Should Your Job Post Require AI Experience, or Does It Screen Out the People You Want?

Should your job post require AI experience?

Require the judgment, not the tools. A line naming ChatGPT, Copilot or Claude filters for people who can name them, which is a fact about someone's browser history rather than about their work. A posting silent on AI hands the whole question to an interview round nobody has written. The requirement that works names an act the role performs with a model, in that occupation's own vocabulary.

Start with what the line does mechanically. A requirement in a posting is a selection procedure: the Uniform Guidelines define one as any measure or procedure used as a basis for an employment decision, and name work experience requirements, educational requirements and unscored application forms among them ¹. So the sentence you write does a test's job with none of a test's discipline. Nobody validated it, nobody scored it, and it runs on every applicant before you read a word.

Three versions of the same requirement, for a marketing role:

• Tool-named. "Proficiency with ChatGPT, Claude and Jasper required." Screens for a claim anyone can make and nothing in your process can check.
• Absent. No mention of AI at all. Screens for nothing, and moves the question to a round nobody has scripted, where every interviewer invents a different bar.
• Act-named. "Comfortable drafting positioning with an assistant, and used to opening the source behind a statistic before it reaches a deck." Screens for a habit, and gives the interviewer something to ask about.

The third costs ten more minutes to write and is the only one you can assess later. It also survives a tool change, which the first does not: named products age faster than the requisition does.

Before writing any of the three, check that the role belongs in this conversation at all. Which roles actually need AI skills right now is a triage on tasks rather than titles, and it removes most of the requisitions this question gets asked about.

What does "AI experience required" actually filter for?

Tool familiarity, which is the cheapest thing a candidate can acquire and the easiest thing to claim. A named product in a requirements list is a keyword, and keywords are what applicants write toward. You get resumes carrying every tool you listed, in roughly the order you listed them, and no information about whether the person ever refused one of the model's answers.

That refusal is the thing worth hiring for, and a tool list cannot see it. Two candidates used the same assistant on the same kind of task. One asked for the deliverable, got a confident draft, tidied the edges and shipped it. The other wrote down what would make the answer wrong before generating anything, demanded a source for the one claim the recommendation rested on, opened it, and cut a section because it was doing no work. Both are honestly experienced with ChatGPT, and the requirement line cannot tell them apart.

Tool names also date badly. A requisition reopened two quarters later still carries the product list from its first draft, and by then it is screening for a stack the team has stopped using. The act underneath, checking a generated figure against its source, did not change.

There is a downstream cost too. Ask for tools and candidates supply tools, which is how a nine-item product list ends up on a resume with nothing you can do about it. What to make of a candidate who lists their AI tools is a problem the posting created one step earlier.

What do you lose by saying nothing about AI?

The bar, and the consistency. A posting that never mentions AI produces a panel where one interviewer treats assistant use as a red flag, another treats it as table stakes, and a third never raises it. Candidates guess in the same way: some disclose, some hide it, and you end up comparing answers that were never written to the same question.

Three costs, in the order they arrive:

• Self-selection you did not intend. People who work this way and want to keep working this way read the posting for permission and do not find it. Silence is not neutral to them; it reads as a shop that has not thought about it.
• An unscripted round. The question surfaces in the loop regardless, usually in a manager's screen, and it gets asked differently every time. That is an unstructured selection procedure by another name, and it is where the bar quietly becomes whichever interviewer spoke last.
• No basis for the offer conversation. A candidate who learns after signing that most of the first-draft work now runs through an assistant has a fair complaint about the posting.

Naming it is not the same as gating on it. A line in the preferred block, written as an act, sets the expectation and costs you nobody. Writing AI skills into job requirements works the wording line by line; the decision here is only whether the subject appears at all.

Write the requirement in the occupation's own language

Take the occupation's published task list and write the requirement against two or three tasks on it. O*NET carries tasks for every SOC code, and the gap between two of them is the whole argument: financial and investment analysts have 26 task statements, including employing financial models and interpreting data on price, yield and investment risk ⁴, while market research analysts have 13, including translating complex findings into written text ⁵.

Run "AI experience" through each and it names unrelated behavior:

Same phrase, three different acts. A requirement written at the altitude of "AI experience" filters correctly in whichever occupation the writer had in mind, and filters for tool trivia everywhere else.

Usage data says the difference is real rather than stylistic. A Microsoft Research team classified 200,000 anonymized Copilot conversations against O*NET work activities and found the most common and most successful AI-assisted activities are information work (creating, processing and communicating information), with applicability cutting across sectors because most occupations have an information-work component ³. Broad reach, a different part of the job in each case.

So write two sentences rather than a skills line. The first names a task from the occupation's own list. The second names the act you would want to see performed on it. Setting a defensible bar for a specific role is the half-day version of the same exercise, and once it exists the requirement falls out of it as a small edit.

Does an AI requirement screen out the people you want?

Written as a credential, yes; written as an act, no. A credential line does its filtering before anyone applies. Harvard Business Review reports the widely cited figure that men apply when they meet 60% of a posting's listed qualifications while women apply only when they meet 100% ⁶, so a requirement nobody in your process can verify still removes the people who read it honestly.

It also has to survive a challenge. Where a selection procedure has a disparate impact, EEOC guidance says the employer must show it is job-related and consistent with business necessity: necessary to the safe and efficient performance of the job, and associated with the skills needed to perform it successfully ². "Three years of AI experience" has no defensible relationship to anything: the phrase names no skill, and the number is arithmetic on a technology younger than that in most workplaces. "Checks a generated figure against its source before it goes into a memo" does have one, and you can say what it is for.

Two adjustments keep the pool and the signal at the same time:

1. Move the requirement out of the posting and into the work. The posting states the act; one exercise in the loop checks it. Replacing the resume screen with a work sample is the fuller version of that trade, and it is the only version where the claim gets tested rather than counted. 2. Keep it preferred unless the role cannot be done without it. A required line filters at the application step, before anyone at your company has read a word. A preferred line sets the expectation and still lets in the candidate who picked this up six months ago and is better at it than your last hire.

The people you want sit on both sides of the line. Someone who has never opened an assistant but reflexively checks a number against its source learns the tool in a fortnight. Someone with three years of assistant use who has never once refused an answer will not acquire the habit because you named a product in the requisition.

See a sample report >