Should You Tell Recipients an Email Was Written by AI?

The question sounds like one thing and is actually two. There is the legal question — am I required to tell them? — and there is the trust question — should I tell them anyway? The two answers rarely line up. For most working email the law says nothing and the relationship says everything, and a rule that treats every AI-touched reply the same will either bury you in useless disclaimers or land you in a conversation where the recipient rightly feels misled.

This guide splits the two. First: where transparency is actually mandated as of August 2026, and where it is not. Then six questions that decide whether a specific message needs disclosure even when nothing forces it, a scoring table, two worked examples, and a verdict that says which reader we build for and which we do not.

The short answer#

For a personal, one-to-one email that you reviewed and sent yourself, no current law requires you to tell the recipient AI helped draft it. The EU AI Act's text-transparency rule, which took effect on 2 August 2026, targets public-interest content and explicitly exempts text that went through human review with editorial responsibility. US federal law is silent on AI drafting in email; CAN-SPAM governs commercial mail generally, not authorship.

The trust question is different. If your recipient would feel deceived to learn a machine wrote the words — a condolence, a personal apology, a hand-written-looking follow-up to a friend — you should disclose or not send it at all, regardless of what the law says. The rule that works in practice: a reviewed and approved message is authored by the sender, until the situation is one where the reader was entitled to assume you personally typed every sentence.

Where the law actually requires disclosure#

Regulation (EU) 2024/1689 — the AI Act — is the first broad law that names AI disclosure directly. The transparency obligations in Article 50 became applicable on 2 August 2026. They cover four situations, and personal email is not one of them:

• Chatbots and other AI systems that interact with people directly. If a natural person is talking to an AI, the deployer must make that clear at the first interaction — unless it is obvious from context.
• Synthetic image, audio or video content — deepfakes. Deployers must disclose that the content has been artificially generated or manipulated.
• Public-interest text. Text generated or manipulated by AI and published to inform the public on matters of public interest must be disclosed as artificially generated — with an exception where the text underwent human review and someone holds editorial responsibility for it.
• Machine-readable marking of AI output. Providers of generative AI systems must ensure output is marked in a machine-readable format so it can be detected downstream. This is a duty on the model provider, not on you as a sender.

US federal law does not mandate authorship disclosure for AI-drafted email. CAN-SPAM regulates commercial mail — accurate headers, no deceptive subject lines, a valid physical address, a working unsubscribe — but says nothing about how the copy was written. A few state laws (California's chatbot-disclosure statute is the oldest) require disclosure when a person is knowingly interacting with a bot for commerce or elections; none treat a reviewed reply between two people as covered.

The pattern across jurisdictions as of August 2026 is consistent: the law targets impersonation, mass automated interaction, and content passing as authoritative journalism or public information. Verify against your own regulator — this is evolving and new state and sector rules keep landing.

Six questions that actually decide it#

When the law is silent, the decision is a judgment about the relationship, and the same six questions work across almost every situation. Answer each honestly for the specific message in front of you.

• Did you read every sentence before it went out? A message you actually reviewed is a message you authored, regardless of who typed the first draft. A message that autopilot sent while you were asleep is a different artefact — and the recipient is entitled to know which one they got.
• Would a reasonable recipient assume you personally typed every word? A condolence, a personal apology, a hand-typed-looking note to a close colleague, a message to your mother — the assumption is high, the disclosure obligation is high. A quarterly customer update or a routine follow-up carries no such assumption.
• Is the message representing you as an individual, or your organisation? A shared-inbox reply from support@ carries weaker personal-authorship assumptions than a one-line note from your own address to someone you know. Organisational voice is more forgiving of AI drafting than personal voice.
• Does a contract, policy, or professional rule require disclosure? Some enterprise contracts and some professional-body rules now specify how AI may be used in client work. Check before you assume the general answer applies to your specific relationship.
• Does the message carry legal weight — advice, warranty, commitment, statement of fact? A message that could be relied on in court or in a compliance audit needs a clear authorship trail. Human sign-off before send, and an audit record showing you approved it, is what makes it defensible.
• What autonomy level actually sent it — you, an approve-before-send draft, or an autopilot with rules? The higher the autonomy, the more the recipient's assumption of personal authorship gets stretched. Autopilot replies, especially those sent while you were not present, are the case where disclosure becomes least optional.

A scoring table you can apply in ten seconds#

The six questions above map cleanly to a small scoring table. For each message, note which row it fits and follow the recommendation. This is not a legal test — it is a working rule for ordinary correspondence where you are choosing whether to add a line to your signature or a sentence in the body.

Two things about the table are worth noticing. First, most working email lands in a 'no' row. If your workflow is approve-before-send and you actually read the drafts, you are almost never in a mandated-disclosure situation. Second, autopilot changes the answer. The moment a message goes out with no human eyes on it, the recipient's reasonable assumption shifts, and a disclosure line is what keeps the exchange honest.

Two worked examples#

The rules read as abstract until you apply them. Consider two messages that arrived a minute apart on the same inbox.

The first is a routine follow-up: 'Thanks for your reply last week — here is the updated proposal, let me know if the pricing works.' AI drafted it in the sender's brand voice; the account owner reviewed and sent. No condolences, no apology, no legal advice — a working note in a business relationship where both sides know how commercial follow-ups are produced. This is a clear 'no disclosure required' case. A disclaimer would be noise and would probably read as less trustworthy, not more.

The second is a note to a customer who has just lost their business partner. An AI drafted a passable version. This is where the tool's help is a trap: the recipient will assume every sentence came from the human, and if they later learn otherwise the memory of the note curdles. Either write it yourself, or use the draft as structure and rephrase every line in your own words. If you send the AI draft as-is, disclose — the awkwardness beats being caught.

A third case is the borderline one: an autopilot reply that clears a routine shipping question overnight, in your brand voice, with no human on that specific message. It is not legally regulated in most jurisdictions today, but the recipient reasonably assumes a person answered. A small footer — 'sent automatically by our AI assistant; reply to this and it reaches a human' — both discloses and gives the reader a way to escalate. That single line converts a potential complaint into a feature.

Red flags — when to disclose without thinking about it#

Some situations are not judgment calls. If any of the following applies, disclose without hesitation, or do not send the AI draft at all:

• The message is a personal apology, condolence, congratulation on a life event, or any note carrying real emotional weight. The value of the words comes from you meaning them.
• You are being paid to write it personally — a paid newsletter under your name, a ghostwritten piece attributed to you where the client did not consent to AI use, a professional statement.
• The message contains legal advice, medical guidance, financial recommendation, or any statement someone might rely on and later dispute. Add the human sign-off record — audit trail, timestamp, approver — regardless of any disclosure to the recipient.
• It is a public-facing statement of fact — press release, official notice, policy document. The EU AI Act's public-interest rule applies, and the safer path everywhere is to disclose or to have a named editor take responsibility.
• You are automating the send with no human review on that specific message. Autopilot without disclosure is where recipient assumptions break most predictably.
• The recipient has asked before whether you use AI. Do not evade the question the second time.

What we'd pick — and where we don't fit#

This is a decision guide, not a roundup, so the recommendation is narrow: pick a workflow, then pick a tool that fits it. For readers whose answer to most rows in the table above is 'no disclosure required because I approved every send', the tool needs to make approve-before-send the default and keep an audit trail so 'I approved it' can be proven later. That is the shape we build AI Emaily around — an AI-native email client with three explicit autonomy modes (Manual, Copilot, Autopilot), mandatory human approval before any send in Copilot, and a full audit log across all three. Voice-matched drafts come from a user-set Context brain and per-client profiles, not from scraping your sent mail.

If your workflow is the other one — a fully autonomous AI persona sending outbound without a human in the loop, or a public-facing chatbot representing your brand — we are not the right shape. A dedicated agent-persona platform fits that case better, and the disclosure obligation there is unambiguous under the EU AI Act. We are built for the reviewed-and-approved case, and we are honest that Autopilot sits inside that same review-and-rules discipline rather than replacing it.

On cost: there is no public free plan — the offer is a 7-day free trial on Pro or Autopilot, card taken, $0 charged if you cancel before day 7. See AI Emaily pricing for the current numbers, and verify on our own page rather than a third-party comparison — these move.