Business & Money
Using AI in Hiring: What Is Defensible and What Is Reckless
By Jim Vernon, Editor, AI Intelligence International · Published 11 February 2026 · Reviewed against our editorial standards · About the author
Hiring is the highest-regulated everyday use of AI in most businesses, and the one where enthusiasm most often outruns caution. Automated ranking of people creates direct legal exposure in many jurisdictions.
The defensible line is straightforward: use it to structure and standardise human decisions, never to make the decision.
Key takeaways
- Where the legal risk actually sits: Risk concentrates in automated ranking or rejection, in inferred characteristics, and in a lack of explanation.
- Safe uses that genuinely help: Writing clearer job descriptions, generating structured interview questions tied to the role's requirements, producing consistent scoring rubrics, and summarising your own interviewer notes are all low risk and improve decision quality.
- Uses to avoid: Avoid automated rejection without human review, personality or emotion inference from video, social media scraping, and any scoring of characteristics not directly job-related.
- Designing assessments for a world of assisted candidates: Assume every take-home is model-assisted.
Where the legal risk actually sits
Risk concentrates in automated ranking or rejection, in inferred characteristics, and in a lack of explanation. Several jurisdictions require notice, bias auditing, or a human review route for automated employment decisions.
The exposure is not theoretical: patterns in outcomes are measurable after the fact, and a system that quietly disadvantages a protected group produces evidence in your own records.
Assume anything the system produces about a candidate is discoverable and could be read aloud in a tribunal.
Safe uses that genuinely help
Writing clearer job descriptions, generating structured interview questions tied to the role's requirements, producing consistent scoring rubrics, and summarising your own interviewer notes are all low risk and improve decision quality.
Structured interviews with a common rubric are among the best-evidenced improvements available in hiring, and tooling makes them cheap to run properly.
Note-taking assistance also reduces the recency and charisma effects that distort unstructured interviews.
Uses to avoid
Avoid automated rejection without human review, personality or emotion inference from video, social media scraping, and any scoring of characteristics not directly job-related.
Avoid asking a model to predict performance from a CV. There is no reliable signal for that, and the model will reproduce whatever historical patterns exist in its training, including the ones you are legally obliged not to act on.
Also avoid using detection tools to reject candidates for suspected AI-written applications. Detection is unreliable, and the policy question is better handled by changing the assessment.
Designing assessments for a world of assisted candidates
Assume every take-home is model-assisted. Redesign around critique, live reasoning about a provided artefact, and discussion of the candidate's own past decisions.
This is better assessment regardless of AI: it tests judgement rather than production, and it is far harder to fake in conversation.
State your policy in the invitation. Candidates behave predictably when the rules are explicit, and ambiguity punishes the honest ones.
Documentation and audit
Keep records of what tooling was used at each stage, what the human decision-maker saw, and the basis for each rejection. If a decision cannot be explained without referring to a model score, it is not defensible.
Review outcome statistics by stage periodically. A drop-off concentrated at one automated stage is the signal you need to catch early.
Give candidates a clear route to request human review, and honour it.
A workable policy in five lines
Tooling may draft, structure and summarise. Humans read every application that reaches screening. No automated rejection. No inferred characteristics. Every decision has a written, job-related reason.
Publish it. Candidates increasingly ask, and a clear answer is a recruiting advantage in a market where most employers are evasive about it.
The distinction that keeps you out of trouble
There is a workable line between using AI to help a human decide and letting AI decide. Summarising an application, extracting stated years of experience, or drafting interview questions keeps a person accountable for the outcome. Ranking or rejecting candidates automatically moves you into territory that several jurisdictions now regulate directly.
Regulators care about three things: whether candidates were told, whether the system was tested for disparate impact, and whether a human can explain an individual decision. If you cannot answer all three for a given step, don't automate that step yet.
The practical consequence is that automated screening is usually the worst place to start and drafting is usually the best. The value of speeding up rejection is small; the exposure is large.
A worked example: a fifty-application role
For a role with fifty applicants, a recruiter used a model to produce a one-paragraph factual summary of each CV — stated role history, stated skills, notable gaps — with an instruction never to infer or score. Reading fifty summaries took about forty minutes instead of three hours.
Screening decisions were still made by reading the CVs of anyone whose summary looked plausible, roughly twenty of the fifty. The AI narrowed reading order, not the candidate pool, so no applicant was excluded without a human ever seeing their application.
The audit trail was a saved copy of every summary plus the prompt used to produce it. When one candidate later asked why they were not progressed, the answer came from the recruiter's own notes, which is exactly the position you want to be in.
Frequently asked questions
Can we use AI to shortlist at all?
To sort and highlight relevant evidence for a human reader, generally yes with notice. To reject automatically, treat it as high risk and check local law before doing it.
Do we have to tell candidates?
In several jurisdictions yes, and it is good practice everywhere. Notice is cheap; retrofitting it after a complaint is not.
Are AI detection tools usable in hiring?
No. Error rates are high and the false positives skew against non-native speakers, which converts a tooling choice into a discrimination problem.
How do we audit for bias?
Compare pass rates by stage across groups over a meaningful sample, and investigate divergence. Vendor bias claims are not a substitute for your own outcome data.
Do we have to tell candidates we use AI?
In a growing number of jurisdictions, yes, and it is good practice everywhere. A single clear sentence in the job posting covers it.
Can AI write the job description?
Yes, with review. Check for language that narrows the applicant pool unnecessarily — inflated year requirements and unexplained jargon are the common ones.