Edition #42: The Friction You Keep: AI, Judgment and Effort

AI removes the effort that builds judgment. A behavioral scientist on the say-do gap, why experience is slowing down, and what to protect.

By Mykel Salomon · 2026-10-17

<h3>AI makes thinking feel optional. Judgment is what we lose first.</h3><hr><h3>🔓 Opening Reflection</h3>There is a version of the future being sold right now that asks you to stop thinking.Not in those words. The words are always freed up, streamlined, so you can focus on what matters. But the mechanism underneath is the same every time. Remove the effort. Shorten the distance between the question and the answer.And let us be honest about what happens next, because we already know.Given an easy path, we take it. Every time. Not because we are lazy, and this is the part my guest this week insisted on, but because that is how the machine between your ears was built.<a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 ember-view" href="https://www.linkedin.com/in/will-leach/">William Leach</a> has spent his career on one question: why do people actually do what they do, usually without knowing why themselves. He ran insights at PepsiCo, wrote Marketing to Mind States, teaches behavioral science at SMU and Texas A&M, and now builds behavioral AI.As a species, we want to spend as little effort as possible. It is a feature, not a flaw. Your mind pushes the vast majority of your choices down into habit and automaticity precisely so the resources are there when something actually matters.Which is a beautiful system, and it has one catastrophic vulnerability.It assumes effort is still available somewhere. It assumes that when the big decision arrives, you will have the judgment to meet it.But nobody is issued judgment at birth. It gets manufactured, slowly, out of effort and consequence. You tried something. It failed. You learned where your own thinking breaks.So here is the thesis of this edition, and it is not the one I expected to be writing.The problem is not that AI thinks for us. The problem is that effort is where judgment used to get made, and we are removing effort from everything without deciding which effort was load-bearing.The question is no longer whether you will take the easy path. You will.The question is which friction you keep on purpose.<hr><h3>🔦 Signal of the Week</h3>The people least equipped to push back are the most likely to accept the answer.Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers and found what they called a critical thinking paradox.Higher confidence in the AI was associated with less critical thinking. Higher confidence in your own expertise was associated with more.People who know their domain argue with the machine. People who do not, take the output.Which means the protection is distributed exactly backwards. The person with twenty years of pattern recognition is the one interrogating the answer. The twenty-two year old who needs the scrutiny most is the one least able to apply it.The machines are good at the middle and bad at the edges.Researchers built a benchmark called ConsumerSimBench to test whether large language models can reconstruct how real consumers actually react. They used 1,553 real social media topics and 23,122 audited criteria across 13 frontier models.The best performer covered 47.8% of real reaction criteria. It missed more than half.And the finding underneath that one is sharper still: performance on technical benchmarks did not transfer. Models that score brilliantly on reasoning tests trailed badly here. Being good at logic turns out to be close to unrelated to understanding people.The platforms have started pricing authenticity.In late July, LinkedIn added a "seems like AI slop" reporting button, deployed new classifiers to reduce machine-written content in recommendations from outside your network, and began privately flagging posters whose content reads as inauthentic.In the same move, LinkedIn removed its "enhance your post" AI writing feature and replaced it with a proofreading tool that does not change your voice.That distinction is the whole argument in one product decision.<hr><h3>🌍 In the Wild</h3><h3>1. The easy path is not a weakness. It is the design.</h3>We talk about cognitive shortcuts as a failure of discipline. He talks about them as efficiency, and he is right.You drank a Coca-Cola once. It delivered. You should never have to run a cost-benefit analysis on a Coca-Cola again. Settle it, automate it, move on, and keep your attention for the things that actually turn on it.Then he said the line I keep coming back to:"The number one decision you make every day is to not decide."You look away from the awkward conversation walking toward you. You close the book. You leave the email unread. None of those register as decisions, and all of them are.So we arrive at this moment already built to prefer the path with less resistance on it. And into that, we introduce the first tool in human history engineered to remove resistance from every step of thinking. Not some steps. Every step.Ask a question, get a plausible answer, in seconds, phrased in a voice it has learned to match to yours.There is no friction anywhere in that experience. That is the product. That is what people are paying for.And it works on us so well because it runs with our nature instead of against it.<h3>2. Two of the numbers in this conversation did not survive</h3>Ego depletion. The idea that willpower is a finite resource, that self-control drains like a fuel tank, and that this is why you make bad choices when you are tired.In 2021, researchers ran a multisite preregistered test across 36 laboratories with 3,531 participants. The preregistered result was nonsignificant, with an effect size of d = 0.06. Bayesian analysis found the data roughly four times more likely under the null hypothesis. An earlier multi-lab replication had already failed to find it.The fuel tank is not there.35,000 decisions a day. This number is everywhere. It is in keynotes and leadership books and it has the satisfying specificity of something somebody counted.Nobody counted it. Trace the citation chain and it runs through a news article, into a magazine column, into a university blog post, which attributes it to "various internet sources" and cites a book containing no original research and no decision count. The nearest real figure is roughly 226 food-related decisions per day, from a researcher whose work has since been subject to multiple retractions over data integrity.Now, before this reads as a takedown, hold on, because here is what actually matters.The observation is real. The mechanism was wrong.You genuinely do make worse choices at the grocery store when you are tired and hungry. Will is right about that, and so is everyone who has ever filled a cart at 9pm. What the research says now is that this is not a fuel gauge hitting empty. The better-supported account, from Inzlicht and Schmeichel's process model, is that after sustained effort your motivation and attention shift. You become less inclined to keep exerting control and more drawn toward whatever is immediately rewarding. Fatigue turns out to be several processes rather than one resource.The practical advice does not change at all. Rest still matters. Do not make consequential decisions depleted. Take the vacation.What changes is that the story we have been telling ourselves about why was wrong for twenty years, and almost nobody noticed, because checking was effortful and the story was satisfying and it was already in the slide deck.That is the thesis of this edition, happening live, in a conversation between two people who think about this for a living.We all run on cached answers. Including about the science of why we run on cached answers.I am not exempt from this and neither is Will, who has forgotten more behavioral science than I will ever learn. The difference between us and the next generation comes down to having been wrong enough times in public to develop a flinch.That flinch is judgment. And it only comes from having paid for being wrong.<h3>3. The say-do gap has moved inside the machine</h3>This is the part of the conversation with the most commercial weight, and the fewest people talking about it.Will's entire discipline exists because of one problem. He calls it the say-do gap.You tell a researcher you want to eat healthier. You mean it. You are not lying. Then on Friday night you go to Five Guys. Market research exists to measure the distance between those two things, because the second one is what your business actually depends on.Here is what has happened over roughly the last four years, and sharply in the last eighteen months. Budget has moved out of talking to customers and into technology and tech stacks. The economics are brutal and obvious. A hundred one-hour interviews is a hundred hours of someone's expensive time, plus the analysis. A ChatGPT subscription is twenty dollars and will produce a persona that answers your questions immediately.And the output looks good. That is the trap. It reads like a customer.But as Will put it, these models were trained on what people said. The say-do gap is baked into every sentence of that training data. We built machines on the written record of human self-report, which is the precise artifact his profession exists to distrust.The validation research backs him up, and it is specific about the failure mode.Synthetic respondents match distributions reasonably well. They fail at individuals. A Dutch panel study found that persona outputs matched aggregate distributions but not individual predictions, and the correlations between variables did not match real data. Multiple studies find the same cluster of problems: reduced variance, inflated positivity toward established brands, and systematic reproduction of the mainstream reaction while missing the critical, the niche, and the culturally specific.Read that last part again, because it is the whole thing.They are accurate about the middle and wrong about the edges.Which brings us back to a phrase from Edition #32, and one I used in this conversation without realizing how literal it was about to become: AI pushes everything toward the middle of the curve. The Microsoft researchers have a term for it. They call it mechanized convergence.The middle is not where brands are built. Nobody ever became irreplaceable by being statistically typical. The edges are the entire asset, and the edges are exactly what the machine cannot see.<h3>4. Judgment is the receipt for a consequence you actually paid</h3>Will's son started his junior year of high school the morning we recorded. He talked about him the way you talk about something you are genuinely worried about and trying not to catastrophize.His worry is narrower and harder than unemployability.In ten years that kid will sit in a room where an AI produces a confident, plausible, well-structured recommendation, and his entire value in that room will be the ability to say no, I would take this a different direction.That ability has one source. You tried something, it did not work, and you learned where your own thinking fails.Will's phrase for what is coming: "a lot of people who can run models but cannot push back on models."He does not think experience disappears. He thinks the cycle gets longer, which is worse in a way, because it is invisible. You still make mistakes. You just make fewer of them, later, with more cushioning in between. His blunt version: you may be twenty-five and still operating with the judgment of a sixteen year old, because the consequences that would have built the difference were absorbed by a tool.Put that next to the Microsoft and Carnegie Mellon finding and it gets worse. Confidence in the AI predicts less scrutiny. The people with the thinnest experience are the ones least likely to challenge the output and most likely to need to.The researchers connect this to Lisanne Bainbridge's ironies of automation, which is forty years old and describes exactly this: automate a task and the operator's skill at that task decays, right up until the moment the automation fails and you need the skill you no longer have.There is one more layer in the conversation worth naming. Will pointed out that those of us who grew up in the nineties were conditioned to believe computers do not make mistakes. That conditioning has not been updated, and it is doing real damage, because these systems do not fail in ways that look like failure. They fail by being plausible. Ask Grok and ask ChatGPT and you will be walked down two different roads, neither of which announces itself as a road.Will's advice to his son was the most practical thing in the hour. Go start something small. Go fail at it fast. Do it without AI, deliberately, so that you own the mistake and the mistake owns you back.Not because the tool is bad. Because the scar is the asset.<h3>5. Execution was the moat. It is gone.</h3>For most of commercial history the advantage went to whoever could execute more, faster, cheaper. More people on the team. More ads out the door. More campaigns tested.Will was blunt about where that stands: he can generate as many ads as the eighteen-person agency down the street, assuming he has the credits. Capacity is not a moat when capacity is a subscription.So what is left?His answer: do you understand people better than your competitors do. Not demographically. Why they actually take the steps they take.The example he used will stay with me. Somebody tells you they are eating salad to be healthier. Fine, accurate, useless. Go deeper and you find they are eating salad because they looked in the mirror and got frightened, and what they actually want is to be alive for their daughter's wedding. That person is negotiating with mortality, not optimizing nutrition.Those are two completely different customers. They need two completely different conversations. And only one of those truths shows up in a survey.This is where Will's distinction between behavioral AI and generic AI lands. Not a different model. The same pattern-finding power, pointed through the social sciences, looking for motivation and identity and fear rather than sentiment and keywords. He used to spend thirty hours interviewing and thirty hours reading transcripts line by line hunting for those tells. The machine does that in a minute and finds patterns he admits he could never see. That is real augmentation, and he is unambiguous that his own thinking is sharper now than it was before these tools, because they push back on him.The barrier is incentives, not capability.Organizations are still measuring output. Will described the performance review logic of about a year ago, and I have seen versions of it myself: Michael used thirty-seven thousand tokens last month, so he must be valuable.We built the dashboard for the old moat and then wondered why everyone optimized for the old moat.And now the market is starting to price the other thing. LinkedIn removed the feature that writes in your voice and kept the one that fixes your commas. Anthropic began embedding an imperceptible, machine-readable watermark in text from models released on or after August 2, 2026, driven partly by the EU AI Act transparency rules that took effect the same day. (Disclosure, since it is the honest thing to do here: Anthropic makes the assistant I use to research and build this newsletter.)Will's concern was that this becomes a badge of shame. I think the nuance cuts differently, and it is worth knowing.The watermark indicates that Claude had a hand in the text. Not that Claude wrote it. Ask it to proofread one paragraph and the trace is there. Heavily rewrite or translate machine-written text and the trace can disappear.So as a signal it is close to backwards from what people want it for. The careful writer who used AI to catch a typo gets marked. The person who generated the whole thing and ran it through a rewriter may not.Which means the watermark will not save anyone from having to judge the work on its merits. Nothing will. The tell was never going to be technical.People can feel the middle of the curve. They have always been able to feel it. That is the actual moat, and it was never for sale.<hr><h3>💬 The Big Question</h3>Which friction in your work is building you, and which is just friction?Because they are not the same, and almost nobody has done the sorting. We are removing both at the same speed and only one of them was ever costing us anything.And the harder version, the one for anyone leading people:If you removed every difficult task from a junior person's first three years, what exactly would they be at the end of it?<hr><h3>🧠 A Small Exercise</h3>1. Name one thing you will keep doing the hard way.One. Not a philosophy, a specific task. Writing the first draft. Doing the math before you check it. Making the call instead of sending the message. Pick it, commit to ninety days, and notice whether anything in your thinking sharpens.2. Answer first, then ask.Before you accept an AI output on anything that matters, write your own answer in one sentence. Then compare. The gap between the two is a live read on your own judgment, and it is the only measurement of it you are going to get.3. Check one number you have been repeating.You have one. A statistic in your deck, a study you cite, a fact you are sure about because you have said it so many times. Trace it to the source. I did it twice this week and neither one survived.4. If you lead a team, stop measuring usage.Tokens consumed, tools adopted, prompts written. None of it tells you anything. Measure the thing that will actually matter in five years: can this person look at a confident, well-formatted, entirely plausible recommendation and tell you why it is wrong?If nobody on the team can do that, you do not have an AI capability. You have an AI dependency.<hr><h3>🎙️ This Week on The Human Protocol Podcast</h3>Behavioral Science, the Say-Do Gap, and What AI Is Quietly Taking From Us Guest: <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 ember-view" href="https://www.linkedin.com/in/will-leach/">William Leach</a> , author of Marketing to Mind States, former Head of Insights at PepsiCo, behavioral science faculty at SMU and Texas A&MWill has spent his career on the question of why people do what they do, usually without knowing why themselves. He now builds behavioral AI, applying social science to the models rather than just pointing models at data.In this episode, you'll learn:<ul><li>The three forces shaping the decisions you never notice making, and why the biggest one is the decision to not decide</li><li>Why taking the easy path is efficiency rather than laziness, and what that means now that a tool exists to remove every obstacle</li><li>The say-do gap, why it is the central problem of his profession, and how it ended up trained into the models companies now use instead of talking to customers</li><li>Why he would no longer tell a young person to learn to code, and the two things he would tell them instead</li><li>The difference between behavioral AI and generic AI, and why understanding a customer's fear beats understanding their demographics</li><li>What happens to judgment when the cycle of consequence gets longer, and the advice he is giving his own son</li><li>The ethical line he drew, the client he took anyway, and what happened when they ended up in front of Congress</li><li>Why execution is no longer a moat, and what he believes replaces it</li></ul>🎧 Available now on <a target="_self" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 WeLUwaPMyhLZZutVswjmZsEjuBomKCc " href="https://open.spotify.com/episode/1BrJ6Jhsf63v9bV06mG0Fn?si=yDi3-NvkR6K68eby_3Jd8g">Spotify</a>, <a target="_self" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 WeLUwaPMyhLZZutVswjmZsEjuBomKCc " href="https://podcasts.apple.com/us/podcast/ai-is-killing-our-ability-to-think-will-leach/id1843402023?i=1000796302922">Apple Podcasts</a>, and <a target="_self" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 WeLUwaPMyhLZZutVswjmZsEjuBomKCc " href="https://youtu.be/xVx_pjkDTBE?si=yCjSS5Mj19KUiJku">YouTube</a>.<img class="max-w-full h-auto rounded-md my-4" src="https://media.licdn.com/dms/image/v2/D4E12AQGGMfZcNJeLfQ/article-inline_image-shrink_1000_1488/B4EaFJyWYhKUAI-/0/1792239298346?e=1793836800&v=beta&t=XStmXQPEpeZHwqqAdUzgW-tyBWZZ7NRjtI4H04GP9qM" alt="Article content"><hr><h3>🧭 Thought to End the Week</h3>I asked Will what still surprises him after a career spent studying human behavior.He did not talk about bias or persuasion or anything from the textbooks.He said what surprises him is how quickly we stop moving toward our own goals.People know what they want. They can name it. They have named it for years. And then they take actions that go directly against it, or, more often, the thing that genuinely saddens him, they take no action at all. Someone carries the goal of starting their own business for forty years and never starts it.Not because they forgot. Because starting is hard, and not starting is easy, and the mind was built to prefer easy.Sit with that for a second, because it predates every piece of technology in this conversation. Long before there was anything to outsource our thinking to, we were already finding ways not to do the hard thing.AI did not create that tendency. It exponentialized it.And this is why I do not think the answer is to use less of it. I use it every day. My own thinking is sharper for it, and so is Will's, and he said so plainly. A tool that argues with you makes you better. A tool that agrees with you makes you smaller. The same product does both, and the only variable is what you bring to it.So the discipline is not abstinence. It is selection.Decide, deliberately, which difficulty you are keeping. The draft you will write yourself. The conversation you will have in person. The problem you will sit inside of for an afternoon before you ask anything to solve it. The number you will check even though you are fairly sure.Keep those. Automate the rest with a clear conscience.Because effort is not the cost of the work. Effort is where the person doing the work gets made.And a brain you never use is a brain you eventually lose.The code may change. And it will.But our humanity is the constant.— Mykel<hr><h3>📚 Sources</h3><ol><li>Vohs et al., "A Multisite Preregistered Paradigmatic Test of the Ego-Depletion Effect," Psychological Science, 2021 (36 labs, N = 3,531, d = 0.06, nonsignificant) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://pubmed.ncbi.nlm.nih.gov/34520296/">https://pubmed.ncbi.nlm.nih.gov/34520296/</a></li><li>Hagger et al., "A Multilab Preregistered Replication of the Ego-Depletion Effect," Perspectives on Psychological Science, 2016 <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://static1.squarespace.com/static/550b09eae4b0147d03eda40d/t/57a0c246ff7c508925c3ddb7/1470153300492/A-multi-lab-pre-registered-replication-of-the-ego-depletion-effect.pdf">https://static1.squarespace.com/static/550b09eae4b0147d03eda40d/t/57a0c246ff7c508925c3ddb7/1470153300492/A-multi-lab-pre-registered-replication-of-the-ego-depletion-effect.pdf</a></li><li>Inzlicht and Schmeichel, "What Is Ego Depletion? Toward a Mechanistic Revision of the Resource Model of Self-Control," Perspectives on Psychological Science, 2012 (the process model: motivation and attention shifts) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://www.researchgate.net/publication/257835702_What_Is_Ego_Depletion_Toward_a_Mechanistic_Revision_of_the_Resource_Model_of_Self-Control">https://www.researchgate.net/publication/257835702_What_Is_Ego_Depletion_Toward_a_Mechanistic_Revision_of_the_Resource_Model_of_Self-Control</a></li><li>"Do We Really Make 35,000 Decisions a Day?" Did The Research (full citation trace showing no primary source) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://didtheresearch.substack.com/p/do-we-really-make-35000-decisions">https://didtheresearch.substack.com/p/do-we-really-make-35000-decisions</a></li><li>Lee et al., "The Impact of Generative AI on Critical Thinking," Microsoft Research and Carnegie Mellon, 2025 (319 knowledge workers; the critical thinking paradox) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/">https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/</a></li><li>R&D World summary of the Microsoft and CMU findings, including mechanized convergence and Bainbridge's ironies of automation <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://www.rdworldonline.com/microsoft-and-cmu-genai-may-undercut-critical-thinking-but-theres-an-antidote/">https://www.rdworldonline.com/microsoft-and-cmu-genai-may-undercut-critical-thinking-but-theres-an-antidote/</a></li><li>"Can LLMs Think Like Consumers? Benchmarking Crowd-Level Reaction Reconstruction with ConsumerSimBench," arXiv (13 models, 1,553 topics, 23,122 criteria, best coverage 47.8%) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://arxiv.org/abs/2605.17079">https://arxiv.org/abs/2605.17079</a></li><li>Summary of 2026 synthetic respondent validation studies, including the NIM and LISS panel findings on homogeneity, positivity bias and individual prediction failure. Note: published by a synthetic research vendor, so read with that interest in mind; the underlying studies are independent <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://personia.ai/blog/what-2026-validation-studies-agree-on-synthetic-research/">https://personia.ai/blog/what-2026-validation-studies-agree-on-synthetic-research/</a></li><li>TechCrunch, "LinkedIn adds a button to report AI-generated 'slop'," July 30, 2026 (classifiers, private dashboard flags, removal of the "enhance your post" feature) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://techcrunch.com/2026/07/30/linkedin-adds-a-button-to-report-ai-generated-slop/">https://techcrunch.com/2026/07/30/linkedin-adds-a-button-to-report-ai-generated-slop/</a></li><li>Fortune, "Anthropic plans to add an invisible mark to AI text," August 11, 2026 (models released on or after August 2, 2026; EU AI Act transparency alignment; limitations) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://fortune.com/2026/08/11/anthropic-claude-watermark-ai-text-police-ai-slop/">https://fortune.com/2026/08/11/anthropic-claude-watermark-ai-text-police-ai-slop/</a></li><li>Will Leach, Marketing to Mind States</li></ol>