What Actually Changes After the EU AI Act: A Builder's Guide to Life After August 2

On August 2, 2026, the EU AI Act stopped being a document people wrote explainers about and started being something you can observe in products. The clearest evidence arrived nine days later, when Anthropic began embedding watermarks in all Claude text output — worldwide, at the model level, with a cutoff set to exactly that date.

That is the shift worth understanding. Our complete guide to EU AI Act and US policy covers what the law says across jurisdictions. This post covers what changes now that it binds: which obligations just landed, which are still coming, and — the part almost everyone gets wrong — which of them are yours rather than your model provider's.

TL;DR — what changed and what's next

The staged timeline, and why August 2 mattered twice

The Act did not switch on at once. It entered into force on August 1, 2024, and its obligations phase in over roughly three years — which is why "is the AI Act in effect?" has had a different correct answer every few months.

August 2, 2026 is the heaviest of those dates, and it carried a second, less obvious significance. The Article 50(2) Code of Practice on Transparency of AI-Generated Content — a voluntary instrument signatories use to demonstrate compliance — applies to models launched from that date. That is why Anthropic's help centre draws its line at "models launched on or after August 2, 2026" rather than at any engineering milestone. A compliance boundary got translated directly into a product changelog.

Understanding that mapping is genuinely useful, because it lets you predict the next round. When a lab announces a behaviour change with an oddly specific date attached, check whether that date appears in a regulatory calendar first.

What already changed, in things you can observe

Regulation usually feels abstract until it shows up in a tool you use. Four changes are already visible:

1. Generated text carries a machine-readable mark. Anthropic marks all output from qualifying Claude models, and an Anthropic engineer confirmed publicly that "other labs are adding similar watermarking." Google has run SynthID for years. The mechanism is a keyed statistical bias in token selection, applied below the model in the sampling layer, so no product surface can opt out of it.

2. Generated files carry signed provenance. C2PA manifests on supported image types, the same standard behind LinkedIn's Content Credentials rollout.

3. Marking is worldwide, not regional. This is the detail that surprised people most. Nothing in an EU code of practice requires marking text generated for a developer in Bangalore or Boise. Providers did it anyway, because maintaining two inference paths is more expensive than applying one behaviour globally. Regulatory scope and product scope are different things, and product scope wins.

4. Detection is becoming a product. Anthropic has announced a text detection API customers can call themselves, because the Article 50(2) code asks signatories to support third-party detection. We break down what that costs and who pays in are AI watermarks monetisable?.

The part most builders get wrong: provider vs deployer

Here is the single most consequential distinction in the whole Act for anyone building on top of a model, and it is routinely missed.

The Act assigns obligations by role, not by company size or by who wrote the model:

Almost every builder reading this is simultaneously a deployer of someone else's model and a provider of their own AI system. Both sets of obligations apply, and they are not the same obligations.

Article 50 makes the split concrete:

• Article 50(1) — providers must ensure people are informed they are interacting with an AI system, unless it is obvious.
• Article 50(2) — providers of systems generating synthetic content must mark output in a machine-readable, detectable way. This is the watermarking clause. It is the provider's duty.
• Article 50(3) — deployers of emotion-recognition or biometric-categorisation systems must inform the people exposed to them.
• Article 50(4) — deployers who generate or manipulate deepfakes must disclose that. Deployers who publish AI-generated text to inform the public on matters of public interest must disclose that too, unless a human reviewed it and someone holds editorial responsibility.

Read 50(4) again if you publish content. A model provider's watermark does not discharge your disclosure duty, because they are different obligations on different parties. Anthropic states this explicitly: if you deploy Claude in your product, you must independently assess what Article 50 requires of you.

That editorial-responsibility carve-out in 50(4) is the most practically important sentence in the article for anyone running a publication, and it rewards exactly the workflow that was already good practice — a human reviewing and owning the output.

Whether it reaches you

Two separate mechanisms, and people conflate them.

Legal reach. The Act applies to providers placing AI systems on the EU market regardless of establishment, and to providers and deployers in third countries where the system's output is used in the EU. That output clause is broad. A US company with EU users is generally in scope.

Practical reach. Even where the law does not reach you, the product does. You cannot buy an unmarked tier of Claude by being outside Europe, because the marking is applied at model level worldwide. As we argue in will every AI model watermark its output?, the only reliable escape is running open weights where you control the sampler — and that is an architectural choice, not a jurisdictional one.

This is the Brussels effect operating exactly as designed: one jurisdiction's rule becomes a global default because compliance is cheaper to apply universally than to segment.

Penalties, and why the tiers matter more than the headline

Coverage tends to quote the biggest number. The tiers are more informative than the maximum:

Whichever is higher applies, with proportionally lower caps for SMEs and startups. Two things follow.

First, transparency violations sit in the middle tier, not the top one. The Act reserves its heaviest penalties for prohibited practices — social scoring, certain biometric surveillance, manipulation exploiting vulnerabilities. Failing to disclose AI-generated content is serious but not catastrophic-tier.

Second, the third tier is a trap worth noticing. Giving an authority incorrect or misleading information carries its own penalty, separate from the underlying violation. In practice that means a sloppy or defensive response to an inquiry can create exposure independent of whatever prompted the inquiry. Accurate records are the cheap mitigation, which is the same conclusion the developer-side analysis reaches from a different direction.

The open-weight carve-out is narrower than advertised

"Just use open weights" is a real answer to model-level marking, but it is not a general exemption from the Act, and the distinction gets blurred constantly.

What the exemption does: free and open-source GPAI models released under a licence allowing access, use, modification, and redistribution — with weights and architecture publicly available — get relief from certain documentation and information obligations placed on GPAI providers.

What it does not do:

• It does not apply to models with systemic risk. Above the capability threshold, obligations attach regardless of licence.
• It does not exempt deployers. Your Article 50(4) disclosure duty does not disappear because your model is open.
• It does not touch the high-risk regime. If your system falls in Annex III, the licence of the underlying model is irrelevant.
• It does not cover monetised models. Commercial terms can pull a nominally open release back into scope.

So the accurate framing is: open weights remove the provider's marking behaviour from your output. They do not remove your obligations. That distinction matters for anyone reading our open-weight versus closed model guide as a compliance strategy rather than an engineering one.

What to actually do

Ordered by cost-to-benefit, not by legal thoroughness:

1. Determine your roles, in writing. For each AI feature you ship, name whether you are provider, deployer, or both. Everything else follows from this and it takes an afternoon.
2. Check Annex III against your use cases. Employment screening, education access, credit scoring, essential services, biometrics — high-risk categories are specific, and most products are not in them. Confirm rather than assume in either direction.
3. Write your Article 50(4) position down. If you publish AI-assisted content on matters of public interest, decide now whether you rely on the human-editorial-responsibility carve-out, and make that real — a named human who reviews and owns it.
4. Add a disclosure clause to client and vendor contracts. Agencies especially: state what AI does, and state that a provenance mark means content was processed, not authored. The marketer-side breakdown covers the specific language.
5. Keep provenance records for your own work. Drafts, prompts, revision history. This is what answers an inquiry, and it also defuses the false-accusation problem that comes with detectors.
6. Do not treat a detection result as proof. Write "detected means processed, not authored" into any policy before your first dispute, not after. How to detect a Claude watermark covers what a result actually supports.
7. Watch for dated announcements. When a provider ships a behaviour change on a suspiciously specific date, check the regulatory calendar. It usually explains the timing better than the blog post does.

What is still unresolved

An honest guide has to mark the edges.

Enforcement is young and uneven. National market surveillance authorities are at very different stages of readiness, and the first significant enforcement actions will tell us far more about real-world scope than the text does.

The Code of Practice is voluntary. Signing it is a way to demonstrate compliance, not the only way. A provider can meet Article 50 differently, which means "all labs will converge on identical watermarking" is a prediction, not a requirement.

Detection obligations lack specifics. The code asks signatories to support third-party detection. Whether that means a free public API, a gated endpoint, or a request process is unsettled — and as of this writing Anthropic's announced detection API has no published pricing, threshold, or false-positive rate.

The interaction with other regimes is untested. California's transparency legislation imposes parallel duties, and reconciling overlapping disclosure obligations across jurisdictions is a problem nobody has had to solve at scale yet.

Bottom line

The EU AI Act's most important effect so far is not a fine or a banned product. It is that a regulatory deadline now shows up as changed behaviour in tools used by people who will never read the regulation — worldwide, because global uniformity is cheaper than regional segmentation.

For builders, the actionable core compresses to one sentence: marking is the provider's job, disclosure is often yours, and nobody's watermark discharges your obligation. Everything else is detail you can look up when it becomes relevant. That one is worth internalising now, because it is the assumption most teams are currently getting wrong.

Related on explainx.ai

• EU AI Act and US policy: complete regulation guide — the full statutory picture across jurisdictions
• Anthropic is watermarking Claude text — the first widely visible consequence of Article 50(2)
• How AI text watermarking actually works — the mechanism behind the compliance behaviour
• Will every AI model watermark its output? — where the convergence holds and where it breaks
• The case FOR AI watermarks — including why marking beats the alternative regulatory mechanisms
• What it changes for developers · for marketers — role-specific consequences
• How to detect a Claude watermark — what a detection result does and does not support
• Europe's AI landscape: sovereign compute and the EU Act — the politics driving enforcement appetite
• Open-weight vs closed models: how to choose — read as engineering strategy, not a compliance exemption
• LinkedIn's C2PA Content Credentials — provenance deployed at platform scale

Primary sources: Regulation (EU) 2024/1689 (the AI Act), Articles 5, 50, 99 and Annex III · EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content · Anthropic Help Center, "How Claude marks AI-generated content" (August 11, 2026)

Accurate as of August 12, 2026. This is a practitioner's orientation, not legal advice — obligations depend on your specific role, system, and jurisdiction, and national enforcement practice is still developing. Penalty figures reflect the Act's stated maximums; SME and startup caps are proportionally lower. Follow @explainx_ai for updates.