EU AI Act Psychometric Testing 2026: HR Obligations for High-Risk Recruitment

A score can inform a hiring decision. It cannot quietly become the decision.

EU AI Act psychometric testing 2026 is not a distant legal topic. It is an operating question for every HR team using automated scores, rankings, recommendations, or candidate screening. On 2 August 2026, many rules for high-risk systems become applicable. Your assessment may remain useful. Your process may still need a redesign. The real issue is simple: does an AI-supported result materially influence who gets interviewed, promoted, assigned, or removed?

EU AI Act psychometric testing 2026: what changes now?

The calendar creates a hard deadline

Regulation (EU) 2024/1689 applies in stages. The first prohibited AI practices applied from 2 February 2025. AI literacy obligations apply from 2 February 2026. Many requirements for high-risk AI systems apply from 2 August 2026. Some product-related rules arrive later, during 2027. This timetable matters because recruitment teams often discover their automated tools only when a vendor renewal, audit, or candidate complaint forces the issue.

The European Commission sets out the regulation and its phased application. Read the dates. Then map them against your actual recruitment workflow. Do not rely on a marketing label such as “AI-enabled” or “human-centred.” What does the platform calculate? Who sees the output? What decision follows?

A test is not automatically prohibited

Psychometric testing is not banned by the EU AI Act. A personality inventory used by an occupational psychologist is different from a black-box tool that rejects applicants after an automated score. The distinction is not the word “test.” The distinction is the use. A questionnaire can support structured discussion. It can also become a hidden filter. Those are not the same activity.

EU AI Act psychometric testing 2026 asks a practical question: can your team explain why a person moved forward or stopped moving forward? If the answer is “the system decided,” your risk is clear. If the answer is “a trained reviewer considered several documented factors,” you have a stronger starting point.

Employment decisions receive special attention

Recruitment decisions shape access to work. They can also amplify bias at scale. Annex III, point 4 of Regulation (EU) 2024/1689 explicitly covers AI systems used in employment, worker management, and access to self-employment. This includes systems used to recruit or select people, particularly to place targeted job advertisements, analyse applications, or evaluate candidates.

Your process may involve one platform or five. A video assessment. A cognitive test. A personality report. An applicant ranking. A recruiter dashboard. Each element needs review. EU AI Act psychometric testing 2026 is about the chain of decisions, not only the questionnaire at the start.

AI Act high-risk recruitment: are you using a covered system?

The outcome matters more than the interface

An assessment can look harmless. A clean dashboard does not reduce regulatory exposure. The relevant question is whether an AI system influences access to employment or a work-related opportunity. Does a score determine interview priority? Does a recommendation guide a recruiter toward one group? Does an automated threshold remove people before human review? These uses may bring the system within AI Act high-risk recruitment rules.

Consider a daily example. A recruiter receives 300 applications. The platform ranks applicants from 1 to 300 using assessment results and CV data. The recruiter reviews only the top 30. The recruiter still clicks the final approval. Yet the ranking shaped the available choices. That is meaningful influence. Document it honestly.

Article 6(3) is an exception, not a shortcut

Article 6(3) AI Act can exclude a system from the high-risk category where it does not pose a significant risk of harm to health, safety, or fundamental rights. The exception can apply where the system performs a narrow procedural task, improves a completed human activity, detects decision patterns without replacing assessment, or prepares information for a human decision.

That exception needs evidence. “Our recruiter remains involved” is not enough. What can the recruiter override? How often do they do so? Can they see the reasons behind a score? Can an applicant be assessed fairly if the algorithm is wrong? Keep written answers. A vague claim will not protect your team.

Create an inventory before you classify

You cannot govern what you have not listed. Build one inventory across recruitment, internal mobility, onboarding, performance review, and workforce planning. Include purchased systems, embedded tools, and supplier-managed services. Record the intended purpose and the real purpose. Those can differ.

• Record the tool. Include supplier, product version, owner, and contract date.
• Record the data. List responses, CV information, video, voice, behavioural data, and inferred traits.
• Record the outcome. State whether the output ranks, recommends, rejects, or supports discussion.
• Record human oversight. Name the reviewer and explain their authority to depart from the result.

HR obligations 2 August 2026: what teams need to document

Risk management cannot remain a supplier promise

Articles 9 to 15 set core obligations for high-risk AI systems. Providers carry substantial duties. Deployers also have responsibilities. HR leaders should not assume that a signed supplier statement solves everything. Your team controls implementation. Your team decides which vacancy uses the tool. Your team sets thresholds. Your team receives candidate concerns.

Article 9 requires a risk management system. In practice, this means identifying foreseeable harms, testing controls, reviewing residual risk, and retaining evidence. Ask a difficult question: what could go wrong for a qualified person in your process? A false rejection has a human cost. A biased ranking can quietly narrow your talent pool.

Data quality and bias testing need proof

Articles 10 and 15 address data governance, accuracy, robustness, and cybersecurity. For psychometric testing, this means more than asking whether the tool has been “validated.” Valid for whom? Valid for which role? Valid in which language? Valid after a scoring model changes? A benchmark from one role cannot automatically justify use in another.

ISO 10667 provides guidance for assessment service delivery in work settings. It supports disciplined thinking about method quality, roles, confidentiality, and feedback. Use it as a practical benchmark alongside your legal review. A credible process records test purpose, job relevance, administration conditions, interpretation rules, and review dates.

Human oversight recruitment needs real authority

Article 14 requires effective human oversight for high-risk systems. A person who merely receives an automated recommendation is not necessarily exercising oversight. They need competence, information, time, and authority. They need to understand the output’s limitations. They need to intervene or stop use when the result appears unreliable.

“Human oversight” means a person can understand, question, and override the output. It does not mean a person clicks “approve.”

Start with a simple operating rule. No applicant should be rejected solely because of an automated psychometric score. Set escalation paths for unusual profiles, accessibility concerns, incomplete responses, and suspected bias. Then train reviewers to use them.

Candidate transparency AI: can applicants understand the process?

Clear notice builds trust before consent screens appear

Candidate transparency AI is not a long legal notice buried under a submit button. People deserve a plain explanation of how an assessment supports the selection process. Tell them what the assessment measures. Tell them how results will be used. Tell them whether AI contributes to scoring, ranking, or recommendations. Tell them who can review the outcome.

Good transparency also improves candidate experience. It reduces confusion. It creates better feedback conversations. It gives recruiters a consistent explanation when applicants ask why an assessment was included. If your team cannot explain the process in plain English, the process is probably too opaque.

Feedback must not overclaim certainty

Psychometric results describe probabilities and preferences. They do not reveal a person’s entire ability, motivation, or future performance. Avoid language that turns a profile into a verdict. Avoid labels that close the conversation. A candidate is more than one score, one response pattern, or one comparison group.

Use structured feedback. Explain the assessment’s purpose. Discuss relevant work behaviours. Invite context. Record any material information that affects interpretation. This protects the individual and improves decision quality. It also reinforces the central discipline behind EU AI Act psychometric testing 2026: the tool supports judgement; it does not replace it.

AI literacy becomes operational on 2 February 2026

From 2 February 2026, organisations using AI systems need to take measures to ensure sufficient AI literacy among staff and others dealing with those systems. This is not a one-hour slide deck for everyone. Recruiters need practical knowledge. Occupational psychologists need model and validation context. Procurement teams need supplier questions. Senior leaders need accountability.

Create role-based learning. Train recruiters to identify automated recommendations. Train managers to avoid overreliance. Train administrators to recognise poor-quality data. Keep attendance and training materials. The evidence may matter later. More importantly, the learning may prevent a harmful decision today.

Psychometric test compliance: what remains valuable?

Use assessments for a defined work purpose

Psychometric test compliance begins with relevance. A personality assessment should connect to documented work behaviours. A reasoning test should relate to the role’s cognitive demands. A structured assessment should support a decision already grounded in clear selection criteria. If you cannot state the purpose in one sentence, pause the deployment.

Ask: what decision will this assessment improve? What evidence would we use without it? What risks appear if we use it badly? These questions are not bureaucracy. They protect candidates from arbitrary selection and protect HR teams from tools that create more noise than value.

Combine evidence instead of chasing one perfect score

Strong recruitment decisions use more than one relevant source of evidence. A structured interview can test examples of behaviour. Work samples can show practical capability. Psychometric tools can add information about preferences, reasoning, or working style. Each source has limits. Together, they support a more balanced decision.

Do not set an automatic exclusion threshold without documented validation and human review. Do not compare scores across unrelated roles. Do not use a report created for development as a hidden selection filter. These small choices determine whether a tool supports fair selection or creates an avoidable risk.

SIGMUND assessments support informed decisions

SIGMUND assessments can help teams structure recruitment conversations and compare job-relevant evidence. The value comes from disciplined use. Define the role. Select appropriate measures. Prepare reviewers. Discuss results in context. Retain human responsibility for the final decision. That is the standard candidates should expect.

Explore the available assessment catalogue before choosing a method. If your team needs a broader view of assessment delivery, review the HR assessment approach. The aim is not more automation. The aim is better judgement, documented clearly.

Explore HR assessments

EU AI Act psychometric testing 2026: your first compliance actions

Complete this review before the deadline

Do not wait until August 2026 to start. An inventory takes time. Supplier evidence takes time. Training takes time. Workflow changes take time. Start with your highest-volume recruitment process and your most automated decision point. That is where a small error can affect the greatest number of people.

1. List every assessment and AI-supported recruitment tool.
2. Map each output to the decision it influences.
3. Assess whether Annex III, point 4 may apply.
4. Request technical, validation, bias, and oversight documentation from suppliers.
5. Remove fully automated rejection routes.
6. Create clear candidate notices and reviewer guidance.
7. Train users before 2 February 2026 and review high-risk readiness before 2 August 2026.

Know the financial and reputational exposure

The AI Act includes serious penalties. Infringements involving prohibited practices can reach €35 million or 7% of worldwide annual turnover, whichever is higher. Other infringements can reach €15 million or 3% of worldwide annual turnover. Incorrect, incomplete, or misleading information may lead to fines of up to €7.5 million or 1% of turnover. The applicable figure depends on the organisation and breach.

Money is not the only cost. A candidate who believes they were screened unfairly may speak publicly, escalate internally, or seek legal advice. A weak process can damage trust with applicants, employees, and hiring managers. Good governance is not merely defensive. It makes your recruitment process easier to explain and easier to improve.

Make the next decision easier to defend

Your next assessment decision should survive a simple review. Why was this tool used? Is it relevant to the role? What data did it use? Was the candidate informed? Could a trained person challenge the output? Was the final decision based on more than one piece of evidence? These questions turn compliance into a working habit.

EU AI Act psychometric testing 2026 is a prompt to raise the standard. Keep the useful parts of psychometrics. Remove hidden automation. Give people meaningful oversight. Build a record that shows your team made a thoughtful decision.

EU AI Act psychometric testing 2026: classify the tool first

EU AI Act psychometric testing 2026 starts with one hard question. Does the tool influence who moves forward, who is rejected, or who receives an interview? If yes, treat that system seriously. Annex III, point 4 of Regulation (EU) 2024/1689 places AI used in recruitment and worker selection within the high-risk category.

A personality questionnaire is not automatically high-risk. A scoring engine using machine learning may be. A cognitive assessment is not automatically high-risk either. The key issue is the AI system, its intended purpose, and its influence on a people decision. Your vendor label does not decide this. Your actual use does.

Map every decision point

List each assessment used during recruitment. Include the tool name, version, role, score, and decision owner. Ask where automated processing enters the journey. Does it rank applicants? Does it produce a recommended shortlist? Does it identify “culture” signals from video, voice, or written answers? Small settings can create large exposure.

• Record the purpose of each assessment.
• Identify whether machine learning produces or changes scores.
• Document the person who makes the final hiring decision.
• Separate assessment evidence from automated rejection rules.

Do not confuse automation with AI

A fixed scoring rule is not necessarily an AI system under the Regulation. A machine-learning model can be. Article 6(3) offers a narrow exception where an Annex III system does not pose a significant risk of harm and does not materially influence decision-making. This is not a shortcut. It requires a documented assessment. If the system profiles people, the exception cannot apply.

EU AI Act psychometric testing 2026 therefore needs a written classification decision. Save it. Date it. Review it after every material change. A new model, new data source, or new role can change the answer.

AI Act high-risk recruitment: what changes on 2 August 2026?

EU AI Act psychometric testing 2026 becomes operational on 2 August 2026. That is the date when core high-risk system requirements apply. The deadline matters because hiring teams often focus on a vendor contract, then forget daily use. Compliance lives in the workflow. It lives in the recruiter’s screen. It lives in the final decision record.

The Regulation requires risk management, data governance, technical documentation, record keeping, transparency, human oversight, accuracy, robustness, and cybersecurity under Articles 9 to 15. A test score alone is not a hiring decision. Yet it can shape one. That is why HR needs clear controls before a score reaches a manager.

Create evidence that survives scrutiny

For every high-risk assessment, retain a technical file. It should show what the tool measures, which role it supports, how performance was validated, and when validation occurred. Include the model version. Include data fields. Include known limitations. Include the owner who approves use.

The EU Artificial Intelligence Act documentation explains prohibited practices under Article 5. Recruitment tools that infer emotions in the workplace or during hiring have been prohibited since 2 February 2025. Do not ask a video tool to score enthusiasm, honesty, or emotional state.

Build real human oversight recruitment controls

Human oversight recruitment is not a person clicking “approve.” The reviewer needs authority, training, and enough context to disagree with the system. They need to understand what a score means. They also need to understand what it does not mean. A low score may reflect a narrow measure, a poor test condition, or a role mismatch. It does not define a person.

• Assign a trained reviewer before any adverse decision.
• Show score explanations and assessment limits.
• Allow an override with a written reason.
• Audit overrides every quarter.

The strongest control is not a score. It is a qualified person who can question it, explain it, and overrule it.

EU AI Act psychometric testing 2026: candidate transparency AI

EU AI Act psychometric testing 2026 changes the conversation with applicants. Candidate transparency AI means people should understand that they are interacting with or being assessed by an AI system when relevant. They should receive useful information in plain language. Not a legal wall of text. Not a vague statement hidden in a privacy notice.

Tell applicants what the assessment measures. State why it is used for the role. Explain whether AI supports scoring or recommendation. Describe the human review stage. Provide a route for questions or accessibility requests. This helps people prepare. It also helps your team detect unclear processes before they become complaints.

Use a clear notice before testing

A good notice answers practical questions. What will happen? How long will it take? Which data will be used? Who sees the results? Can a person review the outcome? Your recruitment team should be able to repeat this explanation in a normal conversation. If they cannot, the notice is too complex.

Review the assessment experience across devices. A timed test may create avoidable barriers. A personality measure may need clear instructions and adequate completion time. Candidate transparency AI also means explaining reasonable adjustments and ensuring that candidates know how to request them.

Avoid prohibited emotional inference

Do not use facial analysis, voice analysis, or behavioral signals to infer a candidate’s emotions during recruitment. The prohibition has applied since 2 February 2025. A claim that the system only “supports” the recruiter does not remove the risk. If the output labels someone as engaged, nervous, confident, or untruthful from biometric signals, stop and investigate.

Psychometric test compliance: run the six-record audit

Psychometric test compliance becomes manageable when each tool has a simple record. Start with six fields. Purpose. Target role. Data used. Validation date. Responsible owner. Human override rule. Add the system version and vendor contact if possible. This is not paperwork for its own sake. It answers the question every HR leader will face: why did we trust this result?

The assessment test catalogue can help teams compare structured assessment options against their intended hiring use. The goal is not to collect more tests. The goal is to use each assessment with a clear, proportionate purpose.

Validate for the role, not for the brochure

Validation needs a date and a role context. A tool that supports customer service hiring may not provide the same evidence for a senior technical role. Ask whether the assessment measures relevant work-related characteristics. Ask how the result is interpreted. Ask which groups were included in validation work. Then record the answers.

ISO 10667 provides a useful benchmark for assessment service delivery. It focuses on agreements, methods, qualified assessors, and reporting. Use that standard to structure questions for providers, even when the AI Act creates additional obligations.

Test fairness after deployment

Bias controls cannot end at launch. Review outcomes over time. Compare progression rates across legally protected groups where lawful and appropriate. Investigate large differences. Look for adverse effects from language, timing, disability access, device requirements, or score thresholds. A model can perform well in testing and still fail in a real recruitment process.

1. Set a review date before deployment.
2. Monitor pass rates and override patterns.
3. Record suspected bias incidents.
4. Pause use where material harm may occur.
5. Revalidate after model or workflow changes.

HR obligations 2 August 2026: assign owners and train users

HR obligations 2 August 2026 cannot sit only with legal counsel or procurement. Recruiters use the tool. Occupational psychologists interpret results. Managers act on recommendations. Technology teams manage access and integrations. Each person needs a defined role. Otherwise, a score moves through the process with no accountable owner.

AI literacy requirements began on 2 February 2026. That means organizations should ensure people dealing with AI have sufficient knowledge for their role. Training does not need to be theatrical. It needs to be specific. Can the recruiter identify an unreliable output? Can the manager explain why a score was overridden? Can the system owner identify a prohibited feature?

Give each role one clear responsibility

• HR owner: approves the intended use and candidate communication.
• Assessment specialist: reviews validity, interpretation, and fairness evidence.
• Technology owner: controls version records, access, and system logs.
• Hiring manager: makes the final decision with documented reasoning.

Keep logs that answer practical questions

Record keeping should show what happened, when it happened, and who acted. Retain the assessment version, generated output, reviewer decision, override reason, and incident notes. This makes internal reviews faster. It also makes vendor discussions more direct. When a tool changes, ask one question first: what changes in our evidence file?

For broader assessment planning, explore HR assessment options built around structured decision support. Clear evidence gives managers confidence. It also gives candidates a more respectful experience.

AI Act sanctions 35 million: turn risk into action now

AI Act sanctions 35 million is not a slogan. For prohibited AI practices, the Regulation provides for fines of up to €35 million or 7% of worldwide annual turnover, whichever is higher. Other infringements can reach €30 million or 6% of worldwide annual turnover. These figures demand attention. Yet the immediate business risk often appears earlier: unfair rejection, damaged trust, poor documentation, or a stalled hiring process.

EU AI Act psychometric testing 2026 gives HR a deadline. Use it as a design moment. Do not wait for a complaint. Do not wait for an audit. Build the process your team would be proud to explain to an applicant.

Complete this 30-day action plan

1. Inventory every assessment and automated recruitment tool.
2. Classify each use case under Article 6 and Annex III.
3. Stop any emotional inference feature in hiring.
4. Create the six-record file for every psychometric tool.
5. Define human oversight and written override rules.
6. Refresh candidate notices before testing begins.
7. Train users on AI literacy and escalation routes.
8. Schedule fairness, performance, and version reviews.

Make the next hiring decision easier to defend

The best recruitment process does not pretend that people can be reduced to a number. It uses evidence carefully. It gives qualified people room to judge. It explains the process clearly. It records what matters. That is the practical standard for EU AI Act psychometric testing 2026.

Want a structured starting point? Read more HR guidance in the SIGMUND assessment blog. Then review one live recruitment workflow this week. One tool. One role. One decision path. That is where compliance becomes real.