Turn the AI Line in Your Job Posting Into One Answerable Question

What does AI fluency required actually screen for?

On its own, close to nothing. A posting line filters only when someone can tell whether an application meets it, and the phrase has no observable act attached, so anyone who has opened a chat window can claim it in good faith. What it screens for in practice is willingness to write the phrase back at you, which selects on confidence and on familiarity with the vocabulary. Nothing in that touches skill.

The underlying framework is clearer about its own limits than the postings quoting it are. The AI Fluency Framework, written by Rick Dakan of Ringling College and Joseph Feller of University College Cork and turned into courses with Anthropic, defines the term as the ability to work effectively, efficiently, ethically and safely within emerging modalities of human-AI interaction, and names four competencies: Delegation, Description, Discernment and Diligence ¹. It is a definitional framework rather than a measurement. No levels, no anchors, no norms, no validated instrument, and nothing in it supports a judgment about who is or is not fluent.

The application-form version fails for a second reason. In a study of 288 Taiwanese teachers who took both a self-report questionnaire and a knowledge test built on one shared AI-literacy framework of concept, use, evaluation and ethics, correlations between the objective and self-reported factors ran from 0.07 to 0.24 ². The population is Taiwanese teachers, and no job candidate was involved. A weak correlation means only that the two instruments do not measure the same thing. It is not proof that people overrate themselves. Either way, a checkbox measures the box.

Write the requirement backwards from the question

Start at the interview and work back to the posting. Decide the question you will actually ask, decide what a passing answer has to contain, and write the line only after both exist. A requirement you cannot reach from a question is a preference, and preferences belong under nice to have or nowhere. Twenty minutes per requirement, once, and the posting stops making promises the loop cannot keep.

Four common lines and what they become when the rewrite is done properly:

1. AI fluency required becomes *can take a model's first draft of a client-facing document and say what in it should not go out, and why*. The question hands over the draft. 2. Experience with AI tools becomes *has shipped work where a model produced the first pass, and can say what changed between that pass and what shipped*. The question asks for one example and one specific change. 3. Strong prompt engineering skills becomes *can restate a vague request precisely enough that a wrong answer would be recognisable as wrong*. The question is a vague request. 4. Comfortable with AI becomes nothing. There is no act underneath it, so the line comes out.

Notice what the rewrite does to the funnel. The original phrases are self-descriptions and every applicant clears them, so the line adds volume without adding information. The rewritten versions name something a person either does or does not do, which is what makes them screenable early and scoreable late. Writing an AI-skills requirement that is not legally vague goes through the same rewrite from the compliance side rather than the interviewing side.

What does a passing answer contain?

Three things, and they fit on an index card before the round starts: what the candidate would not send, why the work makes that the wrong thing to send, and the specific check they would run to settle it. All three is a pass. The first two is partly there and worth a follow-up. An answer that only praises the tool or only distrusts it has not addressed the question.

The framework's own vocabulary lines up with that third item more closely than the popular summaries suggest. Under Diligence it lists Deployment Diligence, described as taking responsibility for verifying and vouching for AI-assisted outputs, including thorough fact-checking, testing for accuracy, and validating claims ³. That sub-competency is the one most often dropped when the four Ds get compressed into a slide, and it is the one an employer is actually buying. Twelve sub-competencies sit under the four, set against three modalities, and the framework calls that its current version rather than a fixed count.

Watch for the two answers that look like passes and are not. The first is fluent process description with no artifact in it: a candidate walks through a careful-sounding routine that never touches a specific claim, a specific source or a specific consequence. The second is blanket refusal, which sounds rigorous and tells you nothing about judgment, because refusing everything requires no judgment at all. Both get the same follow-up: what in this draft, specifically, and how would you find out.

Write the answer down while the candidate gives it. The four lines worth putting on an AI scorecard covers the recording, which is where most of this evidence gets lost. If you would rather test the requirement outside the interview entirely, what AI fluency means on a job description and how to test for it sets out the assessment-side options.

Should the line stay in the posting at all?

Only if it survives the rewrite. An untestable requirement still does something in the funnel, and what it does is invisible to you: whoever reads the line and closes the tab never appears in your pipeline, so the one effect the line has is the one you cannot inspect. If the rewritten version is something the role genuinely needs, publish that. If it is not, delete the line and lose nothing.

The legal frame points the same way, and it is older than any of this. Griggs v. Duke Power, decided in 1971, is the origin of adverse-impact analysis in US hiring: a unanimous Supreme Court held that Title VII proscribes not only overt discrimination but also practices that are fair in form but discriminatory in operation, that the touchstone is business necessity, and that a test must measure the person for the job and not the person in the abstract ⁴. That case concerned a diploma requirement and two aptitude tests at one power plant, its burden-shifting framework was narrowed in 1989 and reworked by Congress in 1991, and it says nothing about software. Quote it for the principle and check the operative burdens with counsel.

Applied here, the principle is practical rather than defensive. A requirement stated as an act, tested the same way for every applicant, and tied to something the job does on a Tuesday is easier to justify and easier to run than a phrase that means whatever the reader assumes. The posting and the interview end up describing the same thing, which is the only real fix for a requirement nobody checks.

See how it works >