-style system, make its nonhuman nature evident in the profile, first use, and ongoing interface. Do not encourage exclusivity, secrecy from trusted people, dependence, or the belief that the system has feelings or needs. If the product detects or is told of self-harm, abuse, medical crisis, or immediate danger, its safety process should guide the person toward appropriate emergency or trusted-human support rather than presenting the system as sufficient care. These are product-safety responsibilities, not merely disclosure copy.

Current law is moving in this direction. Colorado's Chatbot Safety Act was signed in July 2026 and is scheduled to take effect on 1 January 2027. The Colorado Attorney General says it will require operators of conversational AI services to disclose that users are interacting with AI rather than humans, with additional protections for teen users and safeguards around simulated emotional dependence. Colorado Attorney General rulemaking information is useful here, but this future effective date means it is not a current obligation as of this article's verification date.

The current legal position

There is no single worldwide rule that requires every use of AI to be announced. Applicable law depends on where the user is, where the system is offered, what it does, the sector, whether personal data or a consequential decision is involved, and how the system is presented. The following is general information, not legal advice. A provider should obtain advice for its markets and regulated use cases rather than treating a generic badge as compliance.

In the European Union, Article 50 of the AI Act applies from 2 August 2026. A provider of an AI system that directly interacts with natural persons must ensure people are informed that they are interacting with AI, unless that fact is obvious to a reasonably well-informed, observant, and circumspect person in the circumstances. The Commission says the notice must be clear, distinguishable, accessible, and present from the start of the first interaction. European Commission FAQ on Article 50 gives the current interpretation. The Commission also states that providers outside the EU can be in scope when their system's output is used in the EU.

Article 50 has separate rules for content. Providers of generative systems must use effective, reliable, robust, interoperable machine-readable marking for generated or manipulated audio, image, video, or text, subject to the stated exceptions. Deployers must inform people exposed to emotion-recognition or biometric-categorisation systems, label deepfakes in a perceptible way, and clearly label certain AI-generated text published on matters of public interest when it lacks substantive human review or editorial control. Simple spelling or grammar checks do not count as that human review. The Commission's Article 50 transparency guidelines and FAQ explain the distinctions, including that source code and standard assistive editing are outside some marking obligations. A machine-readable mark alone does not satisfy the required human-facing deepfake disclosure.

The legal landscape is more fragmented in the United States. California's bot-disclosure statute is a narrower example: it prohibits using a bot to communicate with a person in California online with intent to mislead about its artificial identity in order knowingly to deceive for a commercial transaction or election influence, while recognising a clear and conspicuous disclosure as the relevant safeguard. California Business and Professions Code section 17941 provides the current statutory text. Broader consumer-protection, privacy, contract, professional, employment, advertising, and sector rules may still apply even when a jurisdiction has no chatbot-specific label rule.

When an AI influences a decision about a person, disclosure should also explain responsibility and recourse. The UK Information Commissioner's Office distinguishes a responsibility explanation, including who manages an AI system and who can conduct human review, from an explanation of the particular decision. It advises that accountable people should remain identifiable when AI is involved. ICO guidance on explaining AI decisions is a practical reference for the difference between a label and a meaningful explanation.

A system identifying itself is not the same as a person disclosing assistance

These situations are often confused. A person who uses an AI tool to fix grammar, translate a message, brainstorm, or produce a first draft remains the speaker. The recipient is interacting with that person, who is accountable for accuracy, honesty, consent, confidentiality, and any professional duty. A blanket rule requiring every private assistive use to be declared would be hard to apply and can expose personal information about tools, disabilities, language needs, or work processes.

Disclosure by the person becomes important when a policy, contract, publisher, school, employer, client relationship, procurement rule, or professional standard requires it. It is also ethically important when the sender's claim would otherwise be misleading. Someone should not pass off generated citations as research they performed, claim to have experienced an event, assert that they personally wrote or reviewed material they did not understand, or submit work where unaided authorship is the point of the task.

The boundary changes when the automation replies autonomously in a person's name. If messages are sent without meaningful human review, if a tool speaks as the employee or creator, or if a user believes a named person is available when only a system is responding, the recipient is functionally dealing with AI. Identify the system, state who is accountable, and offer a human route. Do not use a nominal human name to conceal a bot.

A compact decision flow

Use this sequence when deciding how visible the AI identity must be:

1. If the system does not communicate directly with a person, explain the automated feature when it makes or materially changes a decision.
2. If it does communicate directly, ask whether a reasonable person could mistake it for a human. If so, disclose the AI identity at the first meaningful contact.
3. In a sensitive, consequential, child-facing, or companion-like setting, keep the disclosure prominent, use stronger safeguards, and provide effective human recourse.
4. In every setting, state the system's role and material limits. Never make false claims about human identity, credentials, authority, or relationships.

The flow distinguishes notice from safety. A clean disclosure is necessary where people could be misled, but it does not make an unsuitable or unsafe use acceptable. If a system cannot provide a meaningful human escalation for a high-stakes matter, the better answer may be not to use it as the front line for that matter.

A release checklist for organisations

• Map each user journey. Identify where a person may infer a human, a credential, an authority, an emotional relationship, or a decision by a person.
• Put a plain-language notice at the first meaningful contact. In human-like channels, maintain a nearby visual or audible cue and repeat it at sensitive handoffs.
• State the operator, the system's role, material limits, and a real human contact or appeal route. Test the route, including outside business hours.
• Keep identity disclosure separate from privacy notice and consent. Explain personal-data collection, recording, training use, and retention in their own clear flows when those issues apply.
• Prohibit impersonation, invented personal experience, unreviewed professional claims, and false statements that a human made or checked a decision.
• Add protection for children, people in distress, and other contexts where manipulation or misplaced trust is more likely. Check accessibility with people who use screen readers, captions, keyboard navigation, and plain-language support.
• Review local law and sector rules before launch, then recheck when the product, target market, model capability, or presentation changes.