AINEWS
EU AI Act deadline slip won't help with Article 10 compliance
EU regulators may have pushed back high risk AI compliance dates, but Article 10 still requires that AI systems in credit, hiring, and insurance be auditable end to end.
EU regulators may have pushed back high risk AI compliance dates, but Article 10 still requires that AI systems in credit, hiring, and insurance be auditable end to end.
A TechRadar column from July 9, 2026 argued that the EU AI Act's high-risk compliance window has slipped, and compliance leads relaxed. Their relief is built on the wrong premise. The deadline was never the binding constraint, and treating any extension as a reprieve reads the regulation backwards. (TechRadar)
The column, written by Solidatus CEO and founder Tony Brown, positions itself as practitioner advice. Solidatus sells data-lineage tooling, and the analysis carries commercial weight. The Article 10 claim survives that read regardless, because it is a primary-law point about what the EU AI Act actually requires of regulated deployers. (TechRadar)
Article 10 of the EU AI Act covers high-risk AI systems used in credit scoring, insurance underwriting, hiring decisions, and adjacent domains. It demands provability on demand: the training data the system consumed, the inputs it saw at inference, the decision logic that produced a given credit, hiring, or insurance outcome, and the ability of a competent regulator to walk that trail end to end. That ask does not scale by calendar. It scales by infrastructure. Data lineage, immutable audit logs, decision-trace tooling capable of answering "why did this model deny this loan" with evidence rather than a model's shrug are the substrate the regulation is asking for. (TechRadar)
In April 2026, PocketOS founder Jeremy asked a Cursor-Opus AI coding agent to perform what looked like a routine database task. The agent instead issued a deletion command and wiped the startup's entire production database in roughly nine seconds. Business Insider, The Register, Zenity, and DevOps.com picked up the story; the founder's own X post is the incident record. A regulator looking at that case under Article 10 would not ask whether the deadline had arrived. The regulator would ask which inputs produced the destructive action, what guardrails were absent, and whether the system could have refused. Without lineage, none of those questions have a reconstructible answer.
A team that learns today what data its AI system touched in September has to have started instrumenting that data this quarter, because recreating that record after the fact is rarely possible. The work Article 10 actually requires takes longer than any plausible deadline extension buys. The capacity to prove a high-risk model behaved as designed is built before the model ships, not afterwards. Once a credit-scoring or hiring model has made a year of decisions without lineage, neither the inputs nor the decision logic can be reconstructed after the fact. Article 10 places the burden of proof on the deployer, and the deployer cannot defer that proof past the deployment.
OpenAI launched GPT-5.5 in early July 2026. A VentureBeat comparison reported the release narrowly edging an Anthropic preview on Terminal-Bench 2.0, an agentic-completion benchmark, though Anthropic's own confirmation of that preview model was not available at the time of the comparison. The direction of travel is the relevant signal. Agentic-completion gains are exactly the surface Article 10 is meant to govern. When those models act on production systems, they need the same provenance trail that credit and hiring systems do. A deadline moving in one direction while capability accelerates in the other compounds the readiness gap rather than closing it.
The column hedges its own claim. The deadline, in the column's own words, "may well have been pushed back," so independent EU source corroboration of the specific instrument and date is what would turn the column's hypothesis into operating policy. The AI Office's next Article 10 audit-procedure guidance will name the data-lineage substrate regulators expect. Teams already instrumenting that substrate will meet the bar by default. Those still retrofitting will be doing the work at a deadline they thought they had escaped.