AI Does Not Make You Dumber, Acceptance Does

Hey Everyone - Hope you had a great weekend. The cognitive cost research on AI use has been stacking up for months now, and most of it points in the same uncomfortable direction. Use it heavily, lose something. But a new APA study just put a wrinkle in that story that I think is the most useful finding in the entire pile, because it points at a hedge you can actually run, not just a warning to take to heart. Let me walk you through it.

This week:

Let's dive in.

For the past few months I have been writing variations of the same warning. AI use has a cognitive cost. The cost compounds quietly. By the time you notice the skill atrophy, the muscle is already weaker. A new APA paper just complicated that story in a way I think is more useful than any of the prior research, because it hands you a prescription instead of a warning.

Sarah Baldeo and her team at Middlesex University ran 1,923 US and Canadian workers through 10 simulated work tasks, all the kinds of work where judgment is the point: planning under incomplete information, interpreting ambiguous data, justifying strategic decisions. The participants used commercial AI tools however they normally would.

The headline finding is the one you would expect. 58 percent of participants agreed that AI "did most of the thinking," and those participants reported reduced confidence in their own reasoning and weaker ownership of the ideas they produced.

But the load-bearing finding is the split inside that data. The participants who actively modified, challenged, or rejected AI suggestions reported the opposite. Greater confidence. Stronger sense of authorship. A retained feel for the reasoning that produced the output. Same tool. Same tasks. Opposite outcome. The variable was not whether they used AI. It was the posture they used it from.

Baldeo's exact phrasing is the line worth carrying around. "The issue was not AI use itself but the degree of passive acceptance." Asserting oversight and "active judgement" appears to leave people feeling "more confident in their own reasoning."

That is the entire hedge, in one sentence.

This pairs cleanly with a University of Pennsylvania study referenced in the same coverage. When AI was programmed to give wrong answers in a decision task with 1,372 participants, people followed the wrong answer about 80 percent of the time. Only about 20 percent overruled the bad output. The Penn team called this "cognitive surrender." That is the default failure mode of using AI without a posture, and the Penn data tells you how widespread the default has become.

The mechanism under both findings is the same. When you accept AI output without modification, you have outsourced the judgment work to a system that has no judgment, only a probability distribution over plausible answers. The output looks like reasoning, but the reasoning never actually happened. Your brain notices this. Your confidence drops. Your sense of ownership drops. Both signals are accurate. You did not earn the output, so you do not own it, so you cannot stand behind it.

When you modify the AI output, the math reverses. The judgment is yours. The decisions about what to keep, what to cut, what to push back on, are yours. The output now has your fingerprints on it because you actually put them there. The Baldeo data is, on the most charitable read, evidence that the human cognitive system is correctly tagging passive output as not-yours and active output as yours.

The reframe this gives the whole AI-cost conversation is significant. The warning is not "use AI less." The warning is "use AI from a posture, not as a default." Two people can use the same chatbot for the same number of hours and come out in opposite places on confidence, ownership, and skill. The one who treats output as a starting point ends the week stronger. The one who treats output as the answer ends the week a little weaker, and a little weaker again next week, until the boil completes.

The hedge is not abstinence. The hedge is posture. I will go deeper on the practice in the Survival Skill section below. The short version is to stop opening the chat first. Write your own rough version of the answer, even if it is bullet points on a napkin, before you prompt the model. Your judgment is in the loop from token zero, and the Baldeo study suggests that single change is enough to flip the sign on the cognitive cost.

Six months ago, if you had asked me about starting an agency, I would have told you not to. Trade time for money. Hire employees. Burn out on client work. I think I was kinda wrong…

The wrinkle is what AI is doing to the math. In a world where speed, data, and adaptability are the things that win, the agency model has structural advantages I missed. Agencies focus on one skill, get really good at it, and adapt to market conditions faster than a product team can. Each client makes the agency smarter on the next one. That is a compounding loop SaaS does not have.

The other thing I missed is what is happening to SaaS itself. SaaS is being commoditized in real time. AI features are table stakes. Pricing is collapsing. The old workflow-lock-in moat is getting eaten by agents that move between tools. Most small SaaS bets I was making mentally six months ago look worse to me now than the equivalent service bet.

The personal data point is that agency-style work has been falling into my lap for the past few weeks, mostly around newsletters (distribution, monetization, content). The demand is real, the supply of people who actually understand the operational pieces is small, and every newsletter I run for a client makes my own newsletters better. The skill stack compounds across both sides of the desk.

I am not fully sold yet. The objections I had six months ago are still real. But the right move right now is to test it on purpose, take a few short engagements, and see what the unit economics look like. If it pencils out, great. If not, I will have sharpened infrastructure I would have to sharpen anyway and made short-term revenue in the process. The meta-lesson is that the assumptions I was most confident about are the ones I should be testing first.

The Signal hands us a clean prescription. Active use preserves your confidence and skill. Passive use erodes both. The skill this week is the cheapest, most upstream way I know to enforce active use: write your own version of the answer before you open the chat.

The mechanic is dumb on purpose. Before you prompt the model, write a rough draft of what you think the answer is. It does not have to be good or complete. Three bullet points on a sticky note, a paragraph in your notebook, a one-sentence hypothesis you would have committed to if the chat window did not exist. The point is that your judgment runs first.

Then, and only then, open the chat. Compare what comes back against what you already wrote. Notice where the model went somewhere you would not have. Notice where you were already right. Notice where you were wrong. Make the call about what to keep, what to discard, what to combine. The output that ships is now downstream of your own thinking, not a replacement for it.

This works because of the exact mechanism Baldeo's data points to. The cognitive cost of AI use comes from outsourcing the judgment step. Draft Before Prompt puts the judgment step back at the front of the workflow, where it has always belonged. The AI becomes a sparring partner instead of a substitute.

One example. Every section of this newsletter starts as a rough handwritten outline before I get anywhere near a chat window. The chat shows up at the editing stage, not the thinking stage. The newsletter is the work I am most committed to making mine, and Draft Before Prompt is how I protect it.

This is a small habit with a big compounding effect. The danger is not the tool, it is the posture. Draft Before Prompt is the cheapest posture change you can install, and you can run it on the very next thing you would have opened the chat for.