On July 31, 2026, the European Commission published a press release announcing that from August 2, 2026, the Commission's AI Office, together with national authorities across EU member states, will formally begin enforcing the EU AI Act. The same day, a new set of transparency rules also takes effect: chatbots and other interactive AI systems must tell users they're talking to AI, not a human; deepfakes โ images, video, or audio edited or generated by AI โ must be labeled; and AI-generated or AI-altered content must carry machine-readable marks so it can be detected more easily. Even if your company's headquarters isn't in the EU, if your AI product or service reaches EU users, these rules may still apply to you โ this piece walks through exactly what's happening on August 2, and what businesses should practically check.
Two Things Taking Effect on August 2: Full Enforcement, Plus New Transparency Rules
The official press release is direct about it: from August 2, 2026, the Commission's AI Office, together with member state authorities, begins enforcing the AI Act. That means the Act is no longer just "in force but still in a watch-and-see period" โ it formally enters a phase where providers can be required to submit documentation, have their models directly assessed, be ordered to make corrections, or even have a model restricted or withdrawn. Landing alongside enforcement is a new set of transparency rules, aimed at three situations that come up constantly in practice: chatbots and other interactive AI systems, deepfakes, and AI-generated or AI-altered content.
What the New Transparency Rules Actually Require
Per the official press release, the new rules boil down to three concrete requirements. First, chatbots and other interactive AI systems must tell users they're interacting with AI, not a human. Second, deepfakes โ content edited or generated with AI, including images, video, and audio โ must be labeled. Third, AI-generated or AI-altered content must carry machine-readable marks, making it easier to detect. The Commission states the goal is to reduce deception and manipulation and help people make informed choices, while also giving businesses clearer obligations and a practical way to demonstrate compliance. Worth noting: the Commission has already published a first list of organizations that have signed the Code of Practice on transparency of AI-generated content โ more than 180 organizations so far โ a voluntary tool designed specifically to operationalize these transparency rules.
Why This Matters Even Without an EU Presence
The AI Act's scope is designed to follow the user, not just the provider's registered address: if your AI system is placed on the EU market, or its output ends up being used within the EU, the relevant obligations can apply regardless of whether your company is headquartered in Taiwan, the US, or anywhere else. For a business without EU operations, the two most common exposure points are: first, an AI product you develop or resell that serves EU customers or users; and second, integrating a third-party general-purpose AI model (GPAI) into your own product, where your downstream product in turn serves EU users. Both scenarios are worth checking now, rather than discovering the gap only after being asked to produce documentation.
GPAI Provider Obligations vs. General Business Obligations
Per the Commission's own factpage, the AI Act sets obligations for providers of general-purpose AI (GPAI) models, split across two tiers. All GPAI model providers must: draw up technical documentation, implement a copyright policy, and publish a summary of the model's training content. GPAI models presumed to carry "systemic risk" (trained with more than 10^25 FLOP of compute โ a threshold the Commission notes is currently under review) carry additional obligations: notifying the Commission, conducting risk assessment and mitigation, incident reporting, and implementing cybersecurity protections. The Commission also defines the GPAI threshold itself: any model trained with more than 10^23 FLOP that can generate language qualifies as a GPAI model. These obligations have technically been in force since August 2, 2025 โ a full year before this month's "full enforcement" milestone. What's different now is that enforcement authorities have formal tools to act on it: requiring providers to submit documentation, directly assessing models, and ordering corrections or restrictions.
The Full Timeline: How the AI Act Rolled Out, Step by Step
Laying out the full timeline makes it easier to understand "why now": the AI Act itself entered into force on August 1, 2024; prohibited AI practices and AI literacy obligations became applicable from February 2, 2025; the governance rules and obligations for GPAI models became applicable from August 2, 2025; and this month's August 2, 2026 milestone is when full enforcement, high-risk AI system rules, and these new transparency rules all land together. There's also a newer variable worth flagging: the EU's separate "AI Omnibus" amendment, intended to make the regulation more innovation-friendly, was adopted on November 19, 2025, reached political agreement on May 7, 2026, and entered into force on July 27, 2026 โ just days before this enforcement milestone โ with one specific prohibition under it taking effect in December 2026. In other words, this regulatory framework is still being actively adjusted; it didn't lock into a single final form and stop changing.
What Businesses Adopting AI Tools Should Actually Do
Pulling the points above into a few things you can check right now: First, audit whether any part of your product or service touches EU users, or whether its output could end up being used within the EU. Second, if your product integrates a third-party GPAI model, it's worth proactively asking that model provider whether they've signed the Code of Practice on transparency of AI-generated content, or whether they can supply technical documentation and a training-content summary โ these could form part of your own compliance evidence. Third, audit your own product's interface: if you have a chatbot or other interactive AI feature, does it clearly disclose that users are interacting with AI; if your product generates or alters images, video, or audio, do you already have a corresponding labeling mechanism. Fourth, treat this as an ongoing tracking item rather than a one-time check โ August 2 is just one of several effective dates, with a new prohibition landing in December and the AI Omnibus amendment still being actively adjusted.
FAQ
If my business only serves my home market, do I need to care about the EU AI Act?
If your product or service has no EU users at all and isn't placed on the EU market, in principle you're not directly subject to the AI Act. But if any of your clients or users are based in the EU, or your product's output could end up being used within the EU, it's worth checking the scope at least once rather than assuming you're automatically exempt.
Was the AI Act not in force at all before August 2?
No โ that's not accurate. The governance rules and obligations for GPAI models have technically been in force since August 2, 2025; prohibited AI practices and AI literacy obligations go back even further, to February 2, 2025. What's different about August 2, 2026 is that formal, full enforcement begins, alongside the new transparency rules โ it isn't the first time any part of the Act has taken effect.
What happens if a business isn't compliant?
The official pages state that the Commission's AI Office and member state authorities will be able to issue fines for non-compliance, but the exact fine amounts are not stated on the official pages this piece verified against. We're deliberately not citing an unverified specific figure here โ for the current penalty framework, consult the EU's official legislative text directly or a compliance advisor.
What counts as a GPAI model? Do mainstream large language models qualify?
Per the official definition, any model trained with more than 10^23 FLOP of compute that can generate language qualifies as a GPAI model โ most mainstream large language models on the market today meet this threshold. If a model is trained with more than 10^25 FLOP, it's additionally presumed to carry "systemic risk," triggering extra provider obligations (a threshold the Commission notes is currently under review).
Is the Code of Practice on transparency of AI-generated content mandatory?
No. Per the official pages, it's a voluntary tool developed by independent experts and confirmed by the Commission and member states as "adequate" for demonstrating compliance with AI Act obligations โ but it isn't the only path to compliance. More than 180 organizations have signed on so far.
Source
Every specific fact and date in this piece was verified directly against three official European Commission pages: the "General-purpose AI obligations under the AI Act" factpage (digital-strategy.ec.europa.eu/en/factpages/general-purpose-ai-obligations-under-ai-act), the "Regulatory framework for AI" policy overview page (digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai), and the official July 31, 2026 press release, "Commission starts enforcing AI Act rules and new transparency requirements on 2 August" โ all retrieved and checked against their full raw HTML directly, not summarized secondhand. On the penalty mechanism referenced in this piece: the official pages state only that the Commission and member state authorities can "issue fines for non-compliance," without specifying an amount. This piece does not cite a specific fine figure, because it could not be independently verified against the EU's primary legislative text (EUR-Lex) within this research session โ the EUR-Lex page returned no retrievable content at the time of verification โ so the amount is left out rather than substituted with an unverified number.
