What every PR and communications team needs to know about the EU AI Act’s transparency rules — and the detail almost nobody is reporting.
There is a line buried in the European Commission’s final guidance, adopted on 20 July, that should stop every communications director in Britain mid-scroll.
For AI-generated text on matters of public interest, the obligation to label attaches at the date of publication — not the date of generation.
Read that again with a content calendar open in front of you.
The press release drafted this week and issued on 4 August is in scope. The thought-leadership piece your AI tool helped structure in July, scheduled for the second week of August, is in scope. The public-affairs briefing sitting in a queue, the executive statement pre-written for an event, the synthetic spokesperson video cut last month for a launch after the bank holiday — all of it turns on when it goes out, not when it was made.
Most of the coverage of the EU AI Act’s transparency rules has treated 2 August as a future problem. For anyone who publishes on a schedule, it is a present one. The content already sitting in your drafts folder is the content the rules will land on first.
What actually changes on 2 August
Article 50 of the EU AI Act (Regulation (EU) 2024/1689) introduces transparency obligations for AI systems that interact with people or generate content. Those obligations apply from 2 August 2026, with penalties reaching €15 million or 3% of worldwide annual turnover, whichever is higher.
On 20 July, the Commission adopted the final version of its Guidelines on implementing Article 50 — a 51-page document published less than two weeks before the rules take effect. The Guidelines are not binding, and only the Court of Justice can give an authoritative interpretation, but national market surveillance authorities across all 27 member states are expected to treat them as the primary reference.
For communications teams, four duties matter:
Chatbots must identify themselves. Any AI system that interacts directly with a person must be designed so that person knows they are dealing with a machine. If your client’s website has a conversational assistant, it needs to say what it is.
Synthetic content must be machine-readable. Providers of generative AI systems must mark outputs in a machine-readable format, detectable as artificially generated. This is a provenance obligation — it is about detection tools being able to verify origin, not about a visible watermark.
Deepfakes must be labelled. Deployers publishing AI-generated or manipulated audio, image or video content must disclose it, using harmonised EU labels at first exposure.
AI-generated text on matters of public interest must be labelled where it is published to inform the public — unless it has undergone human review with editorial responsibility held by a person or organisation.
That last exemption is the one worth reading carefully, and we will come back to it.
Three details the coverage is getting wrong
It is not all happening on 2 August
Generative AI systems already placed on the EU market before 2 August have a transitional period until 2 December 2026 to bring their machine-readable marking into conformity, under the Digital Omnibus provisional agreement reached in May. Systems placed on the market on or after 2 August comply from that date.
Everything else — chatbot disclosure, deepfake labelling, public-interest text labelling — applies from 2 August without exception.
If someone tells you the whole regime lands at once, they have not read the transitional provisions.
Content made before 2 August is treated differently depending on what it is
There is no retroactive labelling obligation. But the cut-off works differently by content type, and this is where the publication-date point bites.
For image, audio and video deepfakes, the relevant date is the date of generation. Material generated before 2 August does not need retroactive marking — a change from the May draft, which had used a different test.
For text on matters of public interest, the relevant date is the date of publication. Text generated before 2 August but published on or after that date must be labelled.
For anyone running an editorial calendar, that asymmetry is the whole story. Your video assets are judged on when you made them. Your written content is judged on when you press publish.
The initial signatory window is closing fast — and missing it is fine
There is a rush of urgency building around signing the Code of Practice on Transparency of AI-Generated Content. The Commission’s guidance sets a deadline of 27 July 2026 at 18:00 CEST for inclusion in the initial signatory list, submitted by a senior executive with authority to bind the organisation. Some legal commentators have reported the deadline as 22 July.
If your organisation is not going to make that list, nothing is lost. Missing the initial-list deadline does not prevent an organisation from joining the Code later. The Code is voluntary in any case — the underlying Article 50 obligations are not. The final Guidelines confirm the Code has been assessed as adequate by the Commission, meaning signatories can rely on adherence to demonstrate compliance rather than proving it another way.
Signing is a route to compliance. It is not compliance itself, and not signing is not non-compliance.
Why this is not just a legal problem
Three other things happened in the same fortnight, and together they change the working assumptions of the PR industry more than the AI Act does on its own.
Newsrooms have written their rules
On 16 July, the Columbia Journalism Review published detailed guidance on developing newsroom AI policies, drawing on practice at the Wall Street Journal and others. The framing from Tess Jeffers, the WSJ’s head of newsroom AI and data, is telling: their guidelines open with the opportunity, and the first prohibition does not appear until five paragraphs in.
But the prohibitions are there. The recurring principles across newsrooms are human oversight, accountability, transparency, enterprise-grade tools and continuing review — and CJR’s framing is blunt about the stakes: a newsroom that lets AI-generated mistakes slip into production risks damaging its reputation.
Journalists who have just spent months writing rules for their own use of AI will apply the same scepticism to what lands in their inbox. Mass-produced “personalised” pitches, unverifiable quotes, and synthetic assets sent without disclosure are now a reputational risk to the sender, not just an irritation to the recipient.
The rule worth adopting, before anyone asks you for it:
AI may research, organise, compare and draft. A named human remains responsible for every claim, quotation, source and pitch that leaves the building.
Keep the source evidence attached internally. Disclose material synthetic assets before a journalist has to ask. The agencies that do this will find it becomes a selling point faster than they expect.
Agents can be turned against their owners
On 15 July, OpenAI published details of GPT-Red, an internal automated red-teaming model built to find prompt-injection vulnerabilities in its own systems. The company reported it succeeded in 84% of internal evaluation scenarios, against 13% for human red teamers on the same tests.
The finding that should concern anyone running client-facing automation is not the benchmark. It is the live test. OpenAI aimed GPT-Red at an AI-run vending machine in its office, built by Andon Labs. After practising in simulation, it hit the live agent and met all three goals: it cut a stocked item’s price to the $0.50 floor, listed a new item at that same price, and cancelled another customer’s order.
A chatbot that answers questions is one level of risk. An agent connected to email, a CRM, a publishing tool or a payment system is a different proposition entirely. If it can send, publish or transact, it can be persuaded to.
Youth-facing brands need a separate standard
OpenAI’s teen-safety framework, detailed on 16 July, sets out age prediction, stronger default restrictions, parental controls and break reminders — explicitly framing AI access for teenagers as requiring protections designed for their developmental stage.
For education, entertainment, fashion, gaming and creator-economy clients, the implication is direct: you cannot launch a general-audience chatbot and assume the underlying model handles every safeguarding question. Conversations that stray into emotional, body-image or mental-health territory escalate into brand crises faster than any other category of AI failure.
And the ownership question is hardening
On 15 July, Australia established an Office of AI within the Department of the Prime Minister and Cabinet, with national standards going to National Cabinet in August 2026 and legislation expected in early 2027. Prime Minister Anthony Albanese’s position was unusually direct — that Australian works should not be used to train AI systems without the creator retaining control, including over price and value, and that anything less amounts to theft. The government continues to rule out a text and data mining exemption.
It is not yet an operative licensing regime. But for any AI or technology client, claims implying unrestricted entitlement to scrape or train on publisher content are now a positioning risk. Licensed data, documented permissions and creator partnerships are becoming commercial trust assets.
What to do this week
Audit the publishing queue first. Every piece of AI-assisted text scheduled to publish on or after 2 August that could count as informing the public on a matter of public interest. That is the immediate exposure, and it is sitting in your CMS right now.
Establish where human editorial responsibility sits. The labelling obligation for public-interest text does not apply where content has undergone human review and a person or organisation holds editorial responsibility. Most agencies do this already — very few document it. Documentation is what turns a practice into a defence.
Separate provider duties from deployer duties. If you use a third-party tool, the machine-readable marking obligation generally sits with the provider. Your obligations as a deployer are about labelling deepfakes and public-interest text. Knowing which side of that line you are on determines what you actually have to build.
Check what your agents can do, not what they are supposed to do. If a client-facing agent has permission to send, publish, refund or delete, the question is not whether it behaves correctly under normal use. It is what happens when someone deliberately tries to manipulate it — through a hostile webpage, an uploaded document, or a message crafted to look like an instruction.
Write the disclosure position before you need it. When a journalist asks whether an asset was AI-generated, the answer should already exist in writing, agreed with the client, with a named person able to give it.
The uncomfortable part
Most of the conversation about AI compliance in our industry has the same shape: a cost to be absorbed, a burden arriving from Brussels, another thing to explain to a client who will not want to pay for it.
That framing is about to age badly.
The businesses asking their agencies about AI controls are not doing it because they read the AI Act. They are doing it because their own customers, insurers and procurement teams have started asking them. The question moves down the supply chain, and it arrives at the agency in the form of a questionnaire nobody can complete.
The agencies that can answer clearly — with a documented position on disclosure, human accountability, source evidence and agent permissions — will keep work that others lose in procurement. That is not a compliance story. That is a commercial one.
How we can help
3RD4PR has built AI Human Proof™ — an assurance framework that examines the human controls behind AI systems, agents and automated content. It covers permission architecture, human approval gates, logging and kill switches, adversarial testing, data boundaries, disclosure position and incident response.
For communications teams specifically, that means: a documented disclosure position, an audit of what your client-facing agents can actually do, a human-accountable media-relations policy, and an incident playbook written before you need it rather than at eleven o’clock on a Friday night.
If AI is already touching your client work, the audit is the fastest way to establish where you stand. Fixed fee, two to three weeks, no commitment to what follows.
See how AI Human Proof works →
Sources
- European Commission — Guidelines on transparency obligations under Article 50
- European Commission — Signing the Code of Practice on Transparency of AI-generated Content
- Bird & Bird — First impressions of the final Article 50 Guidelines
- OpenAI — GPT-Red: Unlocking Self-Improvement for Robustness
- Columbia Journalism Review — How to Develop AI Guidelines
- Australian Government — Office of AI
- White & Case — Australian AI Update: Australia changes course
This article is general information, not legal advice. Article 50’s application depends on the specific system, content type, audience and role — provider or deployer. Organisations should take qualified legal advice on their own position.
3RD4PR is a communications agency established in 2004.