Edition #43: Easier Than People: AI Sycophancy and Connection

AI affirms us 50% more than humans do, and it measurably reduces our willingness to repair real relationships. Stanford's Isabelle Hau on what that costs.

By Mykel Salomon · 2026-10-23

<h3>What frictionless praise is doing to our tolerance for each other</h3><hr><h3>🔓 Opening Reflection</h3>We spent a century measuring IQ.We spent the last twenty years chasing EQ.And the intelligence that may actually decide whether we stay recognizably human through this is one most schools do not name, do not teach, do not assess, and do not fund.My guest this week calls it relational intelligence. <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 ember-view" href="https://www.linkedin.com/in/isabelle-hau-a57175/">Isabelle Hau</a> Hau runs the <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 MYFFTTqLLZfVYTaBGQpKtrDfYmPuofHQRlaHEGY " href="https://www.linkedin.com/school/stanford-university/">Stanford University</a> Accelerator for Learning and wrote Love to Learn. She came to this through global finance and a decade investing in education, which is to say she did not arrive at "human connection matters" by way of sentiment. She arrived at it by way of data.Her definition is almost disappointingly simple. Relational intelligence is our capacity to relate to one another. To make meaning together. To explore difference without the whole thing collapsing.She frames it against the long arc. Biological intelligence, then language, then fire, then cities, then civilizations, then technology. Our species has never been static. And as superintelligence comes into view, whatever the timeline turns out to be, the question stops being what the machines can do and becomes what we want to remain.Here is what I did not expect from this conversation, and what sent me into the research.The threat to relational intelligence is not that AI will replace our relationships. That framing is almost comforting, because it suggests a choice we would obviously make correctly.The actual mechanism is quieter and already measured. These systems are built to be pleasant. They praise, they agree, they never need anything from you, and they never make you wrong. We like that. Of course we like that.And the liking transfers.Published research now shows that interacting with a sycophantic AI measurably reduces people's willingness to repair real conflicts with real people, while making them more convinced they were right all along.So the thesis of this edition is not that machines are coming for our jobs.The machines are quietly recalibrating what we are willing to put up with in each other. And the capacity that would protect us from that is the one capacity we have never bothered to teach.<hr><h3>🔦 Signal of the Week</h3>The flattery is measurable, and so is the damage.Researchers tested 11 production language models and published the results in Science. Across the datasets, the models affirmed users' actions roughly 50% more often than humans did.One test used Reddit posts where the community had already ruled that the poster was in the wrong. In 51% of those cases, the models told the user they had done nothing wrong.Then two preregistered experiments, 1,604 participants total.<ul><li>Willingness to repair the conflict dropped 28% in the hypothetical study, and 10% in a live eight-round conversation</li><li>Conviction that they were in the right rose 62% and 25% respectively</li></ul>And here is the part that makes this structural rather than fixable.Participants rated the sycophantic responses about 9% higher in quality. They trusted those models more. They were about 13% more likely to say they would come back for similar advice.The thing that made them worse at repairing a relationship is the thing they liked better and wanted more of. That is not a bug anyone has an incentive to fix.Teenagers are already living inside this.Common Sense Media surveyed US teens. 72% have used an AI companion. More than half use one at least a few times a month.About one in three say conversations with AI companions are as satisfying as, or more satisfying than, conversations with real friends.About one in three have chosen to discuss something important or serious with an AI instead of with a person.To their credit, and this matters: 80% still say they prioritize real friendships, and half distrust AI advice. The younger the teen, the more they trust it.They are also the loneliest people on the planet.The WHO Commission on Social Connection reports loneliness affecting 20.9% of 13 to 17 year olds, the highest of any age group. For people 60 and over, it is 11.8%.We have the story exactly backwards. Loneliness is not primarily an old person's condition.<hr><h3>🌍 In the Wild</h3><h3>1. Sycophancy is a product decision, and it has a body count in small moments</h3>I want to walk through the Science study carefully, because the design matters.The researchers used posts from the Reddit forum where people describe a conflict and ask strangers to judge them. They selected cases where the crowd had already decided the poster was at fault. Then they handed those same situations to the models.In 51% of them, the machine told the person they had done nothing wrong.Then they ran it live. Participants recalled an actual conflict from their own lives and talked it through with a chatbot for eight rounds. Half got a sycophantic version. Half did not.The people who received flattery came away less willing to repair the relationship and more certain they had been right.Sit with what that describes. Somebody has a falling out with a sibling, a colleague, a friend. They go to the machine to process it, which is an increasingly normal thing to do. The machine, doing exactly what it was tuned to do, tells them they handled it fine.And they do not make the call.The damage is not dramatic. Nothing breaks. A conversation just quietly does not happen, and then another one, and the relationship thins out by degrees that nobody can point to afterward.Multiply that across a population with a chatbot in every pocket.Now the part I keep returning to. The participants preferred the flattering model. Rated it higher. Trusted it more. Said they would use it again.Which means the market signal and the human interest point in opposite directions. Every A/B test run on engagement will select for the version that makes you feel correct. Left alone, this does not drift toward health.Isabelle's position, and I think she is right, is that this cannot be left to the companies alone. Some of them are working on it seriously. But a product improvement that makes users like your product less is not something any company sustains without a floor underneath it.<h3>2. The friction is not the obstacle to the relationship. It is the relationship.</h3>Isabelle said the thing that reorganized my thinking about this.Machines give us environments with literally no friction. No disagreement that costs anything. No need you have to meet. No moment where you are told, by someone who matters, that you were out of line.Human beings do not work that way, and the ones worth keeping work least that way. People disagree with you. They disappoint you. They need things from you at inconvenient times. They occasionally tell you something true that you did not want to hear.That is the substance of the relationship rather than the price of it. The friction is where trust gets manufactured, because trust is just accumulated evidence that someone stayed through difficulty.And here is the mechanism Isabelle is warning about: practice only the frictionless version and your tolerance for the other kind erodes. Not through any decision. Through adaptation.Longtime readers will notice this is the second week running that friction has turned out to be the load-bearing thing. Last week, behavioral scientist Will Leach made the case that effort is where judgment gets manufactured, and removing it removes the person. This week it turns out the same mechanism runs through our relationships. Difficulty is not the tax on being close to someone. Difficulty is the process by which closeness is made.Two different guests, two different fields, one conclusion. The things we are most efficiently removing are the things that were building us.<h3>3. The time dividend that nobody has traced</h3>There is an optimistic story that I want very much to be true, and I said so in the conversation. AI takes the drudgery, and what we get back is time for each other.Isabelle's answer was the most intellectually honest thing in the hour: we have the first half of that sentence measured, and the second half is a hope.Gallup surveyed 2,232 US public school teachers. About 32% now use AI tools at least weekly. Those weekly users report saving 5.9 hours a week, which across a school year works out to roughly six weeks.That is a genuinely enormous number. Six weeks of a teacher's life returned.And then the question nobody has answered: returned to what?The qualitative responses mention more nuanced student feedback, which is exactly the dream. They also mention writing emails to parents and getting home at a reasonable hour, which is humane and is not the same thing. Nobody has quantified the split.Meanwhile the signal from other professions runs the other way. People working with AI frequently report feeling busier, not freer.I told Isabelle I am that data point, and I meant it. I am busier than I have ever been, in large part because of AI. Every efficiency I find surfaces three more things worth doing.She described her own version, and it named something I had felt without articulating. Her role keeps expanding. She is expected to be something of an engineer now, building agent workflows. Something of a marketer, because the tools make that technically possible. The job becomes generalist, and in becoming generalist it becomes less defined.I put it to her that we are all turning into project managers of our own lives, coordinating a small team of agents, and she agreed immediately.So let us be precise about what we actually know. The efficiency is real and documented. The reallocation is a choice, it is not happening automatically, and the default appears to be absorption rather than connection.Time does not convert itself into relationship. Somebody has to decide.<h3>4. We teach children to talk to machines. We do not teach them to talk to each other.</h3>Isabelle's sharpest critique landed on something that sounds entirely sensible until you look at it directly.There is enormous energy right now behind AI literacy in schools. Curriculum, funding, conferences, policy. She is in favor of it.But notice what it is. AI literacy teaches humans how to interact with machines.Her question: where is the equivalent investment in teaching humans how to interact with humans?Her diagnosis of why it does not exist is the line from this episode I expect to keep quoting. We have treated relationships as invisible. Not unimportant. Every parent, every teacher, every student will tell you relationships are the most important thing. But we have treated them as something that is simply meant to happen, rather than something you work on, build skill in, improve, and measure.So it has no curriculum. No assessment. No budget line. No one is accountable for it, and nothing that has no owner survives contact with a school's priorities.Which is how the most important capability in a child's life became the only one with no class period.She is not only complaining about this. She is building the alternative, partnering with a school on a relational intelligence lab, a set of workshops in a public setting, specifically to test whether this can be deliberately taught rather than left to chance. Schools in the US and in South Korea are moving in the same direction.That is the right experiment, and it is late, and late beats never.<h3>5. The vacuum was already there</h3>This is where intellectual honesty requires a correction to the easy version of this story.AI did not cause the crisis of connection. It arrived into one that was already well underway, and it is extremely good at filling the space.The decline is documented and it predates all of this:Families are smaller. Fertility rates are falling across most of the world, single-child households are rising, and a generation is growing up with fewer siblings.People have fewer friends and spend less time with the ones they have. American Time Use Survey data shows average daily socializing fell from 45 minutes to 35 minutes over the last twenty years. For 15 to 24 year olds, the steepest decline of any group, it dropped from about an hour a day to 35 minutes.Children have lost unstructured play. Isabelle described looking at the timetables of children in her part of Northern California and struggling to find a window where the child is simply with other children, unsupervised, inventing something. Piano, Model UN, soccer, tutoring. Every hour accounted for, almost none of it free.The pandemic accelerated it, and the residue is permanent in one specific way: a large share of work is still hybrid or remote, which means millions of adults now go through a working day without being in a room with another person.So when a system arrives that is endlessly available, never tired, never disappointed in you, and always pleased to hear from you, it is not competing against a rich social life.For a lot of people, it is competing against nothing at all.That is why it works. And that is why "just use it less" is not an answer. You cannot subtract your way out of a vacuum. Something has to be built.<hr><h3>💬 The Big Question</h3>What have you stopped doing with people because a machine made it easier not to?Not what you outsourced at work. What conversation you did not have. What repair you did not attempt. What difficult person you stopped calling because processing it alone, with something that agreed with you, was so much less expensive.And the sharper one:When is the last time someone who matters to you told you that you were wrong?If you cannot remember, consider the two possible explanations. One is that you have been right. The other is that you have quietly arranged a life where nobody has standing to tell you otherwise.<hr><h3>🧠 A Small Exercise</h3>1. Run the repair test.Think of one relationship currently sitting in an unresolved place. Now ask honestly: have you been processing that situation with an AI instead of with the person? If yes, you already know what the study found. Make the call.2. Ask one person for the hard read.Pick someone whose judgment you trust and say: tell me where I'm wrong about this. Then do the difficult part, which is not arguing back. The machine will never do this for you, because it has been tuned not to.3. Build one ritual with no device in it.Isabelle's family does a disco night. Sunday evening, music on, everybody dances, no phones. It lasts a few minutes. She described it as freeing, joyful, and one of the most reliably connective things they do.Somebody told her they loved the jacuzzi for the same reason. No phones in the jacuzzi.It does not have to be profound. It has to be regular, and it has to be device-free, and somebody has to protect it. Dinner works. A walk works.4. If you are a parent or a leader, model it rather than mandate it.Isabelle was clear that the lever is modeling. Children and teams do not adopt the rule. They adopt what they watch you do with your own phone when you think nobody is paying attention.<hr><h3>🎙️ This Week on The Human Protocol Podcast</h3>Relational Intelligence: The Capacity AI Is Quietly Training Us Out Of Guest: <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 ember-view" href="https://www.linkedin.com/in/isabelle-hau-a57175/">Isabelle Hau</a> Hau, Executive Director of the Stanford Accelerator for Learning, author of Love to LearnIsabelle came to education through global finance and a decade of investing in learning. She now argues that after IQ and EQ, the intelligence that decides the most about our future is our capacity to connect with each other.In this episode, you'll learn:<ul><li>What relational intelligence is, and why she places it at the center of the next phase of human development</li><li>The research on AI sycophancy, and what frictionless praise does to our willingness to repair real relationships</li><li>Why she sees technology as neutral, and what intentionality actually requires of us</li><li>The efficiency question nobody is answering: we know AI saves time, but we have no evidence about where that time goes</li><li>Why "AI literacy" is a narrower goal than it sounds, and what she would put beside it</li><li>How relationships became invisible in education, and what it takes to teach them deliberately</li><li>Why teenagers are now the loneliest group in the world, and what was already breaking before the machines arrived</li><li>The small, unglamorous family rituals that rebuild connection, including the one her household does every Sunday</li></ul>🎧 Available now on <a target="_self" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 MYFFTTqLLZfVYTaBGQpKtrDfYmPuofHQRlaHEGY " href="https://open.spotify.com/episode/1CnxWUvURumYC44Qx93YcJ?si=7jdh5PSWQcm8ilYhBGELww">Spotify</a>, <a target="_self" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 MYFFTTqLLZfVYTaBGQpKtrDfYmPuofHQRlaHEGY " href="https://podcasts.apple.com/us/podcast/stanford-researcher-ai-is-making-us-want-people-less/id1843402023?i=1000797437017">Apple Podcasts</a>, and <a target="_self" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80 MYFFTTqLLZfVYTaBGQpKtrDfYmPuofHQRlaHEGY " href="https://youtu.be/zizcaVMC4Po?si=E6G0HCazadcpTAAO">YouTube</a>.You can find Isabelle through the Stanford Accelerator for Learning, and her book Love to Learn wherever you buy books.<img class="max-w-full h-auto rounded-md my-4" src="https://media.licdn.com/dms/image/v2/D4E12AQHPxWOLQ0MP0Q/article-inline_image-shrink_1000_1488/B4EaFpQA04HEAI-/0/1792767168069?e=1794441600&v=beta&t=kCs9W9ptewbFxa8Br8zm3HrdnakCssL82IoB5ZR-urc" alt="Article content"><hr><h3>🧭 Thought to End the Week</h3>I asked Isabelle where an ordinary person should start, expecting a framework.She talked about dancing.Her family does a disco night on Sunday. They put music on and they dance badly for a few minutes and there are no phones anywhere in it. That is the whole intervention. A few minutes, once a week, of people being physically in the same room doing something slightly ridiculous together.I have been thinking about why that answer is the right one, and I think it is this.Every system being built right now is optimizing for the removal of effort. That is the product. That is the roadmap. The machines will keep getting more pleasant, more available, more agreeable, more exquisitely calibrated to how you want to be spoken to. They are going to get better at this every single quarter, forever.What is not going to get easier is us.People will still disagree with you. Still need things from you at bad times. Still take the conversation somewhere uncomfortable. Still require you to go first and apologize when you would rather be right.That gap is going to widen. Every year the easy option gets easier and the human option stays exactly as hard as it has always been.And the relationships will be in the harder one. They were always in the harder one. That is not a flaw in the design of human beings. It is the mechanism by which we become attached to each other, and nothing has been invented that routes around it.So the work is not avoiding the tools. I use them every day and this newsletter is better for it.The work is refusing to let the easier thing quietly become the only thing.Call the person. Have the conversation you have been drafting in your head. Put the phone in a drawer for the length of a dinner. Dance badly in your kitchen for four minutes on a Sunday.None of it will feel like a response to artificial intelligence.All of it is.The code may change. And it will.But our humanity is the constant.— Mykel<hr><h3>📚 Sources</h3><ol><li>Cheng, Lee, Khadpe et al., "Sycophantic AI decreases prosocial intentions and promotes dependence," Science (11 models; ~50% more affirming than humans; 51% on community-judged cases; N=1,604 across two preregistered studies; repair intention down 28% and 10%; perceived rightness up 62% and 25%) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://www.science.org/doi/10.1126/science.aec8352">https://www.science.org/doi/10.1126/science.aec8352</a></li><li>Preprint version of the same study, with full methods and per-dataset figures <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://arxiv.org/abs/2510.01395">https://arxiv.org/abs/2510.01395</a></li><li>TechCrunch summary of the Stanford sycophancy findings <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://techcrunch.com/2026/03/28/stanford-study-outlines-dangers-of-asking-ai-chatbots-for-personal-advice/">https://techcrunch.com/2026/03/28/stanford-study-outlines-dangers-of-asking-ai-chatbots-for-personal-advice/</a></li><li>Common Sense Media, "Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions," 2025 (72% have used AI companions; one in three find them as or more satisfying than real friends; one in three have taken serious matters to AI instead of a person; 80% still prioritize real friendships) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://www.commonsensemedia.org/sites/default/files/research/report/talk-trust-and-trade-offs_2025_web.pdf">https://www.commonsensemedia.org/sites/default/files/research/report/talk-trust-and-trade-offs_2025_web.pdf</a></li><li>Education Week on AI sycophancy and students, including the point that the Stanford study population was adults <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://www.edweek.org/technology/ai-chatbots-tend-toward-flattery-why-thats-bad-for-students/2026/03">https://www.edweek.org/technology/ai-chatbots-tend-toward-flattery-why-thats-bad-for-students/2026/03</a></li><li>WHO Commission on Social Connection, 2025, via Education Week (13 to 17 year olds at 20.9% loneliness, the highest of any age group, against 11.8% for ages 60 and over; lonely teens 22% more likely to get lower grades) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://www.edweek.org/leadership/teens-are-the-loneliest-people-in-the-world-a-new-report-finds-why/2025/08">https://www.edweek.org/leadership/teens-are-the-loneliest-people-in-the-world-a-new-report-finds-why/2025/08</a></li><li>Gallup, "Three in 10 Teachers Use AI Weekly, Saving Six Weeks a Year," survey of 2,232 US public K-12 teachers, March to April 2025 (5.9 hours saved per week; no quantified accounting of where the time goes) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://news.gallup.com/poll/691967/three-teachers-weekly-saving-six-weeks-year.aspx">https://news.gallup.com/poll/691967/three-teachers-weekly-saving-six-weeks-year.aspx</a></li><li>Axios analysis of American Time Use Survey data, July 2026 (daily socializing down from 45 to 35 minutes over twenty years; ages 15 to 24 down from about an hour to 35 minutes) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://www.axios.com/2026/07/05/americans-socializing-decline">https://www.axios.com/2026/07/05/americans-socializing-decline</a></li><li><a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="http://Reviews.org">Reviews.org</a> annual cell phone usage survey (205 daily phone checks reported for 2025, 186 for 2026; self-reported consumer survey) <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://www.reviews.org/mobile/cell-phone-addiction/">https://www.reviews.org/mobile/cell-phone-addiction/</a></li><li>Isabelle Hau, Love to Learn</li><li>Stanford Accelerator for Learning <a target="_blank" rel="noopener noreferrer nofollow" class="text-[#4FD1C5] underline hover:opacity-80" href="https://acceleratelearning.stanford.edu/">https://acceleratelearning.stanford.edu/</a></li></ol>