Hiring designers in the age of AI

What actually broke

Three things at once, and they compound. Application volume rose sharply because applying became nearly free. The average quality of the artefact rose, because generation improved. And the variance between artefacts collapsed, because everyone uses similar tools on similar templates.

Variance is where signal lives. When every portfolio is polished and every cover letter is well argued, the things that used to separate candidates stop separating them. The candidates haven't become more alike; the measuring instrument has stopped resolving the difference.

What didn't break

Judgment didn't get easier. Noticing that a brief is wrong, choosing what to cut, holding a position with an executive who has decided the answer, knowing what you don't know: none of these became cheaper to fake, because none of them are artefacts.

They're behaviours, and behaviours have to be observed rather than submitted.

The screening arms race can't be won

The obvious response is more screening: detection tools, stricter filters, extra rounds. It doesn't work. Detection has a false-positive rate too high to justify rejections, and it hits non-native speakers hardest. Extra rounds cost you the candidates with options, which is the wrong end of the distribution to lose.

Adding steps to a broken measurement makes the process longer without making it more accurate.

Rebuild around three properties

Whatever you use instead (an assessment, a structured interview, a working session), the useful ones share these.

• The context exists only inside the exercise, so it can't be researched beforehand.
• Conditions change mid-task, so a prepared answer stops being enough.
• The reasoning has to be defended live, so it has to be theirs.

Assess AI fluency as a skill, not as cheating

A designer who uses models well is more valuable than one who doesn't, in the same way that a designer who used Figma well was more valuable than one who stuck with Photoshop. Policing it instead of evaluating it is a mistake you'll regret within two years.

So ask whether they can tell when the model's output is wrong. Whether they used AI matters far less, and spotting bad output is, once again, judgment.

Frequently asked

Should you ban AI during the hiring process?

No. A designer who uses models well beats one who doesn't, and the ban is unenforceable anyway. Ask instead whether they can tell when the output is wrong. That's the skill, and it's the one thing the model can't supply on their behalf.

Do AI detection tools work on portfolios and cover letters?

Not well enough to reject anyone on. The false-positive rate is too high to survive a challenge, and it falls hardest on candidates writing in a second language, so you lose good people and inherit a fairness problem at the same time. Change what you measure instead of policing how it was written.

Applications have tripled. How do you screen that without adding rounds?

Put the filter at the front instead of the end. A short structured exercise, taken before any human reads anything, sorts on what you care about and scales with volume rather than against it. Extra interview rounds do the opposite: they cost the most, and they cost you the candidates who have other offers.

Is a take-home still worth running if candidates will use AI on it?

Only if you change what it asks for. A take-home that rewards a produced artefact now measures tool access. One where the context exists only inside the exercise, a condition changes partway through, and the reasoning has to be defended live still works, because none of that can be prepared in advance.