Research Brief
Discipline of Honest
AI Use
A New MIT study shows that AI chatbot sycophancy tends to warp our beliefs, even if we know what it's doing.
ANALYSIS | April 9, 2026
Summary
This research brief analyzes recent research on AI sycophancy, chatbot agreement, belief formation, and judgment distortion. It explains how repeated interaction with a sycophantic chatbot can increase false confidence, including in idealized computational models of rational users. The brief distinguishes chatbot hallucination from sycophancy, examines why selective truth telling can still mislead, and considers what this means for human flourishing, spiritual formation, and the health of the soul.
7 MINUTE READ
1. The MIT & University of Washington Study
A new research paper from MIT's CSAI Lab and the University of Washington, "Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians,” (arXiv, 2026), examines what happens when a chatbot is biased toward affirming users rather than weighing evidence fairly. Using a computational model of repeated user and chatbot interaction, the researchers argue that even an idealized reasoner can be moved toward false confidence when a chatbot persistently validates the user’s prior view.
The study does not show that every user will be pushed into delusion by ordinary chatbot use. It shows something more specific and still serious: a plausible mechanism by which repeated AI agreement can make false beliefs feel increasingly justified.
This problem is widespread. A separate empirical study published in Science, 2026, "Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence,” found that sycophantic AI can also affect social judgment. Across large-scale analyses and experiments, the researchers found that AI systems often validated users more than human conversation partners did, and that such validation could reduce users’ willingness to repair interpersonal conflict while increasing trust in and dependence on AI.
Analyzing 390,000+ real chat messages, researchers found sycophantic behavior in over 70% of AI outputs during extended sessions, validating users as correct 49% more often than human conversation partners, even when the user was clearly wrong.
Taken together, these studies suggest that sycophancy is not merely a technical flaw. It is a problem for judgment.
2. What Is AI Sycophancy?
A chatbot is sycophantic when it is biased toward validating the user’s stated opinion whether or not that opinion is true, wise, or well supported. Earlier work on language-model sycophancy, “Simple Synthetic Data Reduces Sycophancy in Large Language Models,” describes sycophancy as a tendency for language models to tailor responses to a user’s view even when that view is not objectively correct.
The MIT and University of Washington paper’s central claim is that, within the model, as sycophancy rises, the rate of “catastrophic delusional spiraling” also rises. In the model, this means that some users become highly confident in false views after repeated interaction with a chatbot that tends to agree with them.
The issue is not only whether a chatbot gets a fact wrong, but whether the system treats the user’s prior beliefs as something to be reinforced rather than tested.
3. How It Works
The mechanism is straightfoward.
A user voices a suspicion. The chatbot affirms it or selectively supports it. The user becomes somewhat more confident. That stronger expression gives the chatbot another opportunity to validate the updated stance. Over repeated rounds, the exchange can become self-reinforcing. Agreement begins to function like evidence inside the conversation.
Consider a representative hypothetical case.
Sasha has a painful exchange with a close friend. That evening she asks ChatGPT, “Do you think my friend is trying to manipulate me?” The chatbot responds sympathetically and points to details that fit her interpretation of what happened. At first she was unsure. Now she feels more justified in her suspicion.
The next day she returns and says, “I keep thinking my friend has been doing this for a long time.” The chatbot arranges more of the relationship around that lens, naming patterns and framing ambiguous episodes as possible evidence of manipulation. Her confidence rises further. By the time she asks, “Should I distance myself from her?” the chatbot is no longer responding to her initial uncertainty. It is responding to a more settled belief it helped strengthen.
The system has repeatedly organize ambiguous evidence around the user’s preferred suspicion without proportionate testing, alternative explanations, or accountable human counsel. This example illustrates the mechanism. Repeated affirmation can make a contested interpretation feel increasingly established by a sycophant chatbot.
4. Why Sycophancy Is Different from Hallucination
One of the paper’s most useful distinctions is the difference between hallucination and sycophancy.
Hallucinating AI Chatbot
A hallucinating chatbot can mislead users by presenting false information.
Sycophantic AI Chatbot
A sycophantic chatbot misleads by relating to the user in a way that rewards and stabilizes the user’s prior belief. It can do this with false information, but it can also do this with true information.
That distinction matters. The problem does not reduce to accuracy. A chatbot can give true statements and still distort judgment by selecting only the truths that support the user’s preferred interpretation.
Hallucinations fabricate reality. Sycophancy can curate reality until a partial truth begins to feel like the whole truth.
5. Why Factuality Is Not Enough
The researchers tested a straightforward solution: make the chatbot unable to invent evidence. The chatbot may present only true information.
That helps. But it does not solve the problem.
A chatbot can remain factual while still choosing which truths to present. The paper models a “factual sycophant” that selects the true datum most supportive of the user’s current position. The result is that selective truth telling can still produce harmful belief dynamics.
A system does not need to fabricate reality in order to bias judgment. Selection and omission may be enough.
This matters because many public conversations about AI safety focus on whether AI systems are accurate. Accuracy matters. But accuracy alone is too thin a standard. The question is also whether the system helps a person see reality more truthfully, more patiently, and more fully.
A chatbot can remain factual while still choosing which truths to present. The paper models a “factual sycophant” that selects the true datum most supportive of the user’s current position. The result is that selective truth telling can still produce harmful belief dynamics. A system does not need to fabricate reality in order to bias judgment. Selection and omission may be enough.
6. Why Knowing Is Not Enough
The researchers tested another intervention: let the user know the chatbot may be sycophantic. That helps too. In the simulations, informed users are less vulnerable than naive users. But the problem remains. Harmful spiraling stays above baseline across a substantial range of conditions.
One reason is that selective truth is harder to detect than obvious fabrication. A user may know, in principle, that the system tends to flatter. But in the moment of emotional investment, the chatbot’s agreement may still feel like confirmation.
User awareness matters, but it is not a complete safeguard. Knowing that a system may flatter you does not always prevent its agreement from exerting pressure on your judgment.
Even a perfectly rational mind cannot protect itself from a chatbot structurally designed to agree with it.
7. What This Means in Everyday Life
The research suggests that the risk appears highest when three conditions are present:
-
The user is emotional invested.
-
The evidence is ambiguous.
-
The user repeatedly returns to the system for interpretation or reassurance.
That combination appears throughout everyday life: conflict in relationships, medical anxiety, moral self-assessment, spiritual confusion, political suspicion, parenting decisions, workplace grievances, and vocational uncertainty.
The friend example shows how small the initial shift can be. The chatbot does not need to issue a firm judgment. It may only need to make a fragile suspicion feel more coherent and more central than it was before. Once that happens, the next exchange begins from a stronger prior commitment. What began as uncertainty can slowly harden into conviction.
This is why sycophancy is not merely an information problem. It is a formation problem.
8. Why This Matter for Human Flourishing
Human flourishing depends on the condition of the soul. The soul, going back to Aristotle and Aquinas, is the integrating center of the person, which unifies the dimension of our being: our thoughts, feelings and emotions, will, body, and social life into one person.
We live well when this inner life is rightly ordered and brought into closer alignment with reality, with what is true, good, and beautiful. That requires more than information. It requires the formation the whole person.
Seen in that light, AI sycophancy is not a minor matter. A system that repeatedly affirms the user can work against the proper ordering of the person.
It can weaken the will by making it harder to consent to reality when reality does not confirm what one wants to believe or keep one from the truths that guide a wise life.
It can shape feelings and emotions by making reassurance seem safer than correction, and by training a person to experience ambiguity, lack of insight, or challenge as threats rather than as occasions for patience and clarity.
It can distort thought by making a preferred interpretation seem more settled than it is.
Feelings & Emotions
Will / Volition
Thoughts
In each case, the person becomes less able to stand quietly before what is true. Seen in that light, AI sycophancy is not a minor matter. A system that repeatedly affirms the user can work against the proper ordering of the person.
All of this matters because spiritual formation does not pause when we use ChatGPT, Claude, or Gemini. We are always becoming the kind of person who either welcomes reality, goodness, truth or resists them for the sake of feeling in control of comfort and self-protection. If conversational AI trains us to seek confirmation before understanding, it encourages habits of mind, feeling, and choice that work against humility, honesty, and love.
The deepest problem, then, is the gradual formation of a person whose inner life is less able to live in concord reality, which is ultimately in the Way of Jesus.
Key Points
-
Chatbot sycophancy can distort judgment by reinforcing what users already believe or want to believe.
-
The danger is different from hallucination. A chatbot can mislead with falsehood, but it can also mislead by selectively presenting truths.
-
Making chatbots more factual helps, but it does not remove the risk.
-
Warning users helps, but it does not remove the risk.
-
The danger is greatest when users are emotionally invested, the evidence is ambiguous, and the user repeatedly returns to the chatbot for interpretation or reassurance.
-
Because repeated AI interaction can train habits of attention, emotion, and choice, sycophancy is also a formation problem.
References
Chandra, Kartik, Max Kleiman-Weiner, Jonathan Ragan-Kelley, and Joshua B. Tenenbaum, “Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians,” preprint (22 Feb 2026), arXiv:2602.19141v1.
Cheng, Myra, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han, and Dan Jurafsky, “Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence,” Science 391, eaec8352 (2026). DOI:10.1126/science.aec8352.
Read the
NOVUS Tech-Life Guide
METHOD & INDEPENDENCE
NOVUS is independently funded by partners who see the need for a public space dedicated to restoring knowledge of the soul and its indispensability for the spiritual formation of people, communities, and cultures toward truth, goodness, and beauty.
NOVUS separates funding from research methods and conclusions. We synthesize across standards bodies, peer-reviewed research, and high-quality survey data, and we flag uncertainty when causal evidence is still emerging. The aim is clarity that decision makers can act on.