Are You Good at AI, or Just Using It?

TL;DR. Most AI training teaches tools. The skills that separate people who use AI from people who are good at it are judgement skills: verifying a claim against a source, editing output so it reads like you, and knowing what must never be pasted into a chat box. Only 35% of workers say they have the training they need — down from 45% in 2024 — and more than half say what's missing is practice on their own real tasks. This is the eight-item check, in the order the gaps actually bite.

An Excel trainer wrote something this month that stuck with me. People keep coming to him wanting to "get good at AI." His observation, after enough of these conversations to form a strong opinion: they want to become AI power users before they can reliably find a file they downloaded.

The same week, the front page of Hacker News carried a thread titled, plainly, "Are you good at AI, or just using it?"

And the two largest tech-community threads of the month weren't about models or benchmarks at all. One was called "AI;DR" — AI, Didn't Read. The other was "Don't paste the AI, please". Both were complaints about colleagues sending unedited AI output to other humans.

Put those together and the picture is unambiguous. The problem in most workplaces is not that people can't access AI. It's that access arrived years before judgement did.

The training gap is real, and it's getting worse

This isn't a vibe. Two independent 2026 surveys measure it:

• Only 35% of workers say they have the training and resources they need to use AI in their job — down from 45% in 2024, per Study.com's State of AI Jobs and Skills. More access, less readiness.
• In JFF's 2026 survey, workers in the most AI-exposed roles said the training they receive is generic rather than tied to the tasks AI is already doing in their jobs. More than half said practice on their real work is what's missing.

The second finding is the important one. It's not that training doesn't exist — it's everywhere. It's that almost none of it uses your actual work, so it teaches the tool and skips the judgement.

Meanwhile the reward for closing that gap keeps climbing. PwC's 2026 Global AI Jobs Barometer, across more than a billion job ads in 27 countries, puts the average wage premium for workers with AI skills at 62% — up from 57% in 2025 and 25% in 2024. Jobs requiring AI skills grew 69% against 9% for the market overall.

So: a widening skills premium, and a training pipeline that's teaching the wrong half of the skill.

The eight-item check

Here's the honest self-assessment. Answer yes or no, no partial credit. If you'd like to run it interactively and get it matched against your own profession's demanded skills, the AI Skills Gap Analyzer does exactly that.

They're ordered by how early the gap bites — verification and data hygiene fail loudest and soonest; tool fluency matters later than people expect.

1. Can you check an AI claim against a primary source before you use it?

The gap if not: you can't yet tell a confident answer from a correct one.

This is first because it's the one that ends careers rather than merely embarrassing people. Models produce fluent, plausible, specific text regardless of whether the underlying fact is real — and specificity reads as authority. A fabricated case citation, a subtly wrong dosage threshold, a regulation number that doesn't say what it's claimed to say: all of these arrive looking exactly like the correct version.

How to close it: take your next AI output and try to source three specific claims in it — not the general argument, the specifics. The numbers, the names, the citations. Whatever you can't source, cut. Do this five times and you'll develop a physical sense for which sentences need checking, which is the actual skill.

2. Can you edit AI output so it reads like you wrote it?

The gap if not: your output is recognisably unedited AI — which is what colleagues are objecting to, not AI use itself.

Read those two viral threads again. Nobody was complaining that their coworkers used a language model. They were complaining about receiving text the sender clearly hadn't read. The signal isn't "this was AI-assisted"; it's "you didn't care enough to look at it."

How to close it: rewrite the opening and closing lines yourself — they carry most of the voice. Cut every sentence that states something obvious to your reader. Delete anything you wouldn't say out loud to their face. If the result could have been written about anyone, by anyone, keep going.

3. Do you know which information must never go into an AI tool at your job?

The gap if not: you're relying on luck with client, patient, or company data.

This is the one where the consequences are other people's. Client records, patient details, credentials, unreleased financials, anything under NDA — the boundary needs to exist in your head before you're mid-task and in a hurry.

How to close it: find your employer's AI policy. If there isn't one — and there often isn't — write your own list of categories you will never paste, and hold to it. A written line you can point to is worth far more than a general intention to be careful.

4. Have you turned a task you repeat weekly into a reusable prompt?

The gap if not: you're retyping context every time instead of compounding what already worked.

Most people's AI use never leaves the ad-hoc stage. They get a good result, feel pleased, and then start from scratch next week. All the value evaporates between sessions.

How to close it: pick the document you write most often. Next time you get a genuinely good result, save the prompt that produced it — with the parts to swap out marked clearly. That's the difference between using AI and building leverage with it.

5. Do you know which surface to use — chat, an agent, or something in your browser?

The gap if not: you're using one surface for jobs another does far better.

This one is newly important. Claude alone is now four distinct surfaces — chat, Cowork, Claude Code, and a Chrome extension — and they're genuinely different tools. Doing a multi-step task across your files by pasting things into a chat window one at a time is like emailing yourself files instead of using a shared drive: it works, but you're doing the tool's job.

How to close it: run the surface chooser against a task you actually did this week, and notice whether you'd have picked differently.

6. Can you tell when the model is confidently wrong inside your own field?

The gap if not: your domain expertise isn't being applied as a filter.

This is the skill you already have and aren't using. You have years of pattern-matching in your field. Most people switch it off when reading AI output, because fluent text triggers a reading mode rather than an evaluating one.

How to close it: ask it something you know cold. Not to catch it out — to learn its failure shape in your domain. You'll find it's rarely wildly wrong; it's subtly wrong in a consistent way. That shape is what to watch for everywhere else.

7. Can you get the format you want in one or two tries?

The gap if not: you're negotiating with the output instead of specifying it.

The tell is a long back-and-forth — "shorter," "more formal," "add a section on X" — where each round fixes one thing and disturbs another.

How to close it: say the structure up front. Sections, length, audience, what to leave out. Paste an example of what good looks like. Specification beats iteration almost every time, and it's faster.

8. Could you explain to a colleague why a particular output needed changing?

The gap if not: you can feel that something's off but not name it — which makes it hard to fix reliably and impossible to teach.

This is last because it's the one that turns personal skill into something an organisation values. "It felt wrong so I rewrote it" is a beginner's answer. "The second paragraph asserted a number that isn't in the source, and the register was too casual for a regulator" is the answer of someone who can be trusted with the work.

How to close it: each time you rewrite AI output, write one sentence about what was wrong with it. Five times, and the pattern becomes obvious.

How to read your score

• 0–3: Using it. You're getting output, but mostly taking what you're given. This is where everyone starts — and where the credibility-damaging mistakes happen. Close one or two gaps before learning another tool.
• 4–6: Getting good. You have the habits that matter most and catch the obvious failures. The remaining gaps are what separate someone who uses AI from someone others ask for help.
• 7–8: Good at it. You verify, you edit, you know what not to paste. Your leverage now comes from applying this to the specific tasks your profession is measured on.

Why this order matters

Almost every AI course starts with prompt engineering. That's roughly item four here, and it's deliberate.

If you can't verify an output, better prompts get you to a polished wrong answer faster. If you can't edit into your own voice, better prompts produce more fluent generic text. If you don't know what not to paste in, better prompts increase the volume of your exposure.

Prompting is a multiplier. Multipliers are only worth having once the thing they multiply is sound.

The good news is that judgement skills transfer. A new model ships every few months and the tool knowledge decays with it. The ability to check a claim, edit into your own register, and know what shouldn't leave the building doesn't decay at all — it's the same skill whether you're using Claude, ChatGPT, or whatever replaces both.

Where to start this week

Pick the lowest-numbered gap you answered "not yet" to, and do only that one for a week.

If you'd like it prioritised against your own field — because a gap that's also on your profession's demand list is worth closing first — the AI Skills Gap Analyzer runs this check and maps it onto the skills employers are actually asking your profession for. It's free, instant, and nothing you type leaves your browser.

And if item two is your gap, the specific mechanics are here: how to review an AI draft before you send it.