AI Fluency Is Four Competencies, Not a Tool List

What does AI fluency actually mean?

Working with AI effectively, efficiently, ethically and safely, split into four named competencies. The AI Fluency Framework, written by Rick Dakan at Ringling College of Art and Design and Joseph Feller at University College Cork and turned into courses with Anthropic, calls them Delegation, Description, Discernment and Diligence ¹. That is the version with authorship you can check, which is more than most circulating definitions offer.

In the framework's own words:

• Delegation is "creative vision and selection of the right AI tools and techniques to realize that vision" ¹. In practice it is deciding which part of the job stays yours.
• Description is "effectively describing a vision and/or tasks to prompt useful AI behaviors and outputs" ¹. What context went in before the ask.
• Discernment is "accurately assessing the usefulness of AI outputs" ¹. Including the case where the output is fluent and wrong.
• Diligence is "taking responsibility and vouching for final products created using AI" ¹. Whether the person stands behind what shipped and is straight about how it was made.

Two honest caveats before anyone builds a bar on it. The framework is a definition, not a measurement: it publishes no levels, no scoring anchors, no norms and no evidence that people who rate well on it perform better. And it is authored by two academics and co-produced with Anthropic, so calling it Anthropic's framework overstates the relationship. The four are worked through one at a time in what each of the four Ds looks like in real work.

Who decides what counts as fluent?

You do, and that is not a gap somebody forgot to fill. No board accredits AI fluency and no exam confers it. Even the EU AI Act, which does write an AI literacy duty into law, does not require an employer to measure or test anyone's knowledge of AI ². So the bar exists wherever an employer writes it down, which is usually a job description nobody re-read.

The European Commission's guidance on that duty, revised in 2026, states that no certificate is needed, that no specific training is mandated, and that keeping an internal record of what was done is the suggested practice ². It cuts both ways: nobody has to test staff, and no vendor's badge confers compliance either. That is guidance rather than statute, it has already been rewritten once, and anything touching an EU obligation belongs in front of counsel rather than in a posting.

The credential market fills the vacuum with things that sound general and are not. Microsoft's Azure AI Fundamentals is the credential most often listed as an entry-level AI qualification; its own exam page addresses someone "at the beginning of your career in AI solution development" and states that the candidate needs knowledge of Python syntax and familiarity with Azure resources, with the majority of the exam weight sitting on implementing in one vendor's platform ³. It is a real credential for what it covers. It is not a statement that its holder works well with AI on your material.

Which leaves the employer holding the pen. The requirement is only worth writing if it can be tested, which is the whole of what fluency has to mean on a job description.

Why can't you see fluency in the finished document?

Because the decisions happened before the file existed and left nothing behind in it. The brief shapes what came back, so a deliverable is partial evidence about one competency. It says nothing about what was handed over, what was thrown out, or whether anybody opened the source behind the claim the recommendation rests on. That silence is why the market fills up with proxies.

Every proxy in circulation is cheaper than the thing it stands in for, and each fails in a way worth naming out loud. A certificate says a course was completed. A tool list says somebody can type four product names. Years of AI experience says when they started. A self-rating says how confident they are, which is a genuine measurement of something and not of skill. None of the four survives the question a hiring manager actually has, which is whether this person will notice when the answer is wrong.

The cost of guessing wrong has been measured directly. In a pre-registered field experiment with 758 Boston Consulting Group consultants, on one task deliberately chosen to sit outside the model's capability, consultants using GPT-4 were 19 percentage points less likely to reach the correct answer: 84.5% of the control group got it right, against 60% and 70% in the two AI conditions ⁴. One task, one sample, a 2023 model, and the capability line moves with every release. What travels is that consultants at a top firm could not tell which side of the model's capability the task sat on.

What stays readable is the process. What a candidate kept for themselves, what an assistant produced that they refused, and what they opened in order to check a number are all things a person can describe or a session can record, and none of them can be inferred from a clean document. What being good at using AI looks like in a round turns that into questions somebody can ask on Monday.

Write the bar as four behaviors before you write the word

Turn the four competencies into four sentences about your own work, then delete the word fluency from the posting. A requirement naming a behavior can be tested and defended. A requirement naming a level cannot, because there is no scale underneath it and nothing to appeal to when two interviewers disagree about whether a candidate cleared it.

1. Delegation. "Decides which parts of the analysis to do without a model, and can say why." Satisfied when a candidate names one step they kept deliberately, with a reason attached to the work rather than to a policy. 2. Description. "Briefs a model with the context a colleague would need to do the task." Satisfied when the constraint, the audience and the source material go in before the request for output, rather than dribbling in over a correction loop. 3. Discernment. "Catches a confident claim that is wrong in this field." Satisfied by finding the error in your own material and naming what it was checked against. Not satisfied by saying it felt off. 4. Diligence. "Vouches for what ships and is straight about how it was made." Satisfied by a claim withdrawn or re-scoped rather than shipped unverified.

Who writes those four lines matters more than the wording does. The person currently doing the job knows which step should never be handed to a model and what a wrong answer looks like in this material; a recruiter working from a template does not, and a template is where a bare AI fluency requirement comes from in the first place. Twenty minutes with the incumbent produces four lines nobody has to argue about in the debrief, and the same twenty minutes produces the material the exercise gets built from.

The fourth item is the one most often dropped from summaries of the framework, and it is the expensive one. The framework's own sub-competency closest to an employer's concern covers verifying and vouching for AI-assisted output, including fact-checking and validating claims ¹. That is a disposition rather than a technique, which is why it belongs in the requisition and not in the training plan.

See what gets scored >