Chapter 4
Artificial intelligence and performance appraisal
Measuring work, context and human review in decisions with career consequences.
4.1 What is to be measured
Performance includes outcomes, quality, collaboration and contributions appropriate to the role. Not all these dimensions can be observed through digital records. Indicator selection must begin with a definition of the work and agreement on which outcomes matter.
Proxy indicators can be useful but require caution. Message volume or time online may reflect workload, work style or technical difficulties. Turning them into an automatic productivity measure encourages behavior aimed at the indicator and can undermine the purpose of the work.
4.2 Goals and feedback
Feedback must be understandable, timely and actionable. A system can organize evidence or provide reminders of commitments, but professional conversations need context. Discrepancies between data and experience must be investigated, rather than resolved by assuming that the digital record is always superior.
Goals need review when resources, priorities or responsibilities change. Assessment against criteria that a person could neither know nor influence is problematic. Documentation should make it possible to reconstruct what was expected, what happened and which factors shaped the outcome.
4.3 Bias and comparisons
Comparisons between people can confuse opportunities with outcomes. Teams with different resources or tasks of different difficulty should not be ranked without considering those conditions. Historical appraisals may also contain biases that a model learns and reproduces.
Auditing must analyze relevant errors and effects rather than rely only on average accuracy. Some fairness analyses require processing specially protected information; its necessity and legal basis must be assessed before collection. Measuring inequalities does not authorize any form of data processing.
4.4 Review and disagreement
The person assessed must know the relevant criteria and have a channel for correcting data or providing context. Review requires authority separate from the automatic execution of a ranking. An appeal procedure without the ability to change the result has no practical effectiveness.
Decisions on pay, promotion or continued employment require particular attention. Restrictions on solely automated decisions and other data protection obligations must be analyzed case by case. Human intervention must not be routine confirmation of the system's recommendation [R2].
4.5 Development versus control
Development-oriented appraisal identifies support and learning opportunities. Constant surveillance, by contrast, can reduce trust and autonomy. Before expanding data capture, the organization should ask whether a less intrusive alternative can achieve the same purpose.
Success means improving conversations, decisions and working conditions rather than producing more scores. The organization must review whether the system supports collaboration or introduces unproductive competition, self-censorship and efforts directed solely at appearing active.
Questionnaires
Chapter knowledge & Human Resources dilemmas
Dilemmas receive no moral or employability score and are not used to assess job candidates.