Judgment in the Age of AI · Article 28
Recruitment tools may rank CVs, analyse assessments or recommend applicants, but using the same model for everyone does not automatically make a process fair. Historical data can carry yesterday’s preferences forward, while proxy measures may indirectly exclude particular groups.
Five questions an employer should answer
- What data and measures does the tool actually use?
- What evidence connects those measures to the essential requirements of the job?
- Do selection and error rates differ materially between groups?
- Are reasonable adjustments available for disability or exceptional circumstances?
- Can applicants learn that AI was involved, correct bad data and request human review?
Audit the process, not only the model
A tool with respectable overall accuracy can still produce unfair outcomes when placed at the wrong stage without review. Who sets thresholds, handles exceptions and can reverse a result often matters more than a vendor’s headline score.
Fairness is not making a machine do the same thing to everyone. It is preventing job-irrelevant characteristics from deciding opportunity.
After an automated rejection, an applicant can ask what information was used, whether automation assisted the decision, and whether human review or an appeal is available.
References
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.