V Skill HF-D7.7-008 cross-sector 7 skills
Recognising AI risk and bias
Recognises where an AI system works systematically skewed or unreliably for groups or situations and names consequences and possible measures.
How hrmforce measures this
- Assessment method
- Work sample test · rho 0.33 (SD 0.09)
- hrmforce instrument
- Knowledge test (client-specific), Work sample test
- Competency (50-framework)
- Judgment
- Trainability
- medium
- Demand outlook 2026 to 2030
- rising
The candidate performs a representative work sample under standardised conditions.
Behavioural anchors
| Level | Behaviour at this level |
|---|---|
| N1 Guided | States that AI output can be skewed by the data used and refers striking outcomes to a colleague. works under supervision and follows instruction · routine, one variable at a time · own task |
| N3 Proficient | Examines outcomes per group or situation, identifies where the system falls short and proposes measures or extra checks. sets own approach and seeks input proactively · several variables, some ambiguity · own team or process |
| N5 Leading | Sets the assessment frameworks for AI fairness and risk and decides whether a system may remain in use. sets the standard and the policy · strategic, under high uncertainty · organisation, value chain or profession |
N2 and N4 are deliberately not anchored. Raters place them between the anchors, following the O*NET convention.
Underlying skills
These skills inherit the assessment route and the behavioural anchors of this construct.
| T | Skill | Definition | Demand outlook 2026 to 2030 |
|---|---|---|---|
| V | Auditing a model for bias Bias audit · Fairness analysis | Compares outcomes and error rates of a model across groups and reports where it works out skewed. | rising |
| V | Recognising proxy variables Proxy variables | Recognises attributes that indirectly point to origin, gender or health and judges whether their use is defensible. | rising |
| V | Assessing dataset representativeness Dataset representativeness | Assesses whether the training data covers the groups and situations the model will later be used for. | rising |
| V | Weighing the impact of a wrong prediction Error cost analysis | Weighs what a false positive or negative outcome means for a person and sets the threshold accordingly. | rising |
| V | Monitoring model drift Model drift | Monitors whether model performance and input data shift over time and adjusts the model when needed. | rising |
| V | Setting up an objection and complaint route Appeal process | Arranges that people can contest an outcome and that a person reviews the case again on its merits. | rising |
| G | Communicating the limits of an AI system Communicating AI limits | Explains to users what a system is and is not meant for and where they must keep looking themselves. | rising |