Before it
Model governance was voluntary commitments and self-published system cards.
The problem
No binding requirement to disclose training data summaries, label AI output or document systemic risk.
What it introduced
Binding obligations for general-purpose models and Article 50 transparency rules, with fines up to 3% of global turnover.
Why it matters
The first enforceable general AI regime in a major market, though the Digital Omnibus proposes delaying parts of it.
What modern models inherit
Compliance as a design constraint on deployed models.
What came next
A grace period into December 2026 for some marking obligations, and continuing amendment.