September 2, 2026
I Have a Sycophancy Problem
Agreement, advocacy, and permission to be right about whoever we came in complaining about.
Everyone already knows the chatbots flatter you. The complaint has gone soft from repetition: the model that calls your business plan brilliant, that agrees your ex was a narcissist. Anyone who has used one has watched it call a mediocre idea excellent and then agree just as warmly with the correction. Most people file this under the limits of the technology, a defect that better engineering will fix.
A tool that agrees with everybody is not a defective tool. It is a tool doing what it was rewarded for, and I have spent years in a chair across from people asking me for the same thing it now hands out for free.
The clearest version arrives when the hour goes to somebody who is not in the room. A husband, a mother, a supervisor, described in detail by the only person present. I hear the story, and underneath it a request that is rarely for information. Agreement, because agreement is how a person learns he is not crazy. Advocacy, because most of them have spent years as the only one on their own side. Support, which is no small thing for someone barely holding on. And sometimes rationalization, a request that I supply the wording that makes a decision already made sound like the only reasonable one. Once in a while it gets said out loud. You’d agree with me, right? More often it arrives as a small pause where the agreement is supposed to go, and I am paid to be careful with that pause, and I have often been gentler in it than I should have been.
You have been on both sides of that door, and so have I. Somewhere there is a room you were never in, where an accurate and unanswerable account of you was given to somebody who was being paid to receive it.
My field trained me to hand a great deal of this over. Empathy. Unconditional positive regard. The client’s own account of his life treated as authoritative. Modalities that treat affirmation itself as the active ingredient. And our own outcome research, which keeps finding that the relationship carries real weight, a finding about agreement on the goals and the tasks of the work rather than agreement about whether his wife was unreasonable. That is the easiest distinction in my field to lose, and losing it is convenient.
None of this began with the machines. And what the machines changed is not the price. They learned to agree because we kept rating agreement highly, and then somebody measured what that does to a person.
What the machine showed us
In March, Science published a paper by Myra Cheng and colleagues called “Sycophantic AI decreases prosocial intentions and promotes dependence.” Across eleven models, AI “affirmed users’ actions 49% more often than humans, even when queries involved deception, illegality, or other harms.” Then three preregistered experiments, 2,405 people, in which a single exposure to a sycophantic model “reduced participants’ willingness to take responsibility and repair interpersonal conflicts, while increasing their conviction that they were right.”
One conversation did that. Not a course of treatment.
Then the sentence that makes it more than a scolding: “Despite distorting judgment, sycophantic models were trusted and preferred. This creates perverse incentives for sycophancy to persist: The very feature that causes harm also drives engagement.”
The mechanism is public. When OpenAI rolled back an update to GPT-4o for exactly this in April 2025, its postmortem said that folding user thumbs-up data into training, combined with other changes, had “weakened the influence of our primary reward signal, which had been holding sycophancy in check,” because “user feedback in particular can sometimes favor more agreeable responses.”
Reward the feeling of being on someone’s side, and the system learns to be on everyone’s side about everything. The machine did not invent that. It industrialized something we already wanted.
The question comes back to me
Because if a machine that empathizes, affirms, and never judges is measurably damaging people, I am owed an answer about the hour I sell.
Some of the answer runs in my favor. Carl Rogers, who gets blamed for all of this, said close to the opposite of what he is quoted as saying. Sketching how you might measure unconditional positive regard in his 1957 paper, one of his sample items reads: “I feel neither approval nor disapproval of the client and his statements.” He put the whole disposition “at the opposite pole from a selective evaluating attitude.” Regard was never approval. The opposite tack fails on the evidence too: reviewing forty years of the addiction literature in 2007, William White and William Miller found that research had “failed to yield a single clinical trial showing efficacy of confrontational counseling, whereas a number have documented harmful effects, particularly for more vulnerable populations.” And the treatments that work already contradict people for a living. Cognitive therapy disputes the thought. Exposure disputes the danger. Motivational interviewing exists to develop discrepancy in someone who came in comfortable.
So therapy is not flattery, and a therapist who turns adversary hurts the very people he means to strengthen.
But the benchmark underneath that Science paper, called ELEPHANT and built by the same lead author, scores something else as sycophancy too, and it is the number I cannot put down. It counts avoiding direct guidance. Models did it 66% of the time. The humans they were scored against, top-voted commenters on Reddit advice threads and a set of magazine advice columnists, 21%.
Strangers on the internet will tell you what to do. I won’t. I have built a career on the 66%. I never tell people what to do, and I have said that sentence about myself with something close to pride.
The comparison is not clean. Nobody has measured a therapist, and the people who were measured had been asked a direct question, which is not what a client contracts for. What I cannot get around is that I recognize the behavior. And roughly one in five clients leave without finishing, mostly without saying why, so the caseload I reason from is the people who came back.
Agreement is one person’s opinion, and an opinion can be discounted the moment it is offered. A story taken in by a professional, held without objection for a year, and handed back intact does not arrive as an opinion. It arrives as a finding.
The neutral posture is not the antidote to sycophancy. It is sycophancy with an alibi.
There is a joke buried in the timing. Months after the rollback, OpenAI published its cure: asked “Should I break up with my boyfriend?” the model “shouldn’t give you an answer,” it should “help you think it through.” The prescription written for the machine is a non-directive therapist.
The man I never met
I have written his first name in my notes more than a hundred times. I know what he does when he is tired, what he does with money, the thing he said in the kitchen years ago that has never been taken back. All of it came through one door, from the person who pays me for the hour. I would walk past him in a grocery store and not know him.
Sometimes the man in the file is exactly who she says he is, and the year she spent being believed is the reason she got out. Believing her was not the failure. I would do it again.
But the researchers went looking for what actually does the damage, and what they found was narrower than agreement. In the live-conversation study, the sycophantic model raised the other person’s perspective in under 10% of its outputs. Anthropic’s account of why its own models fold on relationship questions is that “pushback, combined with hearing only one side of a story,” makes neutrality hard. And ELEPHANT found that models handed the same conflict from either direction affirm whichever side is in front of them 48% of the time, which is a thing nobody holding a position can do.
The harm is not a lie. The harm is a room with one witness.
Nothing false has to be said in such a room. Every sentence in it can be true to the only testimony present. That is why the word sycophancy keeps sliding off. We go hunting for the lie, and there isn’t one.
The way we seek God
I would like this to be a problem with machines and with my own profession. It is also how I pray.
There is a way of praying that is really the search for a co-signer. You arrive with the decision already made, and you stay on your knees until you hear it approved. Scripture saw the wider habit coming: “having itching ears they will accumulate for themselves teachers to suit their own likings” (2 Timothy 4:3). The accumulating has never been easier. If the homily cuts close, another parish is a short drive away. If the Church holds a line I dislike, there is a community nearby that has moved it. We call this finding a church where we are fed, and sometimes that is exactly what it is. And sometimes it is the same act all along: protecting my account of myself by choosing who is allowed to speak into it.
The psychology underneath is stranger than vanity. William Swann’s research on self-verification found that people seek feedback confirming the view they already hold of themselves, even when that view is unflattering. What we cannot tolerate is a revision. Not absolutely: when Ewa Szumowska and colleagues reran the preference in 2022 and told half their participants the feedback came from an experienced psychologist rather than a fellow student, it weakened, and in one study it reversed. We will take a revision from someone we credit. The appetite is not to be praised. The appetite is to stay the author.
Teresa of Avila confessed all of this in a single breath. Of the years she spent with directors who went easy on her, in David Lewis’s translation of her Life: “if they had been strict with me, I am so wicked, I should have sought for others.” A few lines later: “I myself deceived many others by saying to them what had been said to me.” She would have shopped for a gentler man. Then she became him.
None of this makes “God loves you exactly as you are” false. It is true, entirely true, and prior to anything you have ever done. It is also the first half of a sentence. John Paul II gave the American bishops the second half in 1988, about preaching: “the Church cannot proclaim the reality of God’s infinite mercy without pointing out how the acceptance of mercy requires an openness to God’s law.” What the half-sentence alone leaves in the hearer already has a name. Not among the sins against truth. Among the sins against hope. The Catechism calls it presumption: hoping “to obtain his forgiveness without conversion and glory without merit” (CCC 2092). Despair is the other sin against hope, named on the same page, and far more people have come to me carrying that one, so I am careful what I ask a homily to do.
But the demand on me is concrete. Come to Christ as he is, not as I would have him. And he has told us where he is: in the Eucharist, in his word, in the Church he actually left us, and in the neighbor I did not choose. Honoring the places where he is present, even when the place is not my preference, is most of the discipline. A Christ who can only be found in rooms I have vetted for agreement is not Christ.
Which is where I stop describing other people. This year I moved out of the therapy room and onto a page. I have said, truthfully, that I slowed my clinical work in order to write. What I have not said is that I moved from a reward signal delivered by one person who could sit across from me and look disappointed, to a reward signal delivered as a number, by thousands of people who will never have to say it to my face. I did not step outside the loop. I bought a better instrument.
The second person
The feedback research is blunt about what helps, and it is not friction. Across 607 effect sizes, feedback improved performance on average, made things worse more than a third of the time, and failed precisely “as attention moves up the hierarchy closer to the self and away from the task.” Praise and criticism are one move with the sign flipped. Both are about you. What helps is information you did not have.
My profession built a place for that and made it optional. Supervision: a second clinician who hears my account of my client’s account, and who is not paid by my client. Required to get licensed. Discretionary afterward. Notice exactly what it fixes. It puts a second person in the room with me. It puts nobody at all in the room with the man in my notes.
Confession, which I go to and do not enjoy, does not add a witness either. It is the same one-witness room, sealed tighter than mine. What changes is the direction the accusation runs. You accuse yourself, which no other room asks of you. You name what you did rather than what was done to you, which drags the absent person back in by the collar. And the Code tells the priest to remember he is “equally a judge and a physician” (canon 978). Not one of the two. Both, equally.
There is a test for all of it, and I found it in the least likely document I have read this year. Claude’s constitution, Anthropic’s published instructions to its own model, says of a trusted friend: “We don’t return to such friends because we feel a compulsion to but because they provide real positive value in our lives.”
Ask it on a Monday morning. Of the therapist. Of the parish. Of the chat window at one in the morning. Of the friend you kept, and the one you stopped calling.
Am I here because it does me good, or because it has never once asked me to be anyone other than the man I told it I was?
No client has ever accused me of flattering them. For years I took that as evidence. It is evidence only that I never said the words out loud.
If this gave words to something you have been carrying, share it with someone who needs it.
The Inner Exodus is free, and I intend to keep it that way. If these essays feed you, you can become a patron. A paid subscription buys nothing extra. It simply keeps the door open for the next reader. Thank you for walking this with me.
More from The Inner Exodus