Ethical AI use in recruiting: four principles to avoid discrimination

Executive overview

AI doesn't eliminate hiring bias — it transforms it into something harder to detect and challenge. When an algorithm screens candidates, its errors are treated as objective outputs, making discrimination easier to deny. This is automation bias: the tendency to trust automated systems even when they're wrong.

The risk is real, but so is the opportunity. Constrained to the right tasks, AI saves time and improves consistency. The goal is using it as an organiser, never as a judge.

Bias doesn't disappear in AI systems — it becomes invisible.

Four principles for ethical AI use

• Assume AI will introduce new bias, not eliminate existing bias
• Maintain human accountability — you own the tool's outputs, not the algorithm
• Monitor outcomes more carefully than you would with human decision-making; track false positives and false negatives by demographic group
• Preserve human judgment — AI can organise information, but humans must make the final hiring call

Where AI adds value in recruiting

• Job description writing: generates structured first drafts quickly; improves formatting and keyword visibility
• Candidate outreach personalisation: crafts tailored messages using resume or portfolio inputs — always review before sending
• Interview question generation: builds consistent behavioural question banks, reducing interviewer deviation and levelling evaluation criteria

Where AI crosses the line

• Ranking or scoring candidates — even unintentional scoring can systematically favour certain groups
• Sorting (filtering by minimum qualifications) is acceptable; ranking (assigning relative scores) is not
• Never assume consistency equals fairness — applying a flawed rule equally still magnifies bias
• Amazon's AI hiring tool is the canonical example: trained on historical resumes skewed male, it penalised phrases like "women's chess club captain" and downgraded graduates from women's colleges

Implementing AI responsibly

• Get legal approval before any AI tool is used; clarify what data is collected, stored, and processed
• Create specific use guidelines — define exactly what AI is allowed to do and where it stops
• Implement mandatory human review for every AI-assisted step; no output moves forward unchecked
• Track diversity metrics and hiring outcomes continuously; bias hides in data, not headlines
• Train your team to question AI, not just operate it — knowing when not to use a tool matters