AI chat on your website and privacy: what to tell visitors

TL;DR

• Visitors need to know three things before they type: that they are talking to an AI, what happens to what they type, and how to reach a person.
• In the EU, the AI disclosure has been a legal requirement since August 2, 2026 (AI Act, Article 50), and the GDPR already required a clear privacy notice for anything you collect in a chat.
• Put a one-line disclosure in the chat window, a short "AI chat" section in your privacy policy, a separate opt-in for marketing, and a retention period you can actually honor. Sample text is below.

This article is general information for website owners, not legal advice. Have your counsel check the final wording for your markets.

Why a chat window is a privacy decision

A chat box looks like a small interface change. From a privacy point of view it is a new way to collect personal data, and often a lot of it: people type names, emails, order numbers, details about their company and sometimes things you never asked for. France's data protection authority, the CNIL, points out that a chatbot usually processes personal data even when visitors don't create an account, simply because the conversation is recorded (CNIL).

An AI chat adds two more questions visitors reasonably ask: "Is this a person?" and "Will my words be used to train something?" Answering both up front is a legal requirement in some places and good manners everywhere.

What the EU AI Act requires: say it is an AI

Article 50(1) of the AI Act requires that AI systems intended to interact directly with people are designed so those people are informed they are dealing with an AI, unless that is obvious to a reasonably well-informed person from the context. Article 50(5) adds that the information must be given "in a clear and distinguishable manner at the latest at the time of the first interaction," and must meet accessibility requirements (AI Act Service Desk).

Three dates are worth knowing as of October 2026:

• August 2, 2026. Article 50's transparency duties started to apply.
• July 20, 2026. The European Commission published guidelines on these obligations, covering who must comply for interactive AI systems and for AI-generated content (European Commission). They are not binding, but they show how regulators read the rules.
• July 27, 2026. The "AI Omnibus," Regulation (EU) 2026/1744, came into force. It postponed several high-risk obligations but left the substance of the Article 50 transparency duties unchanged, according to Goodwin's analysis.

Formally, Article 50(1) is addressed to the provider, the company that builds the AI system. If you use a vendor's chat widget, the vendor is usually the provider and you are the deployer. In practice you control the greeting, the widget title and the settings, so make sure the disclosure is visible on your site rather than assuming the vendor handled it. If you built your own assistant, you are likely the provider yourself.

What a good AI disclosure looks like

• In the window, not only in the policy. A label under the chat title ("AI assistant") plus a first message that says it plainly.
• Before the first answer. Not after the visitor has already shared details.
• Truthful when asked. If a visitor asks "Am I talking to a real person?", the answer is always a direct "No, I'm an AI assistant," followed by how to reach a person.
• Clear about limits. One sentence that answers come from your website content and can contain mistakes.
• Accessible. Real text that screen readers announce, not text baked into an image.

What the GDPR requires: tell people what you collect and why

The AI Act adds the "it's an AI" duty. The data protection duties come from the GDPR, and they apply whether or not the chat uses AI (GDPR).

Who is responsible

You, the website owner, normally decide why chat data is collected and what happens to it, which makes you the controller. The chat vendor processes data on your behalf, which makes it a processor, and Article 28 requires a contract with it: a data processing agreement. Ask the vendor for its agreement and its list of sub-processors, including the company that runs the large language model behind the answers.

What visitors must be told

Article 13 lists what you must tell people when you collect their data. For a chat, that means:

• who you are and how to contact you (and your data protection officer, if you have one);
• what the chat collects and why: answering questions, following up on requests, improving answers;
• the legal basis for each purpose;
• who receives the data: the chat vendor, its model provider, your CRM;
• whether data leaves the EU and on what basis (the GDPR's Chapter V rules on transfers);
• how long you keep transcripts;
• the visitor's rights, including access, deletion and objection, and the right to complain to a supervisory authority.

Choosing a legal basis

Answering a question a visitor asked is usually handled under legitimate interests (Article 6(1)(f)) or as steps taken at the visitor's request before a contract (Article 6(1)(b)). Using the same email address for marketing is a separate purpose and, in many EU countries, needs opt-in consent under the ePrivacy rules. Keep the two apart: answer the question, then offer marketing as an optional, unticked box.

Collect less

The data minimization principle in Article 5 says personal data must be "adequate, relevant and limited to what is necessary." For a chat this means: don't require an email before answering, ask only for what you will use, and don't ask a visitor for their phone number "just in case." If you want to qualify leads, see how to qualify inbound leads without annoying visitors, which keeps questions to the minimum.

Sensitive details people volunteer

Visitors sometimes type health information, financial details or other special categories of data under Article 9 without being asked. The CNIL asks operators to pay particular attention to this. Add a line asking visitors not to share sensitive personal information, and make sure your team can delete a transcript on request.

Retention you can actually honor

The CNIL's advice is to keep conversation data only as long as its purpose requires: delete it once the visitor's question has been answered, and keep it longer only when the request needs follow-up, until that request is closed (CNIL). Pick a default period, configure it in your chat tool, and write the same number in your policy.

Training on conversations

Say whether transcripts are used to improve the service, and by whom. "We review conversations to improve our answers" is different from "our vendor uses your conversations to train its models." Check your vendor's terms before you write this sentence, not after.

Cookies and local storage: when the chat needs consent

Article 5(3) of the ePrivacy Directive requires consent before storing information on, or reading it from, a visitor's device, with an exception for storage that is strictly necessary for a service the visitor explicitly requested (ePrivacy Directive). The European Data Protection Board's guidelines make clear the rule is technology-neutral: it is not only about cookies, so browser local storage used by a chat widget falls under it too (EDPB).

The CNIL gives a practical rule for chatbots: storage placed only when the visitor activates the chat, for example by clicking the chat button, and used only to make the chat work, is strictly necessary and does not need consent. Storage placed before activation, or used for anything else, such as analytics or advertising, needs prior consent (CNIL).

Ask your vendor exactly what the widget stores, when it stores it and why. Then list it in your cookie notice.

Outside the EU: a quick look at California

• Bot disclosure. California's SB 1001 makes it unlawful to use a bot to mislead someone in California about its artificial identity in order to incentivize a sale, and says a disclosure must be "clear, conspicuous, and reasonably designed" to inform people they are talking to a bot (SB 1001). A sales-focused chat on a B2B site is exactly where this matters.
• Notice at collection. If the California Consumer Privacy Act applies to your business, its regulations require a notice at or before the point where you collect personal information, which online usually means a link to the relevant section of your privacy policy (CPPA). A link in the chat window does this.

Other states and countries have their own rules. The same three habits cover most of them: say it is an AI, link to a clear notice, and collect only what you need.

Sample text you can adapt

In the chat window

Title:      Ask Acme
Subtitle:   AI assistant · answers from acme.example and can make mistakes
Footer:     Please don't share sensitive personal information. Privacy notice →

First message

Hi, I'm Acme's AI assistant. I answer from our website and docs and I link
to the pages I use. I can also connect you with our team. What would you
like to know?

When a visitor asks if it is human

No, I'm an AI assistant. If you'd rather talk to a person, I can pass this
conversation to our team. They usually reply within one business day.

When asking for contact details

We'll use your email only to reply about this conversation.
[ ] Also send me occasional product news (optional, unsubscribe anytime)
Privacy notice →

A section for your privacy policy

AI chat on our website
We offer an AI assistant that answers questions from our website content.
It is operated for us by [vendor], acting as our processor, using AI models
from [model provider].
What we collect: the messages you type, the page you were on, and any
contact details you choose to share.
Why: to answer your questions (legitimate interests), to follow up on
requests you make (steps before a contract), and, only if you opt in,
to send you product news (consent).
Where: [regions]; transfers outside the EU rely on [adequacy decision /
standard contractual clauses].
How long: [30] days for answered questions; until a request is closed
when you ask us to follow up.
Training: [we review conversations to improve answers / your messages are
not used to train AI models].
Your rights: access, correction, deletion, objection and complaint to a
supervisory authority. Contact: [privacy email].

Replace every bracket with what is true for your setup. A short, accurate section is better than a long one copied from someone else.

A pre-launch checklist

The last row matters as much as the legal ones. Good human handoff is part of being honest about what an AI agent can and cannot do. So is accuracy: an assistant that says it is an AI but still invents answers has only solved half the problem, which is why grounded answers belong on the same checklist.

Telling visitors the truth costs one line in the chat window and one section in your policy. Do it before launch, keep the wording in sync with how the chat is actually configured, and review both whenever you change vendors, models or what the chat collects.