In this article
Before you start
This page sits in the behavioral interview guide. Answers to it are told as stories, so the STAR method helps.
Why interviewers ask this
Interviewers ask about AI fluency because employers now treat it as a requirement and have to judge it in conversation. A 2026 survey of 1,928 hiring leaders, reported by Recruiting News Network, found that 95% of U.S. organizations have made AI fluency a hiring requirement. The same report says 45% of U.S. employers set the bar at basic tool awareness.
Some companies are stricter. Indeed Lead reports that Zapier assesses AI skills in 100% of candidates, sorting them into four categories: unacceptable, capable, adoptive or transformative.
A weak answer names tools. A strong answer shows judgment: what you gave to AI, what you kept, how you checked it and who answered for the result.
AI fluency from first principles
- AI fluency
- AI fluency is the habit of working with AI in ways that are effective, efficient, ethical and safe. Anthropic's framework splits it into four competencies, so writing good prompts is one part of it, not the whole skill.
- Discernment
- Discernment is the competency of evaluating AI output critically: asking whether it is correct, complete and fit for its purpose before you rely on it or pass it on.
- Diligence
- Diligence is the competency of ensuring responsible AI collaboration: being honest about where AI was used and owning what you ship or share.
- Augmentation
- Augmentation is a way of working in which you and the AI collaborate as thinking and task partners, as opposed to the AI running a single instruction you gave it.
In plain words
Picture a very fast new assistant who knows a lot, never tires and sometimes states wrong things with full confidence. You would decide what to give them, explain it well, read what they hand back and stay responsible for the final work. The analogy breaks in one place. A human assistant tells you when they are unsure. An AI model often does not.
The four Ds are Anthropic's framework, not a standard that interviewers score against, so use them to organize your story rather than naming them as if the interviewer expects them.
Step-by-step description
- Decide what goes to AI and what stays with you.
- Say clearly what you want.
- Check the output before you use it.
- Own the result and be honest about AI's role.
- Delegation is deciding what work to do with AI and what to do yourself.
- Description is communicating clearly with AI.
- Discernment evaluates the output critically before you use it.
- Diligence ensures responsible AI collaboration.
In one precise sentence: Anthropic's lesson defines the four Ds as Delegation, Description, Discernment and Diligence, and says they apply across three ways of working with AI: automation, augmentation and agency.
Automation means the AI executes specific tasks on your instructions. Augmentation is the partnership defined above. Agency means you guide the AI to work independently on your behalf. An interview answer usually describes augmentation or automation.
Step 1: Delegation is deciding what goes to AI and what stays with you
Delegation is deciding what work to do with AI and what to do yourself. It comes first in Anthropic's list of the four Ds. Input: a task with several parts. Output: a split.
A good split has reasons. You hand over work that is routine, easy to check and low in risk, such as a first draft or a summary. You keep work that needs private context, judgment about people or accountability for the outcome. This also corrects a common belief: fluency does not mean using AI for everything. A candidate who says "I used it for every step" shows less judgment than one who says "I kept the audience choice for myself, because it needed customer knowledge." In the story, name the parts of the task and say which side each went to and why.
Step 2: Description is telling the AI clearly what you want
Description is communicating clearly with AI systems. Prompting sits here and nowhere else, which is why prompt tricks alone make a thin answer. Input: your goal. Output: an instruction the AI can act on.
Anthropic's exercise on a marketing-email project asks about the tone you want and the criteria for success. In an answer, say what you told the AI: the goal, the audience, the tone, an example of good work and what "done" looks like. One sentence on what you changed after the first reply shows the loop between Description and Discernment, which Anthropic teaches as a loop, not a straight line. You now have two steps: what went to the AI, and how you said it. The next one is where levels separate.
Step 3: Discernment is checking the output before you use it
Discernment evaluates the output critically before you use it. Input: the AI's draft. Output: a decision to use, fix or discard it. Say what you checked, how, and what you found.
This step is rare in real use. Anthropic's AI Fluency Index (February 2026) studied 9,830 conversations from one week of January 2026. Iteration and refinement appeared in 85.7% of the conversations, but checking facts that matter appeared in only 8.7% of the conversations, and questioning the AI's reasoning in 15.8%. Clarifying the goal appeared in 51.1% of conversations, giving examples in 41.1% and spotting missing context in 20.3%. Conversations that iterated showed reasoning-questioning in 17.9% against 3.2% for those that did not, about 5.6 times as often.1
Finished-looking output made checking rarer. With artifacts such as documents or code, fact-checking fell from 9.2% to 5.5% and spotting missing context from 21% to 15.8%.1 Anthropic labels these findings correlational, from Claude.ai users in a single week who probably skew toward early adopters. They are not population norms, so treat them as evidence of what is uncommon, and so what stands out.
Text description of the diagram
- Iteration and refinement appeared in 85.7% of the conversations.
- Checking facts that matter appeared in only 8.7% of the conversations.
Step 4: Diligence is owning the result and being honest about AI's role
Diligence ensures responsible AI collaboration. Input: the checked work. Output: something you can stand behind. In practice it means being open about which parts were AI-drafted, thinking through what happens if the output is shared or wrong, and taking ownership.
Anthropic's framework defines 24 specific behaviors, but only 11 are directly observable in chats on Claude.ai or Claude Code, so the other 13 cannot be seen in conversation logs.1 They include being honest about AI's role and considering the consequences of sharing AI output. An interview is one of the few places those behaviors can be shown. Say who you told, what you would not paste into a public tool and who answered for the result. Four steps make one story: what went over, how you said it, how you checked it, and who owned it.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →The Staff-level answer
Pick a real task, ideally with a result you can describe. The sample below uses a hypothetical email campaign; it is an illustrative scenario, so replace it with your own.
How do you use AI in your work, and how do you know you're using it well?
Take our last customer email campaign. For delegation, I let AI draft variants of the subject line and body, but I kept the audience and the offer, because those need our customer knowledge. For description, I gave it the goal, the tone and two past emails as examples. For discernment, I never sent a draft unread: I checked every claim and price against our product sheet and asked it to explain its choices, which caught an invented discount. For diligence, I told my manager which parts were AI-drafted and I owned what went out. Next time I'd write the checklist before prompting.
What makes it Staff-level is the mix: a boundary you drew, a check you ran, a failure it caught and an owner named. The spoken structure is delegation, description, discernment, diligence, then one lesson. It is a template to adapt, not a script to paste: interviewers can tell when a story is not yours.
The rubric
| Dimension | Weak answer | Strong answer | Staff answer |
|---|---|---|---|
| Delegation | "I use AI for everything." | Splits the task and says why each part went where it did. | Names what you deliberately kept, and the risk behind that choice. |
| Description | Names a tool or a "prompt trick". | States the goal, audience, tone and an example given. | Shows how the instruction changed after the first draft. |
| Discernment | "I read it over." | Names a specific check and what it found. | Checks reasoning and missing context, not only facts, and has a concrete catch. |
| Diligence | Not mentioned. | Says who knew AI was used and who owned the result. | Describes what you would not share with AI and why. |
| Communication | Vague, no real task. | One real task, told in a clear order. | Ends with what you would change next time. |
Three variants of this question
"Walk me through how you use AI in your work."
Use the four Ds on one task, as in the answer above. Keep to about a minute and give the task, not a list of tools.
"Tell me about a time AI got something wrong."
Describe the situation, how you noticed (a check, a second source, a gap you spotted), what you did and what you now do differently. This is Discernment told as a STAR story and it shows more judgment than a story where nothing went wrong.
"How do you decide when not to use AI?"
This is Delegation from the other side. Name the kinds of work you keep, such as decisions about people, anything needing private data and tasks where you cannot check the result, and say why.
Common mistakes
- "I'm very good at prompting." (wrong) Description is one of four competencies. (right) "Here is the task, here is what I handed over, here is how I checked it."
- "I use AI for everything." (wrong) It signals no Delegation. (right) "I used it for the draft and kept the audience decision."
- "The output looked polished, so I used it." (wrong) Anthropic's data link finished-looking output with less checking. (right) "I verified the claims first."
- "It saved me a lot of time." (wrong when the number is invented) A made-up figure is easy to challenge. (right) "It cut the first draft, and I spent the savings on checking." Label any figure you give as an estimate.
- A list of tool names with no story. (wrong) Tool awareness is the lowest bar. (right) One task, told in order.
For the general method behind these stories, see how to answer interview questions and the wider set of common interview questions.
Follow-up questions
How is AI fluency different from prompt engineering?
Prompt engineering is the Description step. AI fluency also covers deciding what to delegate, evaluating the result and using AI responsibly, so a candidate who only talks about prompts answers one quarter of the question.
What if I do not use AI much at work?
Say so honestly, then use the nearest real example: a personal project, a study task or a small experiment, and walk it through the four Ds. Inventing a workflow is riskier than a small true one, and the same honesty is part of Diligence. You can also connect it to why they should hire you.
Do non-technical roles get asked about AI fluency?
Yes. Indeed's examples include asking a marketer how they use AI to generate content and asking a talent leader about hiring processes rebuilt with AI. Anthropic's course exercises cover marketing, research and creative work.
How do you show responsibility without naming a tool?
Describe the controls, not the product: what you would not paste into an AI model, who reviews the work before it goes out and how you disclose AI-drafted parts. A mistake you caught also pairs well with an honest weakness answer, and the self-introduction question is a natural place to mention how you work.
- Answer with one real task and the four Ds: Delegation, Description, Discernment, Diligence.
- Prompting is only the Description step; tool names alone are the lowest bar.
- Describe a specific check you ran and what it found, because checking is the rarest behavior in Anthropic's sample.
- Say what you kept for yourself and who owned the result.
- Label any invented figure as an estimate, and use a true story of your own.
Sources
- The 4D Framework (AI Fluency: Framework & Foundations, Lesson 3) — Anthropic (Claude Academy), 26 Oct 2026
- Anthropic Education Report: The AI Fluency Index — Anthropic, 23 Feb 2026
- How to Evaluate AI Fluency in Your Team (and in new hires) — Indeed Lead, 10 Mar 2026
- AI fluency is your top hiring requirement, so why are bad hires still happening — Recruiting News Network (reporting a TestGorilla survey), 13 Jul 2026
