AI assessments and games in selection: high-risk, validity and accessibility

Short answer: AI assessments — gamified tests, psychometric analyses, automated case scoring — evaluate candidates and therefore fall under Annex III, point 4 of the AI Act as high-risk. Three questions decide whether such a tool passes muster: does it really measure something related to the job (validity), is it free of bias, and does it not exclude people with a disability? And note: an assessment that infers emotions via image or voice is outright prohibited.

Validity: does it measure what it promises?

The big promise of AI assessments is objectivity. But a model that correlates with "successful past employees" often mainly measures what the past looked like — not what the job needs now. Without substantiated, job-relevant validation, a high "fit score" is false certainty. Ask the supplier for validation evidence and keep it.

Bias: looking objective is not being objective

As with CV screening, assessments learn from data containing past patterns. A game rewarding reaction speed or language use can systematically favour certain groups. The risk of discrimination is real and lies with the employer.

Accessibility: the silent exclusion

Gamified assessments assume a "standard candidate": good eyesight, fast clicking, fluent in the test language. Anyone with a visual, motor or cognitive disability, or less command of the language, scores lower for reasons unrelated to the job. Without reasonable accommodation that becomes indirect discrimination.

The emotion line

Assessments with video or voice analysis sometimes promise to measure "soft skills" or "engagement". To the extent that amounts to inferring emotions, it is prohibited under Article 5 — not "if done carefully", but no.

What to do

• Treat every selection assessment as high-risk: human oversight, monitoring, information to the candidate.
• Demand validation evidence from the supplier: does it measure job-relevant traits?
• Test for bias and for outcomes by group.
• Offer reasonable accommodation and an alternative route for those who cannot take the assessment fairly.
• Drop emotion and "soft skill from video" features.

An assessment that does not measure what the job needs, is not free of bias and is not accessible is not an objective filter but a legal risk with a scientific veneer.