Do I Still Need to Practice If AI Does the Work?

Do I still need to practice reading and listening if AI can do it for me?

Yes, because the tool takes the task and you keep the skill, or you stop keeping it. In Bastani and colleagues' randomised field experiment, students with unrestricted GPT-4 did 48% better during practice and 17% worse on the exam once the tool was gone. Delegating a task is fine. Delegating the practice is what costs you.

That distinction is the whole answer, and it is not the one most people are handed. The loud version of this conversation says you are making yourself dumber. The honest version is narrower: you are not damaged, you are unpractised, and unpractised is a fixable description.

What did the field experiment actually find?

It tested three groups of high school maths students: one with unrestricted GPT-4, one with a guided tutor built to make them do the reasoning, and a control group that had only textbooks. The unguided group did 48% better on practice problems, then scored 17% worse than control on the exam without the tool.

Two details matter here. It was randomised, and the test that mattered happened after access was removed, so it measures what stayed in the students rather than what the tool could produce. The paper reports one more finding that is easy to skim past: the students who performed worse did not believe they had learned less.

Confidence and performance came apart, and only one of those was visible to them. That is the trap. If the work feels smooth and the outcome feels fine, nothing inside you raises an alarm about a skill going unfed.

Why is it so easy not to notice?

Because offloading is a cost calculation rather than a character flaw. Risko and Gilbert describe it as a trade: people move a task to an external tool whenever the tool is available and cheaper than doing it internally. Nothing in that arithmetic asks your permission, and it gets cheaper every month.

Expectation does the rest. In the well-known Google effects work, Sparrow, Liu and Wegner found that when people believe information will stay available, they remember where to find it instead of the content. That was a study of search engines in 2011. Read it with AI in the room and it describes most knowledge work now: your memory keeps the index, not the file.

You can hear the same switch when people describe meetings. In a widely read thread about AI notetakers, one user put it plainly: "I find that after each call I have almost forgotten what the call was about or any of the details." The notes were taken. The memory did not happen. The transcript is doing the job attention used to do.

Does this happen outside school?

Yes, and the workplace version has been measured rather than argued. When AI polyp detection was introduced into routine practice, the detection rate on colonoscopies performed without it fell from 28.4% to 22.4%. Same doctors, same eyes, no intention of getting worse.

Budzyń and colleagues reported that drop in The Lancet Gastroenterology and Hepatology: an observational study across a multicentre group of endoscopists. Those were experienced professionals doing a trained task they were paid to stay sharp at. What changed was not ability. It was whether the skill got exercised on the cases where it used to be.

What does practice actually keep?

Three things, and none of them is a claim that your brain is being upgraded. First, the encoding you do yourself, which is what turns a page into something usable later. Second, recall without the tool. Third, the judgement to catch the tool being wrong, which needs a view of your own to compare against.

There is measurement here too, and it needs a caveat. Kosmyna and colleagues' 2025 MIT Media Lab study recorded brain activity during essay writing and found the weakest engagement among the people writing with an LLM. The paper is small and says nothing about permanent damage. What it does support is modest and still useful: during the task itself, the assisted brain does less, and the effect shows up that afternoon rather than years later.

So where should AI sit in the workflow?

After the attempt, not instead of it. Read the section, close the tab, write the two sentences you think the argument made, then open the summary and compare. The gap between your version and the summary is the real information about what you skipped, and it is free.

Used that way, a summary stops being a substitute for the reading and becomes a correction signal, which is what it was always good at. The same move works for listening: ninety seconds of recap out loud after a call, before you open the transcript. If the recap is thin, you learned something about your attention rather than about the meeting.

Keep the practice, hand over the paperwork. If summaries already feel like the reading, that feeling has its own explanation in why reading an AI summary feels like learning, and the habit that repairs it is set out in how to think without AI.

That gap is what a daily workout is for. Absorb runs three short rounds, one each for reading, listening and analysing, and returns Absorb's three separate scores inside a 10 to 12 minute session. You are not trying to reverse anything. You are trying to see which channel you stopped exercising, because the blind spot you cannot feel is the one that needs a number.