August 2 Makes Chatbots Introduce Themselves

Brussels just turned the EU AI Act's transparency chapter from a deadline on paper into a twelve-day compliance sprint — and the rules are stricter than most enterprises priced in. On July 20, the European Commission published finalized guidelines for Article 50, which takes effect August 2: every chatbot and AI agent must tell users they're talking to a machine, generative outputs need machine-readable detection marks, and deployers must slap human-visible labels on deepfakes and AI-written public-interest text — with fines up to €15 million or 3% of global turnover.

What's new:

• The Commission adopted final guidelines on Article 50 transparency obligations on July 20, 2026.
• Article 50 applies from August 2 — twelve days from publication.
• Four obligation categories hit providers and deployers simultaneously: interactive AI disclosure, generative output marking, emotion-recognition notice, and deepfake/public-interest text labeling.
• A voluntary Code of Practice on AI-generated content signatory deadline falls July 22 at 6 p.m. CEST.

What providers must build:

• Chatbots, AI agents, and avatars that directly interact with people must inform users from the first exchange — unless interaction with AI is "obvious," a standard the guidelines say must be "interpreted in a restrictive manner."
• Generative systems producing synthetic audio, image, video, or text must embed "effective, reliable, robust and interoperable" machine-readable marks.
• Providers outside the EU are in scope if their system's output is used in the EU.
• Standard editing — spell-checking, color correction, routine post-production — is exempt from marking. Going beyond that is not.

What deployers must disclose:

• Deepfakes require human-perceivable labels at first exposure — visible or audible, no special tools required.
• The Commission is explicit: "deployers cannot simply rely on the machine-readable marking embedded in the content by the provider" to satisfy deepfake disclosure.
• AI-generated text on politics, public health, justice, finance, or other public-interest topics needs clear labeling unless a human with editorial responsibility reviewed it.
• Spell-checking alone does not count as human review. Peer review or editor-in-chief sign-off does.
• Emotion-recognition and biometric-categorization systems require notice to anyone exposed — in real time or after the fact.

The grace period trap:

• Legacy systems placed on the market before August 2 get until December 2, 2026 — but only for Article 50(2) machine-readable marking.
• Chatbot disclosure, deepfake labels, and emotion-recognition notices still bite August 2 with no transition.
• Content generated before August 2 does not need retroactive labeling — but the Commission "encourages" deployers to label it anyway.

The enforcement math:

• National market surveillance authorities enforce Article 50 — not the AI Office, except for GPAI-integrated systems from the same provider or DSA-designated platforms.
• Fines reach €15 million or 3% of worldwide annual turnover, whichever is higher.
• Signatories to the voluntary Code of Practice get a documented compliance pathway; non-signatories face individual scrutiny and "more requests for information."

Why today matters globally:

• Singapore published its own chatbot transparency guidelines the same morning — voluntary "nutrition labels" without mandatory opt-out channels.
• Washington's Gold Eagle clearinghouse decides which companies access frontier models. Brussels decides whether your chatbot admits it's a chatbot.
• Moonshot's Kimi K3 and Alibaba's Qwen3.8 ship open weights with no EU marking layer — but any enterprise deploying them to EU users inherits deployer liability on August 2.
• The Commission press release frames the goal plainly: transparency "will help people recognise when they are interacting with AI or when content has been generated or altered by AI, reducing the risk of deception and manipulation."

Convina's view: Europe just did what Washington won't — wrote readable rules for the interface layer instead of hoarding model weights behind guest lists. The deepfake split is the sleeper provision: invisible watermarks satisfy providers, but deployers need labels humans can actually see, which means marketing teams and newsrooms — not just ML engineers — own compliance. Twelve days is not enough to rebuild a product pipeline, but it is enough to panic a procurement team still routing EU customer service through unaudited agents. Sign the Code by Wednesday or document your alternative. Pretending the chatbot is "obvious" is the first fine waiting to happen.