How Data Poisoning Attacks Actually Work, and Why Signature-Based Defenses Miss Them

Any organization deploying a chatbot, a system that generates synthetic audio, image, video, or text content, or an emotion-recognition or biometric-categorization system now has a binding obligation to disclose that fact to end users. AI-generated content that could be mistaken for authentic material must be machine-readably marked as artificial, and deepfakes must carry a clear disclosure. This applies regardless of whether the underlying system is classified as high-risk.

The Commission's authority to enforce obligations on providers of general-purpose AI models - including the systemic-risk tier covering the largest foundation models - is also active from this date. Enterprises building on top of general-purpose models inherit downstream documentation obligations even where their own use case is not classified as high-risk.

Standalone high-risk systems - covering biometric identification, critical infrastructure, education and vocational training, employment decisions, access to essential services such as credit scoring and insurance, law enforcement, migration, and administration of justice - received an additional 16 months. Product-embedded high-risk systems under existing EU product-safety law, such as medical devices, toys, and lifts, received 12 additional months. Neither obligation disappears; both are still coming, on a schedule regulators have now made explicit.

When the high-risk obligations do land, Article 10 requires that training, validation, and testing datasets be relevant, sufficiently representative, and - to the extent possible - free of errors, with documented data governance and management practices appropriate to the system's intended purpose. This is not a checkbox exercise completed once before deployment; it requires an ongoing, defensible record of what data trained the system, where it came from, and whether it has been altered since. Building that evidence trail after the fact, in the final months before a December 2027 deadline, is materially harder than building it into the data pipeline now, while there is still runway.