AI in pre-employment assessments under the EU AI Act: games, personality, video and the Annex III reality

Pre-employment assessments have long been a recruiter instrument to compare candidates objectively. With the rise of AI-driven assessment platforms — game-based tests, video interviews with scoring, personality models, technical assessments with automated evaluation — that instrument has fundamentally changed. For the EU AI Act this means: almost all modern pre-employment assessments hit Annex III point 4(a) high-risk, and the practical implications are significant.

This post explains which assessment types fall under which classification, where the legal hooks sit, and which vendor questions you want answered before every deployment.

Which assessment types we distinguish

In practice we see four main categories of AI pre-employment assessments:

• Game-based cognitive assessments — Pymetrics, Cognify, Plum: interactive games measure cognitive abilities and behavioral indicators via AI analysis
• Video interview AI — HireVue, Modern Hire, Talview: candidates answer structured questions on video, AI scores transcripts and (sometimes) tonality
• Personality and behavioral assessments — SHL, Talogy, Saville, Hogan: questionnaires with AI scoring and match against role profiles
• Technical and skills assessments — HackerRank, Codility, TestGorilla, Karat: technical tasks with automatic evaluation and (increasingly) AI-suggested ranking

Each category has its own risks, but under AI Act Annex III point 4(a) it all sits in the same basic drawer: AI scores candidates for selection purposes.

Why assessments are the sharpest 4(a) example

CV screening can often be defended with arguments down to pure ordering; assessments cannot. The whole product of a pre-employment assessment is a score per candidate that affects selection. Arguments that do not work:

• "The recruiter sees the score but decides themselves" — if the score systematically affects the shortlist, it stays 4(a). Here it always lies
• "It is not AI, it is statistics" — game-based assessments and NLP on transcripts are ML models. The AI Act is technology agnostic but these techniques fall under it explicitly
• "We only use it for one role/pilot" — Article 26 makes no distinction on scale

For regulators and HR lawyers the assessment category is the most evident high-risk area. That also means it is the place where you need the most robust evidence stack.

The legal hooks per assessment type

Game-based assessments — bias risk is known (Pymetrics commissioned research by Northeastern University themselves). The question is not whether bias exists but whether the specific deployment in your population accentuates bias. Validation question is crucial.

Video interview AI — since 2021 without facial analysis (HireVue removed after public criticism), but NLP on transcripts touches language bias, accent bias and cultural differences in expression. Accessibility question (candidates who cannot make videos for medical reasons) is a separate obligation.

Personality assessments — models are often validated for general populations but less for specific EU groups. Ask for validation studies in your market.

Technical assessments — less bias-sensitive on demographics, but on educational context. A candidate without elite CS education may get less recognition despite equivalent skills.

Step-by-step for your assessment AI dossier

What to do now

For employers with assessment AI in their application process — and that is more than you think, including SMEs and scale-ups — the practical order: tool inventory this week, vendor validation in writing within 30 days, candidate notice in your process this month, and dossier via the HR AI hub and the HR AI Evidence Pack. For video interview AI specifically: see the HireVue analysis.