Hiring: where automation is most regulated

Why this one is different

Everywhere else in this course, a mistake costs money or embarrassment. In hiring it can cost somebody a job on a basis that is unlawful, and the rules reflect that.

The direction of regulation is clear even where the detail is not. New York City requires bias audits and candidate notification for automated employment decision tools. The EU AI Act classifies AI used in recruitment and selection as high-risk, with obligations attached. Existing discrimination law in most countries already applies regardless of whether a machine was involved, and an employer cannot delegate responsibility to a supplier.

For a business hiring two people a year, the practical conclusion is simple: use it to prepare, never to decide.

Where the bias comes from

Not malice, and not usually anything visible in the output.

A model trained on decades of hiring text has absorbed which words appear in which contexts. Historic patterns come with it. This is measurable: studies of language models in hiring contexts have repeatedly found different outcomes for identical CVs differing only in name, and differences in how the same role is described.

The subtler version is proxy discrimination. A system never shown a protected characteristic can still infer it from a school, a postcode, a career gap, a language, a club. Removing the name does not remove the signal, and a small business has no realistic way to audit for this.

That is the mechanism, and it is why the answer is not a better prompt.

Where it helps

Writing the advert. Ask for a version in plain language, then ask it to flag wording that might narrow the field unnecessarily: jargon, unnecessary qualifications, phrasing that skews. “Five years’ experience” when three would do excludes people for no reason.

Building a structured interview. Same questions, same order, every candidate, with a written scoring guide. Structured interviews predict performance considerably better than unstructured conversations, and almost no small business runs them because writing one takes an afternoon. This is the highest-value use in the lesson.

Writing the scoring guide. What a weak, adequate and strong answer looks like for each question. This is what makes scoring consistent rather than a memory of how you felt.

Preparing your own thinking. What does this role actually need? What would competence look like in week one?

Replying to unsuccessful candidates. Politely, promptly, individually. Most small businesses do not reply, and drafting removes the excuse.

Where not to use it

Sifting applications automatically. Not because it cannot rank them, but because a ranking you cannot explain is a decision you cannot defend, and in some jurisdictions it triggers audit and notice obligations you will not have met.

Scoring candidates. The scoring guide is a tool for you. The score is yours.

Analysing video interviews. Inferring traits from face or voice is contested, sits close to prohibited practices in some regimes, and does not have the evidence base its vendors imply.

Anything about the person rather than the work. Social media checks, personality inference, background scraping.

Keep records

Whether or not it applies to you, keep the job description, the questions, the scores and a short note of why each decision was made.

If a decision is ever questioned, contemporaneous notes are what you have. They also make you a better hirer, because writing down a reason is a check on whether you have one.

The plain summary

The candidate is a person and the decision is yours. Use the tool on the paperwork around the decision (the advert, the questions, the guide, the replies) and keep it away from the judgement about the human being.

Candidates are using it too

Assume every application you receive was written with help, because most of them were. Two consequences.

Covering letters have converged: they are all fluent, all structured the same way, and they discriminate between candidates far less than they used to. Weighting them heavily is now a way of measuring who used a better prompt.

So move the weight to things that are harder to generate: a short piece of real work, a structured conversation, a practical task, references. If a written exercise still matters to you, make it about something specific to your business that no general tool could produce well.

Do not try to detect AI writing. The detectors are unreliable in both directions, they misfire on people writing in a second language, and accusing a candidate wrongly is a serious matter.

One question worth asking every candidate

How do you use AI tools in this kind of work, and where do you not trust them?

There is no right answer, and the answer is informative either way. Somebody who has never used them may be fine. Somebody who describes checking, and names a specific failure they have hit, is telling you they think about their work, which is most of what you were trying to find out.