The EU AI Act's Transparency Rules Land August 2. What US Businesses Must Do

Starting August 2, 2026, a chatbot operating in the European Union has to tell users it is a chatbot. A tool that generates images, audio, video, or text has to mark that output as machine-made, in a format other machines can detect. Anyone who publishes a deepfake, or an AI-written piece on a matter of public interest, has to disclose it. These are the transparency duties in Article 50 of the EU AI Act, and the deadline does not stop at Europe's borders. A business based in Ohio can be in scope too, depending on where its output ends up.

What Article 50 actually requires

Article 50 of the Act, formally Regulation (EU) 2024/1689, in force since August 1, 2024, sets out several duties rather than one blanket rule. Providers of chatbots and other systems built for direct interaction with people must make the AI's nature obvious, unless it would already be obvious to a reasonably well-informed person. Providers of systems that generate synthetic audio, images, video, or text must mark that output in machine-readable form so it can be flagged as artificial. The European Commission's draft guidance on Article 50, described in a June 8, 2026 client alert from Greenberg Traurig, spells out what will not count. A chatbot calling itself an 'assistant' does not qualify. Neither does a disclosure buried in a terms-of-service page, or a watermark no ordinary user will ever notice. The guidance also reaches agentic systems, tools that act on a person's behalf rather than just chat.

Two more duties fall on deployers: whoever uses the system, not whoever builds it. Publishing a deepfake requires a label, and the definition is broad enough to cover a realistic synthetic image of a person who does not exist. Per Greenberg Traurig's June 8 summary of the guidance, intent to deceive is irrelevant; even an obvious parody needs some disclosure, just a lighter one. Publishing AI-generated or AI-edited text on a matter of public interest, such as public health or an election, requires disclosure too, unless a named person or organization has substantively reviewed the piece and stands behind it editorially. A platform's own AI-labeling feature does not discharge this duty by itself.

Who has to comply, including outside the EU

Article 2 extends the Act's reach past EU borders in two ways. It covers any provider that places an AI system on the EU market or puts it into service there, regardless of where that provider is headquartered. It also covers providers and deployers anywhere if their AI system's output is used in the Union. A US company selling a chatbot subscription to customers in Germany is placing a system on the EU market; that much is not a close call. Cooley's guidance on the Act, published September 9, 2024, gives a similar example: a US developer whose CV-screening tool is used by an EU employer is in scope; open-source models under free licenses, and purely personal use, generally are not.

A US blog or newsletter that uses AI drafting tools and happens to pick up European readers sits in a harder spot. Nothing in Article 2 sets a subscriber count, a traffic percentage, or any comparable threshold that pulls a website into scope. The more defensible reading is that incidental foreign readership is not the same as putting a system into service in the Union. A publisher actively building an EU audience, translating for EU languages, or selling something to EU customers has a weaker claim to staying outside the rule. No regulator has drawn a bright line between the two yet.

Where August 2 fits in the rollout

Article 50 is neither the Act's first deadline nor its last. A set of 'unacceptable risk' practices, including social scoring and untargeted facial-image scraping for recognition databases, was banned starting February 2, 2025, alongside a general AI-literacy requirement, per a January 31, 2025 note from Mayer Brown. General-purpose AI models picked up documentation, copyright, and training-summary duties on August 2, 2025, when the EU's AI Office and AI Board also became operational, per DLA Piper's August 2025 summary. High-risk system rules, covering AI in hiring or credit scoring, were supposed to follow Article 50 on August 2, 2026. Most of that got pushed back. Lawmakers agreed to delay standalone high-risk systems to December 2, 2027 and embedded high-risk systems, in products like medical devices, to August 2, 2028, per Sidley's Data Matters blog on June 22, 2026; the Council of the EU gave final approval June 29, 2026, per Tech Times' July 10, 2026 account. Article 50 itself was not part of that delay, though generative systems already on the market get until December 2, 2026 to meet the Article 50(2) marking duty, per Gibson Dunn's May 27, 2026 client alert.

The penalties, and who is likely to feel them

Article 99 sets three fine tiers. Violating the February 2025 prohibitions tops out at EUR 35 million or 7 percent of worldwide annual turnover. Giving a regulator false or incomplete information tops out at EUR 7.5 million or 1 percent. Article 50 transparency violations sit in the middle: up to EUR 15 million or 3 percent of worldwide annual turnover, whichever is higher, under Article 99(4). Small and medium enterprises, including startups, get the opposite formula: fines capped at whichever amount is lower, under Article 99(6).

Whether a small US operator actually pays is separate from whether the rule applies to them. Enforcement runs through national regulators, and that network is unfinished. At least twelve member states missed the August 2, 2025 deadline to name a competent authority, and seventeen had a thin public enforcement record as of mid-2026, Tech Times reported on July 10, 2026. France had not notified a competent authority to the Commission as of June 2026, and Germany's implementing law was still moving through the Bundestag as of July 2026. A newsletter writer in Ohio is not the likely first target of a fine that needs a cross-border case and a working authority behind it. A US company with real EU subscription revenue and an unlabeled chatbot is a more obvious one.

What to do before August 2

• Disclose any chatbot or AI assistant as AI, clearly and at first contact, not in a footnote or a terms-of-service link
• Mark synthetic audio, image, video, or text output in machine-readable form if your product generates it
• Label deepfakes and AI-generated or AI-edited text on public interest topics before publishing, unless a named editor has substantively reviewed the piece
• Confirm whether your product is sold to, or your output reaches, users in the EU, since that determines whether any of this applies
• Read vendor and platform contracts to see who is responsible for marking AI output; a platform's own label does not automatically cover your disclosure duty

None of this requires a new government filing. It requires an honest label, in place by August 2, 2026, and a second look at that grace period on December 2, 2026.

This is general information, not legal advice. The AI Act's official text and your own counsel are the authoritative sources on how Article 50 applies to your business. Questions go through our contact page.