WhatsApp AI Content Label: Who Writes the AI Disclosure Under EU AI Act Article 50

Does WhatsApp require an AI content label on automated messages?

No. WhatsApp's published terms set opt-in, template and prohibited-content rules but no AI-disclosure clause for third-party automation. The “AI from Meta” notice applies to Meta's own model only. Any AI disclosure in your thread is one you wrote — and in the EU, one the law now requires.

That reading is corroborated, not invented. In a review published on 20 September 2026, mesej.io searched WhatsApp's policy and terms and reported: “On WhatsApp the same company asks for nothing of the kind: we searched its policy and its terms and found no disclosure requirement.” The same review notes Instagram's Messaging API requires automated chats to disclose only “when required by applicable law” — a clause pointing at statute, not at the platform.

Follow the pointer. In the EU, statute speaks: Article 50 of Regulation (EU) 2024/1689 has applied since 2 August 2026. It requires that people be told they are interacting with an AI system unless that is obvious in context. A business running its own model is the party who must tell them.

Be blunt about the failure mode. An operator who assumes the platform handles disclosure ships an agent that discloses nothing. There is no badge to inherit and no vendor switch to flip. WhatsApp's policy can also change; both readings above carry a date, so re-read the current text before relying on the absence.

Four things get called “the AI label”. Which one applies to your WhatsApp agent?

Article 50 of the EU AI Act is the only one of the four that creates a disclosure duty for a self-hosted WhatsApp agent, and it has applied since 2 August 2026. The others are a feed badge on Meta's platforms, a marking Meta applies to its own AI, and a contract for AI-assistant sellers.

1. The AI info badge on Facebook, Instagram and Threads. Meta attaches it when it detects C2PA or IPTC provenance signals, or when the poster self-discloses. Proofcan's platform tracker (read 3 November 2026) and RecurPost's analysis of 24 September 2026 both describe a feed-level marking. It does not render inside a WhatsApp chat, and it binds publishers.

2. The “AI from Meta” in-chat notice. WhatsApp's Help Center (faq.whatsapp.com/6485307734928964, read November 2026) documents a notice reading “AI from Meta receives chats to improve AI quality and generate messages for this business”, plus an AI marking on each generated message. Published scope: Meta's own AI, selected businesses, limited countries, a waitlist, partial language coverage. It binds Meta.

3. WhatsApp's AI Provider terms, reported effective 15 January 2026. Separate terms for businesses whose product is an AI assistant sold to end users. A narrow class of sellers, not every business that automates replies.

4. EU AI Act Article 50. Statute, in force for these systems since 2 August 2026.

Two of the four are Meta labels, and neither is offered to a business running its own model. One is a contract. One is law — and law is the one that follows a self-hosted stack home.

Which route fits? A one-line verdict on each of the four ways to run an AI agent on WhatsApp

Four routes reach an AI agent on WhatsApp: Meta's own AI on the Cloud API, a BSP layer, a self-hosted open-source API, and SocialMate. They differ less on capability than on two questions — who writes the AI disclosure, and whether cost scales with volume. One line each.

• Meta's own AI on the Cloud API. You inherit the in-chat marking and Meta's scope limits together; metered pricing and business verification apply.
• A BSP layer such as WATI or Twilio. Fastest to a working bot without engineers, but the disclosure is still yours and it lives inside a vendor's system. BSPs add a per-message or per-conversation platform fee on top of Meta's charges — check the vendor's current pricing before you model a volume like 2,000 messages a month.
• A self-hosted open-source API such as Evolution API or WAHA. No per-message fee and full ownership, and every risk control plus the disclosure is yours to build and maintain.
• SocialMate, self-hosted. Flat licence, your number, your local storage, and the Article 50 line seeded into the system prompt by Agent Studio, on by default on Free and Pro. Bans remain possible and warming is never optional.

The table below holds the detail: pricing model, real cost at 2,000 messages a month, who writes the disclosure, and account risk for each route.

Why a self-hosted agent will never carry the “AI from Meta” marking

The “AI from Meta” notice is generated inside Meta's infrastructure for messages Meta's model writes. WhatsApp's Help Center scopes it to selected businesses, limited countries, a waitlist and partial language coverage. A model running on your own machine sits outside that pipeline, so no badge renders above its reply.

So the disclosure, if one appears in the thread, is yours. No platform checkbox exists for this, and nothing hands a third-party agent a compliance toggle.

That is the structural consequence of self-hosting: you keep the data and you keep the duty.

The honest trade. A business on Meta's own business AI inherits the marking for free, and inherits Meta's country limits, language gaps and waitlist with it. A business running its own LLM gets none of that marking and gains what a transparency duty actually asks for: control of wording, timing and placement.

That control is only worth something if you use it. A badge you cannot configure and a sentence you never wrote are equally absent from the customer's screen.

What did WhatsApp's AI Provider terms of 15 January 2026 actually change?

Reported as effective 15 January 2026, WhatsApp's AI Provider terms target businesses that sell an AI assistant as their product, not every business that automates replies. They create no consumer-facing label and transfer no transparency duty to WhatsApp. Read them for your own case; this is reporting, not legal advice.

Coverage of the change was louder than the change. moca-tech.net's analysis of 24 October 2025 and setscale.io's review of 30 September 2026 both characterise the terms as narrower than the headlines suggested — a separate contracting track for AI-assistant sellers on the WhatsApp Business Platform.

Scope matters for a founder running an internal agent. Order updates, booking confirmations and triage on your own number are not the target. A business reselling a WhatsApp AI assistant to clients may fall inside. That is the line to check against your own model.

Note what the terms do not do. They do not create a consumer-facing AI label, they do not transfer your Article 50 duty to WhatsApp, and they do not describe your obligation to the person on the other end of the chat. Treat any secondary review as a pointer to the primary text, and read the terms yourself before relying on the absence.

Does EU AI Act Article 50 require you to disclose an AI on WhatsApp?

Article 50 of Regulation (EU) 2024/1689 requires that people be informed they are interacting with an AI system unless that is obvious in context. It has applied since 2 August 2026. It is a transparency duty about interaction, not a content label — the distinction most headlines blur.

The operative wording is plain. Providers must ensure systems intended to interact directly with natural persons are designed so that those persons “are informed that they are interacting with an AI system”, unless that is obvious to a reasonably well-informed, observant and circumspect person given the circumstances and context of use. The European Commission's Article 50 FAQ at digital-strategy.ec.europa.eu is the primary explainer to read alongside the text.

Read the drafting honestly. It is written for providers, and a business that builds its own agent with its own model and prompt occupies that role in practice. Note the exception too: an instant, fluent, always-available reply to a routine question is rarely obvious as a machine to a reasonable customer.

Article 50 also carries a separate duty on marking synthetic content in machine-readable form. That duty does not require a chat bubble saying “AI”. Operators who conflate the two build the wrong control.

Penalties sit in Article 99 of the Regulation. Read Article 99 and the Commission's published guidance for the maximum fines that apply to a breach of the transparency obligations, and for the provisions that cap a fine for an SME or start-up. This is a summary of a statute, not legal advice.

Is a WhatsApp chatbot “high-risk” under the EU AI Act?

No. A conversational WhatsApp agent is not high-risk per se. Classification follows the Annex III use cases — employment screening, creditworthiness, access to essential services, education assessment and similar — not the fact that a model writes chat replies. The Article 50 duty applies regardless, so a “not high-risk” conclusion does not clear you.

That independence is the point. A small business running a booking bot sits squarely outside high-risk territory and still owes the disclosure. Compliance reviews that stop at classification routinely miss the one obligation that actually attaches.

If your agent starts scoring applicants, ranking tenants or deciding credit, the analysis changes and you need counsel rather than a blog post. Public explainers tracking the Act, such as aiactverdict.com's review updated 19 August 2026, are useful orientation and not a substitute for advice on your facts.

One practical detail operators miss. The Article 50(1) exemption turns on what a reasonable person would notice, not on what your privacy policy says. A disclosure behind a link is not a disclosure, and an FAQ page nobody opens is not a disclosure either.

Does Article 50 reach a non-EU business messaging EU customers?

Often, yes. Territorial scope is wider than establishment: Regulation (EU) 2024/1689 reaches providers and deployers established outside the Union in defined cases, including where the output produced by the system is used in the Union. A US or UK business whose agent answers an EU customer is the case the text addresses.

Where that lands exactly depends on facts: where the customer is, where the output is used, and whether you are treated as provider or deployer. Take it to counsel, and start from the Commission's Article 50 FAQ at digital-strategy.ec.europa.eu.

A second scope question: do one-way template notifications count? A utility template is not an interaction with a natural person, so the transparency duty is unlikely to attach to the send itself. The moment the customer replies and an AI answers that reply, it does.

Keep a routine. Store the prompt version, the exported workflow and a dated copy of a real thread showing the disclosure. The Act prescribes no Article 50 record format; evidence that your system was designed to inform is the practical defence.

What does an honest AI disclosure look like in a WhatsApp thread?

One plain sentence at the start of the thread, in the thread, in the customer's language: that this is an AI, on whose behalf, and how to reach a person. Not a link, not a footer, not a policy page. The statute prescribes no wording, only that the person is informed.

A workable template, adapted from the kind of copy a support team can actually maintain:

Hi — before we start: you are talking to an AI assistant on behalf of [Business]. A person can take over at any time. Reply HUMAN and I will hand you to the team.

Placement rules. Say it once at the start of a thread — the first outbound message, or the first reply to an inbound one. Say it again on request, whenever someone asks whether they are talking to a bot. Say it again whenever the content, not just the sending, was generated: a summary, a drafted answer to a complaint, a recommendation.

One rule decides most cases. If every message is written and sent by a human, there may be no AI interaction to disclose. If an AI drafts replies a person waves through, the duty can still attach. Take the facts to counsel.

Three failure modes. Disclosing nothing and assuming the platform covers it. Burying the sentence in a linked policy that never surfaces in the chat. Labelling human-written replies as AI, which teaches customers to distrust the label you depend on.

Where does the AI disclosure actually live in a self-hosted stack?

In a self-hosted stack the disclosure is not a platform setting. It is text in four places you own: the system prompt your model reads, the Agent Studio seed that writes it, the n8n AI Agent system message if you orchestrate there, and the conversation memory the agent reads back. Each must carry it.

• The system prompt. The seed instruction your model reads before every turn. Put the disclosure sentence and its trigger conditions here.
• Agent Studio. The 10-step wizard writes the EU AI Act Article 50 disclosure line into the seeded prompt, on by default on Free and Pro, then exports a paste-ready n8n workflow or an MCP config (SocialMate docs, /docs/ai-agent-prompt, read November 2026).
• The n8n AI Agent system message. The native node n8n-nodes-socialmate exposes every operation as an n8n AI-Agent tool (usableAsTool), so the agent's system message is the enforcement point when it chooses actions itself (/docs/n8n).
• The memory the agent reads back. GET /v1/accounts/:id/ai-context returns the thread role-mapped and token-windowed, so the agent can see the disclosure was made in turn one and need not repeat it clumsily in turn nine.

That last point is a real failure mode, not a theoretical one. A token-windowed transcript eventually drops the opening turn, and the agent loses sight of its own disclosure. Long-running threads need the sentence restated in the system prompt, not only once in history.

Two honest notes. SocialMate runs no model: no replies, summaries or triage are generated on our side — that remains coming soon. Agent Memory stores what your own model produced; it produces nothing. And the MCP server's socialmate_human_agent prompt (setup guide at /docs/mcp-server) teaches reply cadence — mark read, recall the thread, show typing, reply — which is free on every tier but is not your legal disclosure. Paste your sentence alongside it. Human-like pacing is a delivery behaviour; impersonating a person when asked directly is a disclosure failure.

Can you automate WhatsApp without Meta approval, and does that change the disclosure?

You can run WhatsApp automation on your own number without WhatsApp Business verification or Meta approval, because the connection is made from your own machine rather than through the Cloud API. That removes an approval step. It removes no disclosure duty, no consent requirement and no ban risk — those follow the operator, not the approval status.

Be precise about what self-hosting buys. Data locality, a flat cost, and no per-message meter. SocialMate is a desktop app you run yourself — or, since v2.0, headless on a VPS via Docker or systemd, managed from the same browser admin console — and it uses the unofficial WhatsApp Web protocol. It is not the official WhatsApp Business API, and it is not affiliated with, endorsed by or partnered with WhatsApp or Meta.

The trade is real and should be named. Unofficial protocol automation can violate WhatsApp's terms, and account bans are always possible regardless of tooling. SocialMate's anti-ban engine — randomised delays, pacing profiles, per-number rate limits, session warming, read receipts and a typing indicator before replying, an adaptive throttle, a duplicate-content guard and live risk scoring — reduces that risk. It does not eliminate it. Any vendor promising otherwise is selling you a future ban.

The compliance habit that follows: message people who are already waiting on you — they ordered, booked or asked — keep consent records, and put the disclosure in the first message of the thread rather than in a policy page. That is the whole posture. It costs one sentence per conversation.

Route	Pricing model	Cost at 2,000 messages / month	Who writes the AI disclosure	Ban / account risk	Best for
Meta's own AI on the Cloud API	Per-message, metered, billed by template category and market; business verification required	2,000 × Meta's per-message rate for your category and country — it varies by market, so there is no flat figure; no flat cap either	Meta, where the in-chat marking has rolled out; you inherit its scope limits	Policy enforcement, template rejections, quality-rating limits; verification gates the account	Businesses wanting a managed agent and accepting metered cost
BSP layer (WATI, Twilio and similar)	Per-conversation or per-message, on top of Meta's fees	Meta's per-message charges for your category and market, plus the BSP's own per-message or per-conversation fee — check the vendor's current pricing; there is no flat cap	You — no Meta marking applies to third-party AI; a BSP supplies templates, not the legal duty	Same Cloud API enforcement, plus vendor policy layers and account review	Teams without engineers
Self-hosted open source (Evolution API, WAHA)	Open-source licence, no per-message fee; you pay hosting and your own maintenance	$0 in message fees; hosting and engineering time are yours to carry	You — nothing is supplied	Unofficial protocol; bans possible. Pacing, warming and rate limits are yours to build and tune	Developers who want to own the whole stack
SocialMate (self-hosted: desktop, or headless on a VPS)	Flat licence, one device included, add seats for more machines — Free $0; Pro $99/yr ($8.25/mo equivalent) or $10/mo; extra seats $4/mo each; no per-message fee	Fixed: $8.25/month equivalent on Pro Annual, or $10/month Monthly, at any volume inside your tier's daily cap	You — seeded into the system prompt by Agent Studio, Article 50 line on by default on Free and Pro	Unofficial protocol; bans possible. Anti-ban engine, adaptive throttle and 72-hour warming reduce risk, never eliminate it	Operators wanting flat cost and a visible disclosure before the first send

Ship a self-hosted WhatsApp agent with the AI disclosure already in place

1. Install the app or the server build, and accept the risk frame Download SocialMate free for Windows, macOS or Linux — no card — or self-host headless on a VPS with the Docker image or the systemd installer, managed from the same browser admin console. Before you send anything: this uses the unofficial WhatsApp Web protocol, automated use can violate WhatsApp's terms, and bans are always possible. No tool eliminates them.
2. Seed and edit the prompt in Agent Studio Run the 10-step wizard. The EU AI Act Article 50 disclosure line is already written into the seeded system prompt, on by default on Free and Pro. Rewrite it to sound like your business, and keep a dated copy of the version you shipped.
3. Export to n8n or MCP and carry the sentence across Agent Studio exports a paste-ready n8n workflow (install n8n-nodes-socialmate from Community Nodes) or an MCP config for Claude Desktop, Cursor, Claude Code or Cline. Whatever orchestrator you use, put the same disclosure sentence in the AI Agent system message — that is where a tool-calling agent actually reads its instructions.
4. Switch on the local API and create a key The local API server is off by default, which is the most common reason tools do not appear. Enable it under API & Integrations, then create an API key. Text send and read endpoints run on Free; Pro adds media, groups, scheduled messages, the named tunnel and Agent Memory.
5. Test on your own number, then warm before any batch Send to yourself first and read the thread as a customer would, disclosure included. Warm a new number over 72 hours before sending to anyone else, and keep batches personalised — the duplicate-content guard blocks identical text to many contacts by design. Bans are always possible; SocialMate reduces risk rather than eliminating it.

Frequently asked questions

Does WhatsApp automatically label AI-generated messages?

Only for Meta's own AI. WhatsApp's Help Center documents an “AI from Meta” notice inside chats, scoped to selected businesses, limited countries, a waitlist and partial language coverage. A business running its own model gets no such marking. Whatever your customer reads is text you placed.

If I self-host the agent, isn't the disclosure the software vendor's problem?

No. The tool supplies a mechanism, not the obligation. Article 50 places the duty on the provider or deployer of the AI system — the party who chose to run that agent against that audience. Self-hosting relocates the duty onto your machine and your prompt.

Is 2 August 2026 still a future date I can plan around?

No. Article 50 has applied since 2 August 2026 and is in force now for systems that interact with people. If your agent talks to EU customers today, the transparency duty applies today, not at some later transition.

Do I need a lawyer to write one sentence?

No. Write the sentence, ship it in the first message, and keep a dated copy of a real thread showing it. Counsel matters when you are near a high-risk use case, drafting for multiple jurisdictions, or selling an AI assistant as the product — not for choosing the wording of a plain disclosure.

What is the exposure for breaching the transparency obligations?

Breaching them is enforceable under Article 99 of Regulation (EU) 2024/1689, which sets maximum fines and includes provisions capping a fine for an SME or start-up. Read Article 99 and the Commission's guidance for the figures that apply to your case; scope turns on facts.

Does an AI label make customers trust my bot less?

Trust, not reach, is the exposure. An undisclosed bot that gets caught costs more than a disclosed one that sets expectations. Disclosing also gives you the handoff line — reply HUMAN — which converts a dead end into a support ticket.

Will the local API and n8n node work on Free, or is automation a Pro feature?

The local HTTP API runs on Free: read endpoints, plain text send, quoted replies, reactions, mark-as-read and the typing indicator all work there, plus 9 of 35 webhook events. Pro adds media, polls, locations, contact cards, group operations, the smart queue, scheduled messages, unlimited webhook endpoints and deliveries, the named tunnel and Agent Memory.

Do I need WhatsApp Business approval to do any of this?

Not with a self-hosted app running on your own number — there is no Meta approval step. You also get none of the Cloud API's official status, and automated use of the unofficial protocol can violate WhatsApp's terms. Bans are always possible; warming and pacing reduce the risk and never remove it.