AI in HR and recruitment: fair, lawful and open to challenge

Cue 1 A recruitment funnel with AI at each stage: advert, sift, video score, ranking

EMMA 00:00 George, our applicant system scores every CV and gives me a shortlist. The supplier calls it decision support. Am I fine?

GEORGE 00:08 Depends what you do with the shortlist. Let me tell you what the Information Commissioner's Office found. On 31 March 2026 they published a report on automated decision making in recruitment, after working with more than thirty employers.

EMMA 00:23 And?

GEORGE 00:24 Employers underestimated when tools were making the decision. Many tools described as supporting a human showed no meaningful human involvement in practice.

Cue 2 A report dated 31 March 2026 with its main finding: no meaningful human involvement

EMMA 00:32 I do look at the shortlist.

GEORGE 00:35 Do you ever change it? Do you ever pull someone back in that it rejected?

EMMA 00:41 ...No. I've never seen the ones it rejected.

GEORGE 00:44 Then for those candidates, the tool decided. If nobody ever overrules it, it's deciding. Count how often a reviewer changes its outcome. A figure of zero tells you the answer.

Cue 3 A reviewer with training, information, time and authority, beside a rubber stamp crossed out

EMMA 00:56 What would proper human involvement look like?

GEORGE 00:59 The ICO's draft guidance says active, not tokenistic. The reviewer has to be trained to understand the system's logic, outputs, limitations and risks. And they need the information, the time and the authority to reach a different outcome.

EMMA 01:14 Our IT lead configured it. Doesn't that count?

GEORGE 01:17 No. The guidance says someone who designed or built the system isn't meaningful human involvement. That happened before any real decision was made.

Cue 4 Four safeguards as four cards: told, representations, human intervention, contest

EMMA 01:26 So is a fully automated rejection illegal?

GEORGE 01:29 Not necessarily, since 5 February 2026. The law was amended. A solely automated decision with a significant effect is permitted on ordinary personal data, with safeguards. And rejecting a job application is a significant effect.

EMMA 01:43 Which safeguards?

GEORGE 01:44 Four. The person is told about the decision. They can make representations. They can obtain human intervention. And they can contest it.

Cue 5 A candidate notice in plain words with a route to a human review

EMMA 01:53 And if the tool uses health or ethnicity?

GEORGE 01:56 Special category data. Then solely automated decisions are generally prohibited, unless there's explicit consent or a narrow legal condition.

EMMA 02:03 What do I have to tell candidates?

GEORGE 02:06 Clearly, that automated decision making is used and how it works. Not a technical essay. And how to ask for a human review. Plus a data protection impact assessment where the risk is high, kept up to date.

EMMA 02:21 Is that guidance final?

GEORGE 02:23 No. It went out for consultation, which closed on 29 May 2026, so it may change. The four safeguards are in the legislation and they're in force now.

Cue 1 Nine protected characteristics and a scoring model labelled a criterion

EMMA 02:34 George, our supplier says the tool is unbiased because it never sees sex or ethnicity. Is that enough?

GEORGE 02:41 No, and it's the most common misunderstanding. Start with the law. The Equality Act protects nine characteristics, and it doesn't care whether a decision was made by a person or by software.

EMMA 02:54 So what's the risk with a tool?

GEORGE 02:57 Indirect discrimination. A criterion that looks neutral, applied to everyone, which puts people with a protected characteristic at a particular disadvantage. Unlawful unless you can justify it. And a scoring model is a criterion.

Cue 2 A CV with the name removed and three details still pointing to who the person is: a gap, a postcode, a year

EMMA 03:11 But it can't see the characteristic.

GEORGE 03:13 It doesn't need to. Other details stand in. A gap in employment can stand in for maternity or illness. A postcode for ethnicity. A graduation year for age.

EMMA 03:24 That's uncomfortable.

GEORGE 03:25 It gets more so. A tool trained on your past hiring learns your past preferences. If previous hires were mostly of one kind, it favours that kind, and presents it as a score.

Cue 3 A funnel showing who applies and who passes each automated stage, with a gap marked at one stage

EMMA 03:38 So how would I know?

GEORGE 03:40 Look. Compare who applies with who's passed through at each automated stage, by the characteristics you lawfully collect for monitoring. A large unexplained gap at one stage tells you where to look.

EMMA 03:53 Doesn't collecting that data make it worse?

GEORGE 03:56 Not if you keep it separate. The people and the tool making decisions never see it. Do it before go live, using past applications, and regularly afterwards.

Cue 4 A candidate asking for an adjustment before a timed test and a video interview, and a human route

EMMA 04:07 Anything simpler?

GEORGE 04:07 Yes. Every so often have a person review a sample of the candidates the tool rejected. It's the cheapest test there is.

EMMA 04:16 What about disabled candidates?

GEORGE 04:18 You must make reasonable adjustments. And automated stages create new disadvantages. A timed online test for someone with dyslexia. A video interview scored on speech or eye contact, for someone who's deaf, who stammers, or who's autistic.

Cue 5 Six questions for a supplier, and a calendar marking 2 December 2027

EMMA 04:33 So I need a way round the tool.

GEORGE 04:36 A way to ask for an adjustment before the automated stage, clearly offered, and a human route. A process with no alternative can't meet the duty.

EMMA 04:46 And if it goes wrong, is it the supplier's problem?

GEORGE 04:50 Yours. The employer is liable for discrimination in its recruitment, whoever built the tool. The software chose is not a defence in a tribunal.

EMMA 05:00 What do I ask the supplier, then?

GEORGE 05:03 What it does at each stage and what data it uses. What it was trained on and how it was tested for bias, with the results. Whether it infers any characteristic. And what you can see about why a candidate got their score.

EMMA 05:20 We recruit in Ireland too.

GEORGE 05:22 Then note the EU AI Act. Recruitment AI is high risk there, with duties from 2 December 2027. And emotion recognition at work is already banned.