EU AI Act Article 50: What Became Law on August 2, and What It Doesn't Cover

Since August 2, 2026, Article 50 of the EU AI Act has been binding law: AI systems that generate synthetic text, audio, image, or video must mark that output in a machine-readable format, detectable as artificially generated. The obligation applies globally to any provider whose AI output reaches users in the EU — not just companies headquartered there.

What Actually Changed on August 2

The European Commission published its final guidelines on Article 50 on July 20, 2026, and confirmed a voluntary Code of Practice — built around technical standards like C2PA Content Credentials — as an accepted path to demonstrating compliance. Systems already on the market before August 2 get a grace period on the specific marking requirement until December 2, 2026, but every other transparency obligation in Article 50 — disclosing that a user is talking to an AI, labeling deepfakes, flagging AI-generated public-interest text — applied starting on the original date with no delay.

The Exemption That Matters for Everyday Editing Tools

Article 50(2) specifically does not apply where an AI system performs "an assistive function for standard editing" that doesn't substantially alter the input's meaning — the Commission's own guidance names grammar correction as the clear example. A spell-checker fixing typos, or a grammar tool cleaning up punctuation, isn't required to watermark the result; the obligation is aimed at content a system generates or substantially rewrites, not text a person wrote that gets lightly corrected.

The distinction the law draws isn't "did AI touch this text at all" — it's whether the AI substantially changed what the text says. A comma fixed by a grammar checker and a paragraph rewritten by a generative model sit on opposite sides of that line, even though both technically involve AI processing the same sentence.

Where This Connects to Recent AI Provider Announcements

Several major AI companies rolled out invisible watermarking for generated text in the weeks surrounding this deadline — a pattern of timing that lines up directly with Article 50's effective date rather than being coincidental. For anyone who noticed a provider's terms suddenly mentioning watermarks or content marking around early August 2026, this regulation is very likely the reason, whether or not the announcement explicitly named the EU AI Act as the driver.

What This Doesn't Require

The obligation doesn't require retroactively labeling anything generated before August 2, 2026, doesn't apply to purely internal or non-EU-facing use, and carries a specific carve-out for AI systems authorized by law for detecting or investigating criminal offenses. It's also worth being precise about scope: this is a marking and disclosure requirement, not a ban — the law doesn't prohibit AI-generated text, it requires that it be identifiable as such through machine-readable means.

What a Tool Built Around Local, Manual Editing Means Here

A text editor that only counts words, compares two drafts, or converts a file format — without generating or substantially rewriting content — sits outside Article 50(2) by design, since nothing about that kind of tool produces synthetic text in the first place. That's a structural distinction worth understanding regardless of which specific tools someone uses: the question Article 50 asks is whether a system generated or substantially altered content, not whether any software touched the text at all.

Staying Informed as Guidance Develops

The Commission's guidelines are non-binding interpretive guidance rather than the final word — courts will ultimately settle disputed edge cases, and the voluntary Code of Practice is still gathering signatories. Checking a provider's own disclosures, rather than assuming a category of tool is automatically in or out of scope, remains the more reliable approach given how recently these specific rules took effect.

---

Article 50 became binding EU law on August 2, 2026, requiring AI-generated text, audio, image, and video to carry machine-readable marking — with a specific, deliberate exemption for assistive editing that doesn't substantially change what was written. Understanding that distinction, rather than treating every AI-touched piece of text as equally in scope, is what separates an accurate read of the regulation from a broader, less precise anxiety about it.