AI in CV screening and its risks for a small employer

5 days ago
3 min read

Updated: 4 days ago

Introduction

A single advertised vacancy in a small business can produce two hundred applications, most of them unsuitable, and the reading falls to whoever has least time to do it. The temptation to automate the first pass is obvious and the efficiency argument is real.

It is also the application in this whole area with the greatest capacity to cause harm and legal exposure. A screening system learns from past decisions, which means it reproduces the patterns in past decisions, including the ones you would not defend. Efficiency in screening is worth having; it is worth having only with the constraints understood, because the failure mode here affects people rather than numbers.

1. AI in CV screening and its risks come from learning your own history

The central problem.

A model trained on who you hired before will favour candidates resembling them. Where past hiring reflected bias, the system automates it and adds a veneer of objectivity, which makes it harder to challenge.

2. Screen against stated requirements, not against a model of success

The safer design.

Checking whether a candidate holds a required qualification, licence or specific experience is a rule you wrote and can defend. Ranking candidates by similarity to previous hires is not.

3. Write the requirements before you see any applications

Discipline that improves the whole process.

Essential and desirable criteria, agreed in advance. This improves screening quality regardless of any automation, and it is what makes an automated first pass defensible.

4. Never let a system reject without review

The line to hold.

Automated sorting, ranking or flagging is defensible. Automated rejection of a person is not, and in some jurisdictions decisions of this kind made solely by automated means carry specific legal obligations.

5. Be careful what the system reads

Proxies are the trap.

Names, addresses, dates, school names, employment gaps and photographs all correlate with characteristics you must not discriminate on. A system given these will use them, and the correlation is invisible in the output.

6. Keep a record of why each decision was made

Both protection and improvement.

Which criteria each candidate met or did not, and who decided. If a decision is ever challenged, this record is the defence, and its absence is a serious problem regardless of how fair the process was.

7. Watch the composition of the shortlist

The practical check.

Compare the applicant pool against the shortlist by any characteristic you can lawfully monitor. A shortlist that consistently narrows the pool in a particular direction is telling you something about the screening that no audit of the criteria will.

8. Tell candidates how their application is handled

Increasingly an obligation.

Whether automated processing is used, what it does, and how to ask for a human review. Transparency requirements around automated decision-making differ by jurisdiction and are tightening in several.

9. Use the time saved on the interviews

Where the value should land.

Screening is not where hiring quality comes from. If automation frees a day, spending it on better structured interviews and proper reference checks improves outcomes far more than screening precision ever will.

Employment law, discrimination law, data protection and rules on automated decision-making all apply here and all vary considerably by jurisdiction. This is an area where the compliance position should be confirmed before implementation rather than after a complaint.

Conclusion

Understand that a screening system learns your past hiring, and design around that.

Screen against written requirements rather than similarity to previous hires, agree the essential and desirable criteria before applications arrive, never allow an automated rejection without human review, keep proxies for protected characteristics out of what the system reads, record which criteria each decision turned on, compare your shortlist composition against the applicant pool, tell candidates how their application is processed and how to request human review, and spend the saved time on better interviews.