AI Act: transparency obligations affecting your business since August 2

Since August 2, 2026, software talking to your customers must tell them it is artificial intelligence, and its output must be marked. The regulation's heavier provisions were postponed sixteen months. Here is what applies today, with dates and sources.

AI Act transparency obligations have applied to your business since August 2, 2026, while the central high-risk systems provisions were postponed to December 2, 2027 (Toute l'Europe, article updated August 3, 2026). Four obligations concern anyone having software built that talks or generates content.

The four obligations applying today

According to Toute l'Europe's August 3, 2026 summary, chatbot users must be informed they are addressing artificial intelligence unless obvious. Generated images, video, sound, and text must carry machine-readable technical marking. Content presenting a real person as making statements they never made requires visible disclosure. AI-produced text of public interest must be identified. The article also notes the Commission can now demand information from providers. General fines reach €35 million or 7% of worldwide turnover; transparency failures have a separate scale of €15 million or 3%. These amounts concern established failures and prejudge no particular situation: check yours with your adviser.

The sixteen-month reprieve, and what it excludes

Regulation (EU) 2026/1744, the AI Digital Omnibus, entered into force July 27, 2026 after publication in the EU Official Journal on July 24 (Hunton Andrews Kurth analysis). It shifts the standalone high-risk systems deadline from August 2, 2026 to December 2, 2027, and safety components to August 2, 2028. The Commission cites practical implementation difficulties, delayed national authorities, and missing harmonised standards. Business owners should read this two ways. Heavy compliance documentation concerning recruitment, credit, or education is not required before December 2027. Transparency has been required since August.

The December deadline for existing systems

A second date awaits software marketed before August 2, 2026. The Future of Life Institute's application timetable, consulted September 2, 2026, states that providers of systems generating synthetic audio, images, video, or text marketed before then must comply with Article 50(2) by December 2, 2026. A spring-launched product writing messages or making visuals falls into that situation, detailed in three months to mark what your AI writes.

Turn an obligation into a specification line

A regulatory obligation gets lost when it remains in a legal note and never reaches the product description. Three sentences often suffice: visitors see a disclosure that they are interacting with AI before their first message; every generated file carries technical marking; the screen identifies the model producing the answer. Written into the specification, they become verifiable acceptance criteria. Left in a lawyer's email, they resurface six months later when change is expensive. No tool makes a company compliant, including ours: at best, a tool makes the requirement visible when someone decides.

What Europe fundamentally asks

Article 14 already requires effective human oversight for high-risk systems, a principle we examined in human approval as real security. Transparency obligations share that spirit: a person must know what they face, and someone must answer for the decision. Another related deadline affected digital product security this year, covered in the Cyber Resilience Act. The three texts converge on one request: make what happens inside understandable.

Inventory it in an hour

List places where your software talks to someone or produces content: contact forms with automatic responses, website assistants, generated emails, product descriptions, visuals. For each line, note whether the other person knows they address a machine and whether the output carries a mark. The list fits on a page, starts discussion with your adviser, and converts directly into specification lines for the next version.