Regulation Watch: Article 50 is live - a 90-day plan for AI transparency
Of everything that changed in the AI Act this summer, one item belongs at the top of an AI governance lead's list: Article 50. The high-risk obligations moved to December 2027 and August 2028. The transparency duties did not move. They have applied since 2 August 2026, and the next hard date falls within the coming 90 days.
What applies
Providers must tell people when they are interacting with an AI system, and must mark synthetic audio, image, video and text in a machine-readable way.
Deployers must label deepfakes and certain AI-generated text on matters of public interest, and must disclose emotion recognition and biometric categorisation.
The soft law is complete. The Commission published a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026 and guidelines on Article 50 on 20 July 2026. According to a law firm summary of the guidelines, the information must be given no later than the first interaction or exposure, and generic references in terms and conditions or website footers are unlikely to be sufficient.
Why it comes first
The deadline is close. Generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to meet the Article 50(2) marking duty. Systems placed on the market from 2 August 2026 must comply immediately.
The fines are material. Breaches of transparency obligations can cost up to EUR 15m or 3% of worldwide turnover.
Supervision has started. The Commission's enforcement powers over GPAI providers have applied since 2 August 2026. In Poland the national supervisor, KRiBSI, can inspect and fine from 28 October 2026.
The scope is wide. Any chatbot, voice agent or content generator offered in the EU is covered, whether or not you run a single high-risk system.
Two more dates sit in the same window. From 2 December 2026 two new prohibitions apply to AI systems that generate non-consensual intimate imagery or child sexual abuse material. General-purpose image, video and audio generators are in scope where such output is reasonably foreseeable and reproducible and safeguards are not adequate; the top fine tier applies, up to EUR 35m or 7%. And 9 December 2026 is the transposition deadline for the revised Product Liability Directive, under which software and AI systems become products under strict liability.
A 90-day plan
Days 1-30: find every touchpoint
List every system that talks to people and every system that generates text, image, audio or video, including features embedded in vendor products.
For each, record your role (provider or deployer) and the date it was placed on the market. That date decides whether the grace period to 2 December applies.
Note where emotion recognition or biometric categorisation is used.
Days 31-60: fix disclosure and marking
Put the AI disclosure at the first interaction, in the interface, not in the terms and conditions.
Ask each generative AI vendor in writing how outputs are marked and whether it follows the Code of Practice.
Set a labelling rule for deepfakes and for AI-generated text on matters of public interest, and build it into publishing workflows.
For image, video and audio generators, document the safeguards against prohibited output and the process for correcting misuse.
Days 61-90: prove it
Test disclosure and marking on live systems and keep the evidence.
Add a transparency check to release gates, so that no new feature ships without it.
Review supplier contracts for the liability change.
Report to the board: what was in scope, what was fixed, what remains open.
What not to do
Do not wait for the high-risk classification guidelines. They are due at the end of 2026 and concern a different chapter of the Act. Article 50 applies today.
Sources
This is information, not legal advice.