EU AI Act Article 50: When Communications Teams Must Disclose AI-Generated Content

AI now drafts, trims, and reshapes much of the copy communications teams publish. Starting August 2, 2026, Article 50 of the EU AI Act sets out when that involvement has to be disclosed to the public, and the rule is narrower than the headlines suggest. This piece walks through what Article 50 actually covers, why the exception for human-reviewed public-interest text is the part communications teams should care about, and how to build a review process that can stand behind what it publishes.

KEY TAKEAWAYS

• Article 50 of the EU AI Act does not require organizations to disclose every use of AI.
• It takes effect on August 2, 2026 under Regulation (EU) 2024/1689.
• For AI-assisted text on matters of public interest, no label is required when a person has genuinely reviewed the content and holds editorial responsibility for publishing it.
• The test is substantive human judgment, not a minimum percentage of human-written words.
• Deepfake images, audio, and video, and provider-level content marking, are separate obligations the text exception does not cover.

The question every AI-assisted team now has to answer

AI has become ordinary infrastructure for communications work. It drafts first passes, condenses reports, adjusts tone, and turns raw notes into usable copy. That convenience carries a regulatory question with a date attached: when does an organization have to tell the public that AI was involved?

For teams operating in or reaching the European Union, Article 50 of the EU AI Act, the transparency provision of Regulation (EU) 2024/1689, takes effect on August 2, 2026. It sets transparency requirements for several kinds of AI systems and synthetic content: direct interactions with AI, machine-generated media, deepfakes, and certain text published to inform the public about matters of public interest.

It is not a universal disclosure rule. What it requires depends on what the system does, what kind of content comes out, and whether your organization is acting as a provider or a deployer. For PR and communications teams, one detail carries most of the weight: public-interest text is exempt from disclosure when a human has genuinely reviewed it and someone remains accountable for publishing it. What follows is a practical overview, not individualized legal advice.

What Article 50 covers

Article 50 is the AI Act's main transparency provision for certain AI systems and AI-generated content. Its aim is plain: people should be able to tell when they are dealing with an AI system, or with material a machine generated or altered. The mechanics are more specific: the article sets several distinct obligations, not one.

Direct interactions with AI

Providers of systems that talk directly to people must make sure users know they are dealing with AI, unless that is already obvious. For communications teams, this reaches customer-facing chatbots, virtual spokespersons, and conversational assistants on a website.

Machine-readable marking of synthetic content

Providers of AI systems that generate synthetic text, images, audio, or video have a separate duty to mark their outputs as artificially generated, in a machine-readable format. These are provider-level technical obligations. A communications team's review of a finished asset does not remove duties that sit with the company building or offering the underlying system.

Deepfake images, audio, and video

Deployers face their own disclosure requirement when they use AI to generate or manipulate image, audio, or video content that is a deepfake. The public-interest text exception does not stretch to cover synthetic photographs, cloned voices, or altered video.

Public-interest text

Article 50 also covers AI-generated or manipulated text published to inform the public about matters of public interest: public health, consumer safety, finance, science, the environment, regulation, and policy. Whether a given publication falls inside the provision depends on its purpose and content. This is where human review and editorial responsibility decide the outcome.

Provider or deployer: know which one you are

The distinction between providers and deployers runs through the whole AI Act, and both terms are defined in Article 3 of Regulation (EU) 2024/1689.

A provider develops an AI system and places it on the market or puts it into service under its own name. A company offering a generative AI platform is the provider of that system. A deployer uses an AI system under its own authority in a professional setting. A corporation, PR agency, publisher, or communications department using a third-party generative tool is a deployer.

One organization can be both, in different situations. A company might deploy a vendor's writing assistant while providing its own AI customer-service system. Article 50 hands each role different duties. The human-review exception belongs to the deployer's obligation on public-interest text, and it settles nothing about the other transparency requirements in the article.

The human-review exception

Here is the whole of it. For public-interest text, disclosure is not required when the content has had genuine human review or editorial control and a natural or legal person holds editorial responsibility for publishing it.

That is not an authorship test. The regulation sets no ratio of human sentences to machine ones. AI can draft, rewrite, summarize, restructure, or adapt the copy, and it can do most of the typing. The exception asks whether a person examined the content, held real authority over it, and accepted responsibility for the decision to publish. The machine proposes; the accountable human disposes.

Substantive review means the reviewer does the work: checks names, dates, quotations, statistics, sources, and links against the source material; flags language that is misleading, exaggerated, or unsupported; notices when the framing leaves out context that changes the meaning; rewrites structure, emphasis, and wording where it needs it; escalates sensitive claims to legal, scientific, or executive review; and can delay or reject publication when something does not hold up.

The reviewer needs actual authority: to change the content, challenge its premise, pause the workflow, or decide it should not run. Scanning the first paragraph and clicking approve is not that.

What this looks like in practice

A worked example makes the line concrete. The scenario below is illustrative, not a real case.

A biotech company is announcing results from a mid-stage clinical trial. The communications team feeds the trial summary, the investigator quotes, and the safety data into a generative tool, and it returns a polished release. AI wrote perhaps ninety percent of the prose. It reads beautifully.

One line claims the therapy "significantly outperformed the standard of care." The reviewer has the trial tables open. She notices that the study measured that outcome as a secondary endpoint, and the comparison was not powered to support the word "significantly." The prose is more confident than the data. She cuts the claim, replaces it with what the results actually show, and sends a question back to the medical team before anything goes out.

That single act, catching a confident but unsupported claim and refusing to publish it, is the thing Article 50 is asking for. The volume of AI-written text was beside the point. A person read the substance, had the authority to change it, and took responsibility for what shipped. Had the release gone out untouched, a human name on the byline would not have made the review real.

The mistake most communications teams will make

Most teams will over-correct. Faced with a new rule, they will disclose far more than it requires.

When Article 50 lands, legal and compliance will feel a familiar pull: label everything. Stamp "AI-assisted" on every release, every post, every caption, and the exposure seems to vanish. It is the safe-looking move, and it is the wrong one.

Blanket labeling backfires two ways. Readers who see the label on everything learn to ignore it, so the disclosure that genuinely matters, on the synthetic video or the deepfake, loses its signal. And a label on work your own team reviewed and stands behind tells the reader you are not sure you do. It reads as a shrug. The exception exists so that reviewed, accountable content does not need the label, and using it is a statement of editorial confidence. For the marketing and communications leaders now navigating this shift, the smarter instinct is narrower: label what the law requires, review what you publish, and let the absence of a label mean something.

Building a process you can stand behind

A defensible process rests on three things: clear authority, documented judgment, and a named owner. In practice:

• Name the person or legal entity responsible for publication before drafting starts, not after an error surfaces.
• Review the substance, not just the grammar: the claims, the evidence, the quotations, the context.
• Give reviewers real power to revise, delay, escalate, or reject. A process that rewards speed and punishes scrutiny is not editorial control.
• Return to original documents and approved quotations. AI renders uncertain information in remarkably certain prose.
• Keep a simple record of who reviewed what, what changed, and who approved. Article 50 prescribes no format, but a traceable trail shows the review was substantive rather than nominal.
• Treat each media format on its own. Written copy clearing review says nothing about the synthetic image or video sitting next to it.

Article 50 does not ask communications teams to write every sentence by hand. It asks them to be clear about who answers for the work when AI helps produce it. AI can draft, condense, and rearrange, and it has made fluent copy easy to produce. Someone still has to read it, decide, and put their name to it.

Frequently asked questions

Does Article 50 prohibit AI-generated press releases?

No. It does not ban AI in press releases. It sets transparency requirements for certain AI systems and content, and public-interest text that has had genuine human review is exempt from the disclosure requirement.

Does the text exception cover AI-generated images, audio, or video?

No. Deepfake media has its own disclosure rule and needs a separate assessment. A reviewed press release and an AI-generated video published alongside it are two different questions.

When does Article 50 take effect?

It takes effect on August 2, 2026, under Regulation (EU) 2024/1689. The application dates are set out in Article 113 of the regulation.

Sources

• Regulation (EU) 2024/1689 (the EU AI Act), official text on EUR-Lex, including Article 50 (transparency obligations), Article 3 (definitions of "provider" and "deployer"), and Article 113 (application dates; general application from August 2, 2026): https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689