Regulation
The EU AI Act clock: what actually changes on 2 August 2026
On 2 August 2026, two sets of EU AI Act duties become enforceable: the European Commission's general-purpose AI (GPAI) enforcement powers and the Article 50 transparency obligations. So yes - if you serve EU users, from that date you must disclose when someone is interacting with a chatbot rather than a person, and mark generative AI output in machine-readable form, with deepfakes labelled. These are not the obligations the June "delay" headlines were about, and the penalty behind them is real.
Dated analysis. Checked 18 July 2026, fifteen days before the deadline. No first enforcement action or fine has occurred yet, so nothing below reports one; the figures are the statutory maxima, not observed penalties. Compliance specifics for any given product should be taken from the regulation and qualified counsel, not a news post.
What actually changes on 2 August 2026?
Two things activate on the same date. First, the Commission's active GPAI enforcement powers - the ability to issue information requests, demand model access, and compel recall - come into force one year after GPAI's substantive obligations took effect on 2 August 2025. Second, the Article 50 transparency duties begin applying to anyone deploying AI that interacts with, or generates content for, EU users. The first is about the Commission's teeth against foundation-model providers; the second is about what ordinary deployers must tell their users.
Wasn't the AI Act delayed?
Part of it was - but not this part, and that is the distinction the headlines blurred. On 16 June 2026 the European Parliament approved amendments pushing most high-risk AI system obligations out to December 2027 and August 2028. That delay explicitly did not touch Article 50 or the GPAI enforcement powers, which remain fixed at 2 August 2026. If your reading of the June coverage was "the AI Act got pushed back, I have time," that applies to the high-risk regime, not to the transparency and GPAI duties landing next month.
| Obligation | Applies from |
|---|---|
| Article 50 transparency duties | 2 Aug 2026 (unchanged) |
| GPAI enforcement powers | 2 Aug 2026 (unchanged) |
| Most high-risk system obligations | Dec 2027 / Aug 2028 (delayed) |
Do you need to label AI-generated content?
For EU-facing services, in most cases yes. Article 50 sets out four duties: (1) chatbot disclosure - users must be told they are interacting with AI, and the disclosure must be perceivable within the interaction itself, not buried in the terms of service; (2) machine-readable marking of generative AI output, covering text, image, audio and video; (3) deepfake labelling for synthetic media depicting real people or events; and (4) an editorial exemption for AI-assisted content that passes documented human review under a named responsible editor. That last carve-out is narrow and conditional - it rewards an auditable review process, not a claim of one - which is exactly the kind of documented-human-in-the-loop discipline worth having in place before the date rather than after a complaint.
What is the penalty?
Both GPAI and Article 50 violations sit under the same Article 99 maximum tier: the greater of EUR 15 million or 3% of total worldwide annual turnover (EU institutions are capped separately at EUR 750,000). For a large multinational the turnover-percentage limb is the one that bites, and it is calculated on global, not EU, revenue. This is a materially larger exposure than the transparency rules are often assumed to carry, which is why they deserve engineering and editorial attention now rather than treatment as a cosmetic disclosure line.
Who enforces it?
At EU level the primary GPAI enforcement body is the EU AI Office, which sits within DG CONNECT and has been operational since February 2024. Article 50's transparency duties are enforced through the wider AI Act supervisory structure across member states. No enforcement action or fine has been recorded yet - the powers are simply becoming available on 2 August - so any specific "first target" claim you see before then is speculation.
Questions people ask
What EU AI Act rules take effect in August 2026? The Commission's GPAI enforcement powers (information requests, model-access demands, recall) and the Article 50 transparency obligations, both on 2 August 2026.
Do I need to label AI-generated content for EU users? Under Article 50, generative AI output must be marked in machine-readable form, and deepfakes labelled; there is a documented-human-review editorial exemption for AI-assisted content, but it is conditional on an auditable review process.
What is the penalty for not disclosing a chatbot is AI? Up to the greater of EUR 15 million or 3% of total worldwide annual turnover, under Article 99.
Has the EU AI Act high-risk deadline been delayed? Yes - the Parliament's 16 June 2026 amendments pushed most high-risk system obligations to December 2027 and August 2028. That delay did not affect Article 50 or GPAI enforcement, which stay on 2 August 2026.
Does the EU AI Act apply to US companies serving EU customers? The Act's transparency duties attach to AI systems interacting with or producing content for people in the EU, so a non-EU provider serving EU users is within scope; the turnover-based penalty is calculated on global revenue.
Sources & caveats. Transparency-obligation detail from a fetched Article 50 checklist, cross-checked against EU AI Act timeline trackers (Legiscope, DLA Piper, DataGuard) for the 2 August 2026 activation and the 16 June 2026 high-risk delay. Article numbers (50, 99) and penalty tiers are drawn from those sources; no first enforcement action had occurred at the time of writing, so no penalty figure here is an observed fine. This is reporting, not legal advice.