Use It and Lose it

Hey Everyone - Hope you had a great weekend. There was a study published last August that resurfaced this week with new coverage, and I think it is the most important piece of evidence in the entire AI cognitive-cost literature so far. It is short, it is clinical, and it ends an argument I have been hedging on for two months. Let me walk you through it.

This week:

Let's dive in.

In a multicentre study published last August in The Lancet and re-cited heavily this week in TIME, researchers across four Polish endoscopy centers tracked what happened to expert clinicians' ability to detect precancerous growths after they had been using AI assistance for routine colonoscopies. The participants were not novices. Each had performed more than 2,000 procedures. The measurement was simple. After three months of routine AI use, how good were they at the same job when the AI was taken away.

The answer is the cleanest piece of evidence we have on this entire question. Unassisted adenoma detection fell from 28.4 percent to 22.4 percent. That is a roughly 20 percent relative drop in a patient-relevant outcome, in three months, among seasoned experts. The authors describe it as the first study to show a measurable negative impact of regular AI use on a healthcare professional's ability to complete a real clinical task.

The Lancet paper is the clinical proof. A new MIT study from this month is the psychological one. Researchers tracked 67 participants over four weeks as they judged the credibility of news headline-and-image pairs. With AI assistance, accuracy jumped 21 percent. Once the AI was removed, unassisted performance on new items fell 15.3 percentage points below where they had started before the experiment. The sharpest finding is the metacognitive one. About a quarter of participants reported feeling they were getting better at the task even as their measured skill eroded. The atrophy was invisible from the inside.

Put the two studies together and the picture changes the frame the newsletter has been carrying for two months.

I had been writing the deskilling story as a slow thing. Years of low-friction AI use quietly hollowing out a muscle that you would notice if you tried hard enough. That story is wrong on both halves. The Lancet data says the timeline is months, not years. The MIT data says you will not notice, because confidence and competence decouple as the skill fades. By the time you sit down without the tool and realize you cannot do the work, the loss has already happened, and your sense of how good you are has been padding the gap.

This changes what the hedge actually is. The hedge is not "use AI less." That was always a half-measure. The hedge is to put deliberate unassisted reps on the calendar for the specific skills you most want to keep. Treat the unassisted work the way an athlete treats a maintenance lift. It is not optional. It is the price of keeping the muscle.

The honest list for most readers in 2026 is short. Three to five skills, max, that you would lose the most by losing. Writing in your own voice. Reading something hard and forming an original take on it. Doing math on paper. Reading a contract. Coaching a person through a problem without a script. Diagnosing a system. Whatever they are for you, name them, and put them on a recurring schedule where the tool stays closed. The Lancet study suggests the minimum dose is weekly, not monthly. The MIT study suggests the diagnostic value is at least as important as the practice. You will not feel the loss before the rep tells you it is there.

The boiling frog has a temperature now. It is three months.

The thing I keep coming back to in my own work is that testing beats theorizing every time. I have spent the last few weeks running a steady stream of A/B tests on my websites and newsletters, and the surprises keep coming. The headline I was sure would win underperforms. The opt-in placement I was about to remove turns out to convert. The subject line I almost did not send hits the highest open rate of the quarter. None of it was predictable from the outside. All of it was learnable from one experiment.

The reason this works is that the only real way to know is to test, and the key to making testing actually pay is volume. One test tells you almost nothing. Ten tests start to draw a curve. A hundred tests turn into a personal market-research system, and the system compounds because each test sharpens the next hypothesis.

The version of this that I am most thankful for is that the business gives me the volume. I have enough surface area now (multiple newsletters, multiple landing pages, multiple paid funnels) to run two or three meaningful experiments a week without anything breaking. That is the unfair advantage of running an operation at this scale.

The meta-lesson, and the one I keep applying outside work, is that opinions are cheap and data is expensive. The cheapest way to make the right call on almost anything is to set up the smallest test that would change your mind, run it, and then act on the result. The people who try to reason their way to the answer always lose to the people who let the next experiment tell them.

Pick the skills you most want to keep, and put unassisted practice for each of them on a recurring schedule. That is the whole skill. The Signal does the work of explaining why it matters. The mechanic is dumb on purpose.

Three things to make this real.

One. Name three to five skills, no more. Trying to maintain everything is the same as maintaining nothing. The skills worth naming are the ones whose loss would change the kind of work you are able to do or the kind of person you are. For me right now the list is: writing in my own voice, reading a long piece and forming an original take, holding a strategy conversation without a tool open, and one or two domain-specific items I will not bore you with.

Two. Schedule the reps. Weekly is the floor based on the Lancet timeline. For the highest-value skill, twice a week. Put the rep on the calendar with a real time block, not as a "when I get to it" item, because the whole point is to do the work when you would have reached for the tool.

Three. Treat the unassisted output as the diagnostic, not the deliverable. The MIT data is the key here. You cannot tell from the inside whether the skill is still there. The unassisted rep is the only test that gives you an honest read. Some of your reps will surprise you. Some of them will be worse than you expected. Use the gap as data, not as a verdict.

The chatbot economy is going to keep getting better at handing you finished work. The hedge is to keep producing the unfinished work yourself, on purpose, so the muscle does not quietly leave.