On 2 August 2026, the EU AI Act's core requirements for high-risk AI systems become enforceable. Annex III classifies AI systems used to analyse and filter job applications, evaluate candidates, and affect employment decisions as high-risk.

The banned category is explicit: AI that infers sensitive attributes, including emotional states, from behavioural proxies. This is precisely what a generation of hiring tools was built to do.

Compliance documentation will force a question most vendors have never had to answer publicly. What psychological construct does this score represent? What is the theoretical basis for the claim that this signal predicts job performance? Has that construct been validated against relevant criteria in this population?

These are not new questions. They are the foundational questions of rigorous psychological assessment, asked and answered in clinical, educational, and occupational contexts for most of the last century. Hiring AI largely skipped that step.

You cannot recover months or years of behavioural data collection and retroactively establish what it was measuring or under what conditions. The evidence is either there from the start or it is not. For high-stakes decisions like hiring, where outcomes are discrete and often irreversible, that is not a technical limitation. It is a fundamental validity problem.

August 2026 is forcing a conversation that should have happened at the point of purchase.