Using AI a Little Bit for Everything Is the Worst Way to Use It

A 2026 study split 912 students into groups by how they used AI. Which group do you think did worst?

The ones who dabbled, using AI a little bit for everything. They learned less, and they saved no time.

Have you touched AI twenty times today, with a line of an email here and a summary you skimmed there? Then that group is probably you, and the email still took an hour.

You feel busier and a little less sharp, and you couldn't say what the AI actually did for you.

Dr Philippa Hardman is a learning scientist who writes about AI and how people learn. She cites that study, and she has a name for what's going on.

First, though, let's hear from the dabblers, because their case is better than it looks.

The case for a little of everything

Most people try a new tool on small things first, because small things are cheap to get wrong. A bad two-line rewrite costs you nothing, but a bad report could cost you a client. Using AI little and often is how you learn what it's good at, before you trust it with anything that matters.

Small use feels safer, too, because you write the email, AI tidies one sentence, and the email is still yours.

The case against

Hardman calls this the cognitive offloading paradox. In plain words, you hand small bits of thinking to AI all day, and you lose the thinking without getting the time back. Your attention keeps jumping between doing the work and handing it over, so neither one gets your full focus.

You end up as tired as if you'd done the work, with only half the work to show for it.

That safe feeling is the problem. Handing over a whole job is a decision you can see, but twenty little hand-overs don't feel like decisions at all.

Nobody chooses to stop writing their own sentences. The habit slips away one line at a time.

And when AI has done a bit of everything, you can't remember which bits were yours, so none of the work gets read properly. You check less at the exact moment you should check more.

Meanwhile, the saved time never arrives. You save two minutes here and three there, and then every one gets eaten by the switching, as you open the chat, explain, read, judge, paste and fix.

So the savings are real, but too small to spend on anything.

The students who came out ahead in the study did the opposite. They used AI heavily for a few specific jobs, then spent the freed-up time on the parts that needed them. They beat the dabblers, and they beat the students who didn't use AI at all.

Whole jobs or none

I agree with Hardman that at work, a half-handed-over task is worse than one you hand over fully or one you keep. The hand-over itself is the cost.

Pay that cost once, for something big, or don't pay it. I'd rather you used AI hard on two jobs and left it alone for the rest.

Still, the dabblers keep one point. Small use is a fair way to learn the tool, as long as you treat it as a short trial period with an end date.

I'll also admit that one study of students isn't the whole of working life, and heavy use only wins when someone checks the work. Hand over a whole job and skip the check, and you've just found a faster way to be wrong.

What settles it for me is another line from Hardman, "judgement is the product".

That judging means deciding what the piece is for and spotting what's wrong with the draft. It means knowing which of three decent options is right for this reader, this week.

Finding your two jobs

For one week, write down every time you reach for AI, on paper or in a text file, and don't change anything you do. You might count around forty reaches, bunched around three or four jobs.

Then paste the list into AI with this prompt. A prompt is just the message you type to the AI.

Here is a list of every time I used AI this week. Group them into the jobs they belong to, for example "first drafts of client emails" or "summarising meeting notes". For each job, tell me how many times I reached for AI. Tell me whether you could do that job whole, from start to finish, if I gave you the inputs once. And tell me what I'd need to give you. Rank the jobs by how much of my week they'd free up. Ask me anything you need to know first.

Pick one or two jobs to hand over properly. Set them up once, for example with a Project, a Skill or a saved prompt, which are ways to save your instructions so you don't repeat them. Give AI the whole job and everything it needs, and let it run from start to finish.

How do you check the work? The same way you'd check a colleague's, by reading all of it, fixing what's wrong and testing whatever can be tested.

Everything else goes back to being your work, on purpose. Write your own two-line reply and do your own thinking before you ask, because that's the skill you'll need to judge the jobs you handed over.

Can't see any clear jobs in your list? Let AI interview you. Try "Ask me ten questions about my working week, one at a time, until you can tell me which three jobs you could do whole."

Some jobs shouldn't go to AI at all, however much time they'd save. So make a list of what you refuse to hand over to AI, and keep that list.

Then judge yourself by where the saved hour went. Say you hand over the report draft, then spend the time reading the draft hard and fixing the argument. That's the point.

If you spend the hour on twenty more small reaches, you've bought the paradox back.

AI does the whole job. You do the judging, with the hour AI gave you.

Credit. The cognitive offloading paradox and "judgement is the product" are Dr Philippa Hardman's ideas, from her newsletter pieces in April and May 2026. The 912-student study is the one she cites. Read her write-up for the detail.