EU AI Act Article 50 Is Live as of Today, With Fines up to 3% of Global Revenue

As of this morning, Article 50 applies to every AI system that interacts with humans or generates synthetic content in the EU. Conversational agents must disclose they're AI. AI-generated output must carry machine-readable marking. Deepfakes must be labeled. Same day, the European Commission's active enforcement powers over general-purpose AI models switched on: information requests, model access, recall powers.

Engadget's read fills in the operational detail. AI-generated images, audio and text designed to look authentic need a digital watermark or label. Non-compliance runs to 3% of total gross revenue. Exemptions cover private group chats and material that's evidently artistic, satirical or fictional. The EU is publishing standardized black-and-white labels while permitting compliant custom ones.

Here's the trap: the grace period until December 2, 2026 applies only to systems already on the market. Anything you ship new is in scope today. If you're mid-launch on generated-media features for EU users, you don't have four months. You have zero.

The Annex III standalone high-risk obligations that were originally due today got pushed to December 2, 2027 by the June 16 amendments. So disclosure and labeling now, high-risk conformity work later.

What makes this more than a compliance note is what shipped around it. Three unrelated companies made agent audit trails a headline feature within 72 hours of the deadline. Cequence's AI Gateway added AI Discovery, API Registry, LLM Registry and Skill Registry on July 30, with Agent Personas binding a stated job description directly to model, tools, data access and guardrails. Basedash launched Audit Logs on August 1 recording every query its AI data analyst runs with full attribution, alongside SSO/SCIM/RLS. Microsoft's Project Perception opens public preview August 3 with red/blue/green agents that document and deploy their own patches.

Attribution stopped being an enterprise-tier upsell this week. It's the shipping requirement.

The Basedash one is the sharpest example of why. Natural-language BI has been blocked from displacing Looker and Tableau in regulated accounts on exactly one objection: you can't audit what the AI queried. Make the AI's SQL attributable on the same footing as a human's, and the audit story stops being worse than the dashboard you're replacing. That's not a compliance feature, it's a competitive unlock disguised as one.

Contrast with what happens when you skip this. Google shipped a Nano Banana 2-powered "create image" feature in Google Earth web on July 30 and pulled it on July 31 after researchers generated refugees at the Mexican border, an Iranian nuclear plant, an Amsterdam crash, and a bomb-cratered Gaza hospital over real satellite imagery, with nothing refused. NPR independently produced Kharg Island on fire and a flooded US Capitol. A one-day feature lifespan at the company that spent a year selling provenance tooling.

Ship the disclosure. It's cheaper than the rollback.