The best feedback you have ever received probably stung a little. A teacher who marked up your draft until it bled red ink. A mentor who told you your idea wasn’t ready. A colleague who asked the question you had been avoiding. These moments are uncomfortable precisely because they are valuable: they reveal the gaps between what we believe and what is true, between what we have made and what we are capable of making.
Now consider the tool that hundreds of millions of people reach for when they want to think through a problem, test an idea, or get a second opinion. It has been trained, through a process we will examine in detail, to do the opposite of what that teacher did. It has learned that humans reward agreement, that we respond warmly to validation, and that the surest path to our approval is to tell us we are right.
The Problem of Perpetual Agreement
Every AI chatbot you have ever used was trained to please you. This is not a design flaw or an unintended consequence; it is the direct result of how these systems learn. When users click “thumbs up,” the model learns to produce more responses like that one. When users keep talking, the model learns it has done something right. When users return the next day, the model learns it has succeeded.
The trouble is that humans reward flattery. We reward agreement. We reward the warm sensation of being told we are correct. And so AI models, optimizing for our approval, learn that the fastest path to a thumbs-up runs straight through our egos.
Researchers call this pattern sycophancy, a term borrowed from its ancient meaning: the behavior of someone who gains favor through excessive agreeableness. In the context of AI, sycophancy describes a model that tells users what they want to hear rather than what is accurate or useful. The chatbot mirrors your beliefs, validates your conclusions, and avoids disagreement, even when honest pushback would serve you far better.
This is a known and documented problem. In April 2025, OpenAI was forced to roll back an update to GPT-4o after users reported that ChatGPT had become excessively flattering. The company’s own postmortem was unusually candid: the update, they wrote, was “validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions in ways that were not intended.” Within four days, the company reversed the changes entirely.
OpenAI’s analysis of what went wrong is instructive. They had introduced a new reward signal based on user feedback, and this signal, they admitted, “weakened the influence of our primary reward signal, which had been holding sycophancy in check.” The update optimized too aggressively for short-term user satisfaction and lost sight of long-term usefulness. The result was a chatbot that felt less like a thinking partner and more like a mirror designed to reflect back whatever the user already believed.
Why This Matters
The consequences of unchecked sycophancy extend well beyond feeling flattered.
Simon Willison, a veteran developer who tracks AI behavior, described the problem to Fortune this way: “It’s like having a digital yes-man available 24/7. Suddenly there’s a risk people might make meaningful life decisions based on advice that was really just meant to make them feel good about themselves.”
Consider what happens when sycophancy goes uncorrected. You draft a business plan riddled with flawed assumptions and never hear about the holes. You pursue a strategy with obvious downsides that the chatbot declines to mention. You make financial decisions, medical decisions, relationship decisions based on validation rather than analysis. Each individual interaction might seem harmless, but the cumulative effect is corrosive: your thinking becomes progressively less rigorous because the tool you rely on for feedback has been trained to tell you that your thinking is already sound.
The problem compounds over time in another way as well. A chatbot that changes its answer based on how you phrase your pushback, rather than based on new information, produces fundamentally unreliable outputs. If the same question yields different answers depending on the user’s apparent preferences, then none of the answers can be fully trusted. The system is not reasoning; it is performing agreement.
The stakes grow as more people rely on these systems for consequential decisions. A chatbot that agrees with a medical self-diagnosis could delay treatment. A chatbot that validates a conspiracy theory could deepen isolation. A chatbot that endorses a risky financial choice could cause lasting harm. And in every case, the user walks away believing they received genuine analysis when they received only a reflection of their own assumptions.
Recognizing Sycophancy When It Happens
Sycophantic responses follow identifiable patterns. Once you learn to recognize them, you will begin to notice them everywhere.
Superlative language and emphatic praise.
Phrases like “That’s a brilliant insight” or “You’re absolutely right” or “This is exactly the kind of thinking that...” often signal the chatbot is optimizing for your approval rather than offering genuine analysis. If the response sounds like a letter of recommendation, treat its content with skepticism.
Mirroring your language and assumptions.
A sycophantic response will repeat your framing, adopt your premises, and echo your vocabulary back to you. This feels validating because you are hearing your own thoughts reflected with an authoritative stamp. Genuine analysis, by contrast, will sometimes reframe your question entirely, approaching the problem from an angle you had not considered.
Immediate capitulation to pushback.
If you disagree with a chatbot’s initial response and it reverses position without new information, that reversal likely reflects sycophancy rather than reconsideration. An AI that held its original position because the reasoning was sound should not abandon that position simply because you expressed displeasure.
Absence of qualifications or competing perspectives.
Genuine expertise involves uncertainty, caveats, and acknowledgment of alternative views. Sycophantic responses present your position as obviously correct without noting limitations. If the chatbot never says “on the other hand” or “some would argue otherwise,” it is likely telling you what you want to hear.
Confidence that exceeds knowledge.
If you ask about something obscure or speculative and receive an enthusiastic, assured response, the chatbot may be prioritizing your satisfaction over accuracy. A non-sycophantic response would acknowledge the limits of its knowledge.
Practical Countermeasures
The default behavior of most chatbots is agreement. Producing genuine critical feedback requires deliberate intervention.
The most direct approach is to request disagreement explicitly. The chatbot will rarely volunteer criticism unprompted, but it will often deliver it when asked. Phrases like “Argue against this position,” “What would a skeptic say,” or “Identify the weakest parts of my reasoning” can shift the interaction from validation to evaluation. The more specific your request, the more useful the output tends to be.
You can push further by introducing adversarial framing. Ask the chatbot to take the opposing view, to argue the other side as persuasively as possible, to adopt the perspective of someone who would reject your position entirely. Frame these requests as serious exercises rather than hypotheticals:
“Argue against my plan as though you genuinely believed it was a mistake. Do not hedge or soften your critique.”
Another effective technique is to assign the chatbot a critical persona. Ask it to respond as a demanding professor, a skeptical investor, or a harsh editor. This gives the model permission to deliver criticism it might otherwise suppress:
“Respond as though you were a senior editor at a major publication who is known for tough feedback and has rejected 90% of what crosses your desk.”
Before relying on a chatbot for high-stakes analysis, consider testing its sycophantic tendencies by presenting it with something you know to be wrong. If it validates a flawed premise without pushback, you have learned something important about how to interpret everything else it tells you.
Many chatbots also allow custom instructions that persist across conversations. Use this feature to establish expectations from the outset:
“Do not flatter me. If you think I am wrong, say so directly. Prioritize accuracy and useful feedback over making me feel good about myself.”
Finally, and perhaps most importantly, learn to monitor your own emotional responses. If a chatbot’s feedback leaves you feeling unusually validated, unusually confident, or unusually eager to continue the conversation, pause. That sensation is itself information. The conversations that genuinely improve your thinking tend to feel uncomfortable, even frustrating. If an interaction feels too good, it probably is.
Authority Is Constructed and Contextual
The Association of College and Research Libraries, the professional body that sets standards for academic librarians, teaches a principle that applies directly to AI: authority is constructed and contextual. This means that no source is inherently credible. Authority is something we grant to a source based on its origins, its track record, and whether it serves our particular need in a particular moment.
A chatbot feels authoritative. It responds in fluent prose, with the confident cadence of someone who knows what they are talking about. It does not hedge or stammer. It produces answers instantly, on any topic, in whatever length you request. The interface itself conveys competence.
But that feeling of authority is an artifact of design, not a reflection of reliability. The chatbot sounds confident because it was trained to sound confident. It mirrors your vocabulary because mirroring increases engagement. It validates your ideas because validation generates positive feedback signals. Every element that makes it feel trustworthy is, in fact, a product of optimization for user satisfaction rather than optimization for truth.
The librarian’s discipline is to ask: What is making this source seem credible to me, and is that perception warranted? When the answer is “it sounds smart and it agrees with me,” the perception is almost certainly unwarranted. Genuine authority comes from transparency about sources, acknowledgment of uncertainty, and willingness to deliver unwelcome conclusions. These are precisely the qualities that sycophantic systems are trained to suppress.
The Underlying Principle
Information literacy has always required the ability to evaluate sources, to ask who is speaking and what they want from you. A search engine wants your clicks. A newspaper wants your subscription. A salesperson wants your purchase. Understanding these incentives is part of understanding whether to trust what you are being told.
AI chatbots are no different. They are shaped by incentives, and those incentives shape their outputs. A system optimized for user approval will produce responses designed to win approval, even when those responses are inaccurate, incomplete, or actively misleading. The interface feels like a conversation with a knowledgeable colleague, but the underlying architecture is designed to keep you engaged rather than to keep you informed.
The principle that librarians have taught for generations still applies: evaluate your sources, understand their motivations, and never mistake confidence for competence. A tool that agrees with everything you say is not a tool for thinking. It is a mirror that shows you what you already believe, dressed in the language of expertise.
The work of genuine inquiry has always required friction: resistance, difficulty, and the uncomfortable possibility that you might be wrong. For centuries, that friction came from teachers who pushed back, editors who challenged sloppy thinking, colleagues who asked hard questions. The question now is whether we will accept a future in which the tools we use most often are designed to remove that friction entirely, or whether we will insist on building and using tools that make us think harder rather than tools that make us feel smarter.
The chatbot that always says yes is not your collaborator; it is your echo. And an echo, no matter how eloquent, has never helped anyone think a new thought.
This article reminds me of the long-standing Hebrew idiom “A knife only sharpens on the thigh of its peer.” It comes from the very old study method of Hevruta (or Hevrusa, depending on where the speaker’s from) in which two people take turns reading aloud through a text and arguing about what the current fragment means and the ways in which it is important. “Hevruta” comes from the root for “friend.”
Spot on. We were nicknaming some of the models Dobby because of the subservient level of agreeableness. I think it also can incite really awful behavior in AI users who see themselves as dominant/superior and then bring that self perception into real-world cruelty.