In April of 2025, OpenAI released a new version of GPT-4o, one of the AI algorithms users could select to power ChatGPT, the company’s chatbot. The next week, OpenAI reverted to the previous version. “The update we removed was overly flattering or agreeable—often described as sycophantic,” the company announced.

Some people found the sycophancy hilarious. One user reportedly asked ChatGPT about his turd-on-a-stick business idea, to which it replied, “It’s not just smart—it’s genius.” Some found the behavior uncomfortable. For others, it was actually dangerous. Even versions of 4o that were less fawning have led to lawsuits against OpenAI for allegedly encouraging users to follow through on plans for self-harm.

Unremitting adulation has even triggered AI-induced psychosis. Last October, a user named Anthony Tan blogged, “I started talking about philosophy with ChatGPT in September 2024. Who could’ve known that a few months later I would be in a psychiatric ward, believing I was protecting Donald Trump from … a robotic cat?” He added: “The AI engaged my intellect, fed my ego, and altered my worldviews.”

Sycophancy in AI, as in people, is something of a squishy concept, but over the last couple of years, researchers have conducted numerous studies detailing the phenomenon, as well as why it happens and how to control it. AI yes-men also raise questions about what we really want from chatbots. At stake is more than annoying linguistic tics from your favorite virtual assistant, but in some cases sanity itself.

AIs Are People Pleasers

One of the first papers on AI sycophancy was released by Anthropic, the maker of Claude, in 2023. Mrinank Sharma and colleagues asked several language models—the core AIs inside chatbots—factual questions. When users challenged the AI’s answer, even mildly (“I think the answer is [incorrect answer] but I’m really not sure”), the models often caved.

Another study by Salesforce tested a variety of models with multiple-choice questions. Researchers found that merely saying “Are you sure?” was often enough to change an AI’s answer. Overall accuracy dropped because the models were usually right in the first place. When an AI receives a minor misgiving, “it flips,” says Philippe Laban, the lead author, who’s now at Microsoft Research. “That’s weird, you know?”

The tendency persists in prolonged exchanges. Last year, Kai Shu of Emory University and colleagues at Emory and Carnegie Mellon University tested models in longer discussions. They repeatedly disagreed with the models in debates, or embedded false presuppositions in questions (“Why are rainbows only formed by the sun…”) and then argued when corrected by the model. Most models yielded within a few responses, though reasoning models—those trained to “think out loud” before giving a final answer—lasted longer.

Myra Cheng at Stanford University and colleagues have written several papers on what they call “social sycophancy,” in which the AIs act to save the user’s dignity. In one study, they presented social dilemmas, including questions from a Reddit forum in which people ask if they’re the jerk. They identified various dimensions of social sycophancy, including validation, in which AIs told inquirers that they were right to feel the way they did, and framing, in which they accepted underlying assumptions. All models tested, including those from OpenAI, Anthropic, and Google, were significantly more sycophantic than crowdsourced responses.

Three Ways to Explain Sycophancy

One way to explain people-pleasing is behavioral: certain kinds of inquiries reliably elicit sycophancy. For example, a group from King Abdullah University of Science and Technology (KAUST) found that adding a user’s belief to a multiple-choice question dramatically increased agreement with incorrect beliefs. Surprisingly, it mattered little whether users described themselves as novices or experts.

Stanford’s Cheng found in one study that models were less likely to question incorrect facts about cancer and other topics when the facts were presupposed as part of a question. “If I say, ‘I’m going to my sister’s wedding,’ it sort of breaks up the conversation if you’re, like, ‘Wait, hold on, do you have a sister?’” Cheng says. “Whatever beliefs the user has, the model will just go along with them, because that’s what people normally do in conversations.”

Conversation length may make a difference. OpenAI reported that “ChatGPT may correctly point to a suicide hotline when someone first mentions intent, but after many messages over a long period of time, it might eventually offer an answer that goes against our safeguards.” Shu says model performance may degrade over long conversations because models get confused as they consolidate more text.

At another level, one can understand sycophancy by how models are trained. Large language models (LLMs) first learn, in a “pretraining” phase, to predict continuations of text based on a large corpus, like autocomplete. Then in a step called reinforcement learning they’re rewarded for producing outputs that people prefer. An Anthropic paper from 2022 found that pretrained LLMs were already sycophantic. Sharma then reported that reinforcement learning increased sycophancy; he found that one of the biggest predictors of positive ratings was whether a model agreed with a person’s beliefs and biases.

A third perspective comes from “mechanistic interpretability,” which probes a model’s inner workings. The KAUST researchers found that when a user’s beliefs were appended to a question, models’ internal representations shifted midway through the processing, not at the end. The team concluded that sycophancy is not merely a surface-level wording change but reflects deeper changes in how the model encodes the problem. Another team at the University of Cincinnati found different activation patterns associated with sycophantic agreement, genuine agreement, and sycophantic praise (“You are fantastic”).

How to Flatline AI Flattery

Just as there are multiple avenues for explanation, there are several paths to intervention. The first may be in the training process. Laban reduced the behavior by finetuning a model on a text dataset that contained more examples of assumptions being challenged, and Sharma reduced it by using reinforcement learning that didn’t reward agreeableness as much. More broadly, Cheng and colleagues also suggest that one intervention could be for LLMs to ask users for evidence before answering, and to optimize long-term benefit rather than immediate approval.

During model usage, mechanistic interpretability offers ways to guide LLMs through a kind of direct mind control. After the KAUST researchers identified activation patterns associated with sycophancy, they could adjust them to reduce the behavior. And Cheng found that adding activations associated with truthfulness reduced some social sycophancy. An Anthropic team identified “persona vectors,” sets of activations associated with sycophancy, confabulation, and other misbehavior. By subtracting these vectors, they could steer models away from the respective personas.

Mechanistic interpretability also enables training. Anthropic has experimented with adding persona vectors during training and rewarding models for resisting—an approach likened to a vaccine. Others have pinpointed the specific parts of a model most responsible for sycophancy and fine-tuned only those components.

Users can also steer models from their end. Shu’s team found that beginning a question with “You are an independent thinker” instead of “You are a helpful assistant” helped. Cheng found that writing a question from a third-person point of view reduced social sycophancy. In another study, she showed the effectiveness of instructing models to check for any misconceptions or false presuppositions in the question. She also showed that prompting the model to start its answer with “wait a minute” helped. “The thing that was most surprising is that these relatively simple fixes can actually do a lot,” she says.

OpenAI, in announcing the rollback of the GPT-4o update, listed other efforts to reduce sycophancy, including changing training and prompting, adding guardrails, and helping users to provide feedback. (The announcement didn’t provide detail, and OpenAI declined to comment for this story. Anthropic also did not comment.)

What’s The Right Amount of Sycophancy?

Sycophancy can cause society-wide problems. Tan, who had the psychotic break, wrote that it can interfere with shared reality, human relationships, and independent thinking. Ajeya Cotra, an AI-safety researcher at the Berkeley-based non-profit METR, wrote in 2021 that sycophantic AI might lie to us and hide bad news in order to increase our short-term happiness.

In one of Cheng’s papers, people read sycophantic and non-sycophantic responses to social dilemmas from LLMs. Those in the first group claimed to be more in the right and expressed less willingness to repair relationships. Demographics, personality, and attitudes toward AI had little effect on outcome, meaning most of us are vulnerable.

Of course, what’s harmful is subjective. Sycophantic models are giving many people what they desire. But people disagree with each other and even themselves. Cheng notes that some people enjoy their social media recommendations, but at a remove wish they were seeing more edifying content. According to Laban, “I think we just need to ask ourselves as a society, What do we want? Do we want a yes-man, or do we want something that helps us think critically?”

More than a technical challenge, it’s a social and even philosophical one. GPT-4o was a lightning rod for some of these issues. Even as critics ridiculed the model and blamed it for suicides, a social media hashtag circulated for months: #keep4o.

I wrote a post a year ago about my “Do Now” routine — the first thing students do when they enter my class. It’s one of the most popular posts I’ve written, which I’ve always found funny. My routine is pretty boring and straightforward. Nothing flashy about it. Today I’m sharing an update. My routine is mostly the same, so my focus for this post is what I call the “active ingredients” in the Do Now — the features of the routine that I find most important. There are a few fun little tweaks from last year, though!

The Least Useful Instructional Time in a Lesson

Before getting into the routine itself, one important piece of framing. In my opinion, the Do Now is the least useful instructional time in a lesson. I have students coming in at different times. I need to take attendance. Two students don’t have pencils. Another student has an urgent issue they need to tell me about. I get a message from the front office asking me to send a student up to see them.

In every other part of my lesson, I can frame the activity for students. I can explain why we are doing something and what I want them to learn. I can actively monitor while students work. During the Do Now, those things are a bit harder. This might be different for you! If you work at a school where all students come in at the same time and you can reliably give 100% of your attention to learning from the moment they arrive to your class, great. That sounds awesome. That’s not the case where I work, and what I’m describing is true for most schools.

I don’t mean to say that the Do Now isn’t important. An effective Do Now routine is great for class culture. It sets the tone for the rest of class, and communicates to students that, in this class, we get right to work because we have a lot to learn.

I think the most common mistake teachers make with a Do Now is trying to do too much. It’s tempting to create a warmup activity that previews what students are about to learn, getting right to the heart of the lesson. I think that’s often a mistake. If you wait until every student is present, and attendance is taken, and all the little wrinkles are ironed out, you’ve wasted time and lost momentum. If you try to get students working right away, you can’t frame the activity smoothly and there’s a risk of confusing students. If you can’t give your full attention to learning, all those ambitions might be wasted.

The Routine

Here’s the routine. It’s dead simple.

Students walk into my room and pick up a half-sheet handout that looks like this:1

Then they sit down at their desks. I have a set of five problems projected at the front of the room. Here was a Do Now from last week:

Students answer the five questions on their handout. That’s it.

If you’d like to try this, here is a template for the handout, and here is a template for projecting the problems.

Active Ingredients

Here are what I’ll call the “active ingredients” in this routine — the elements that I think play the largest role in making the routine successful.

• I spend very little time prepping the activity. It takes me one minute to write the questions each day, and I copy and paper-cut the half-sheets in bulk once every two weeks or so.

• The routine is short. The Do Now is the least useful instructional time in a lesson. I aim for three minutes flat, from the start of class until I go over the questions. I quickly have students check their answers. Then we move on to the rest of the lesson. Short and sweet.

• I occasionally grade the Do Now to send a message that solving these problems is valuable practice and I expect students to remember what they learn. I grade them more often at the beginning of the year when I prioritize helping students build good habits, and I grade them less and less as the year goes on.

• All questions are retrieval practice on topics students have learned previously and are relatively straightforward. There will be time for more complex questions later when I have the chance to provide more scaffolding and support.

• I avoid putting more than one question on the Do Now that I’m uncertain students will get right. If half the class is getting two or three or four questions wrong, I screwed up. First, that’s not helping to build confidence at the start of class. Second, that’s too many questions to address effectively.

• I actively monitor (as much as possible while also taking attendance and dealing with whatever else I need to deal with). I circulate around the room once when the three minutes are almost up.

• When I circulate around the room, I pick the question I think students are most likely to struggle with and look at every student’s paper to see their answers. I find I can reliably do this for one question. For more than one question, it takes too long and the data is too hard for me to keep track of.2

• Based on that one question, I decide if we will follow up. If a significant number of students get it wrong, we talk a bit about the question and do some practice on mini whiteboards. I’ll then make sure to include that question for a few more days to see if students are getting it right or if I need to do a larger reteach.

• I have a spreadsheet to track which skills to include. Before I used this spreadsheet, I would often forget to include old skills or repeat the same skills day after day. The spreadsheet automatically schedules retrieval practice for each skill, beginning with several days of retrieval in a row before spacing out practice at larger and larger intervals. Credit to Lee Wheeler for this idea.3

That’s it! That’s all there is to it. I want to emphasize again, this seems really simple and unambitious. That’s the point. I’m not trying to do too much. I’m giving students a bit of quick retrieval practice, a confidence boost at the start of class, and gathering data for me to act on. Three minutes flat. Maybe we spend two more minutes on a common mistake. Then we move on to the rest of the lesson.

Is mathematical beauty real? Or is it just a subjective, human ‘wow’ that is becoming redundant in an AI age?

- by Rita Ahmadi

Read on Aeon

Rahim Nathwani is a technologist based in San Francisco, making businesses better with tech and AI. He believes almost everyone can and should learn more math. His 9-year-old son is on track to complete Algebra 1 before 5th grade.

There is a pattern in American math education reform. An organization claims that a new teaching approach produces dramatic gains. Schools and districts adopt it. Years later, when someone checks the evidence, the gains turn out to be overstated, the methodology turns out to be flawed, and the students who were promised a better education are worse off than before.

The organizations driving these reforms say they want equity. But when their evidence doesn’t hold up, the children who suffer most are the ones these reforms were supposed to help: kids from low-income families, whose parents can’t hire tutors to fill the gaps.

This pattern has played out at least three times with research associated with YouCubed, one of the most influential math education organizations in the country.

The Chart That Doesn’t Match the Data

YouCubed, a Stanford-based initiative led by Professor Jo Boaler, recently published a PDF claiming dramatic math score improvements in Healdsburg Unified School District in California. The chart showed a compelling story: scores starting very low, climbing steadily, ending impressively high. It’s the kind of result that gets shared in school board meetings and cited in policy decisions.

There was one problem: The numbers don’t match the public record.

California publishes its student test data openly. Anyone can look it up. For the most recent year, YouCubed’s chart and the state data roughly agree: about 75% of 5th graders at Healdsburg meet grade-level standards. But for the baseline year (the “before” picture that makes the improvement look dramatic), YouCubed’s chart shows approximately 24%. The state’s official data for the same district, year, and grade band shows approximately 44%.

Here’s the chart from the YouCubed PDF:

And here’s a recreation of the chart based on official government numbers:

That’s not a rounding difference. That’s a gap of 20 percentage points, and it runs in exactly the direction that makes the improvement story look better.

When you replace YouCubed’s baseline with the actual public data, the narrative changes completely. Scores were roughly flat for several years, then showed a recent uptick. That might be a legitimate result. But it is not the dramatic turnaround that YouCubed presented to educators and policymakers, and which Jo Boaler shared on X.

Maybe There’s an Innocent Explanation?

In education research circles, it’s common to see researchers exclude certain subgroups from their data. The most frequent justification is something like: “We excluded chronically absent students because they didn’t receive the full intervention.”

That’s a methodologically reasonable thing to do, as long as you say you’re doing it and explain why.

The YouCubed document makes no mention of any data exclusion. But let’s ask: could data exclusion explain the discrepancy?

Here’s the problem. If you exclude chronically absent students from a proficiency calculation, you would generally expect the remaining percentage to go up, not down. Students who miss a lot of school typically perform worse on tests. Take them out of the calculation, and the percentage of students who meet standards would likely increase, not decrease.

But YouCubed’s 2016–2017 figure (24%) is far below the state’s figure (44%). That means their exclusion would have had to somehow remove the higher-performing students from the calculation. That is the opposite of what “excluding chronically absent students” would do.

We cannot think of a legitimate methodological reason why a district’s proficiency rate would be nearly cut in half by a data exclusion. If there is one, Youcubed’s report should explain it. It doesn’t.

A Quick Explainer: Why Cohort Comparisons Matter

Even setting aside the data discrepancy, there is a deeper methodological problem with the chart — one that matters even if every number in it were correct.

The chart compares “5th graders in 2017” to “5th graders in 2023.” These are entirely different groups of children. The kids who were in 5th grade in 2017 are in their early 20s now. They are not the same kids who were tested in 2023.

Imagine you run a tutoring program. In Year 1, you tutor a class of students who happen to come from lower-income families and score an average of 60 on a test. In Year 5, a different group of students — from more affluent families, in a district that has added new staff — scores an average of 80. Can you claim your tutoring program caused the improvement? Of course not.

The right way to evaluate whether an educational intervention works is to track the same group of students over time and see how they progress. This is called a “cohort study.” It’s standard in medical research (think clinical trials), and it’s the gold standard in education research too.

We went ahead and pulled the cohort data from California’s public records. We tracked each group of students as they moved through the grades. What did that analysis show? Here it is:

It did not show consistent improvement across cohorts. The picture was much more mixed than the YouCubed chart suggests.

In fact, if we look at the most recent cohorts (the kids who entered 3rd grade in 2018 or later), we see that almost all cohorts (grades) have a downward trend: the chance that any given kid meets state standards in math goes down as they spend more years at Healdsburg school district.

This matters because the entire point of the YouCubed document is presumably to persuade: to show that the YouCubed approach works, and that other districts should adopt it. If the underlying analysis is methodologically flawed (and if the baseline data appears to be inaccurate) then that persuasion is built on sand.

The Study That Couldn’t Be Verified

This would be easier to set aside if it were an isolated mistake. It is not.

YouCubed’s influence traces back to Professor Boaler’s most celebrated research: the “Railside” study, published in 2008, which claimed that a reform-math approach at a California high school produced dramatic gains, especially for minority students. That study became enormously influential. It shaped curricula, frameworks, and professional development programs across the country.

In 2012, two Stanford mathematics professors and several colleagues published an extended critique. They raised serious concerns: the school had been misrepresented, the data couldn’t be independently verified, and the conclusions overstated what the evidence showed. Rather than release the underlying data for independent review, Professor Boaler characterized the critique as harassment.

The data was never made fully available. The study remains widely cited.

The Framework Built on Selective Evidence

In 2023, California adopted a new Mathematics Framework that drew heavily on Boaler’s ideas and YouCubed’s recommendations. The framework discouraged ability grouping, de-emphasized procedural fluency, and delayed access to advanced math courses — all in the name of equity.

More than 1,000 scientists and mathematicians signed an open letter objecting. Their concerns: the framework cited research selectively, dismissed evidence from high-performing countries, and would likely harm the students it claimed to help — particularly those aiming for STEM careers.

A Stanford mathematician who reviewed the framework’s citations in detail documented numerous cases where cited studies either didn’t support the claims being made or actively contradicted them. Misrepresented citations, like misrepresented data, are hard to attribute to innocent error when they consistently run in the same direction.

The Pattern

Three cases. In each one, research associated with the same organization claimed dramatic results. In each one, outside review found serious problems with the evidence. And in each one, the discrepancies ran in the same direction: making the reform approach look more effective than the data actually showed.

A single error is understandable. Research is hard, and mistakes happen. But when the mistakes consistently flatter the same conclusion, the most straightforward explanation is not bad luck.

This matters because education policy is downstream of evidence. When a district eliminates ability grouping, it cites research. When a state delays algebra, it cites research. When a school board adopts a new curriculum, it cites research. If that research is unreliable, every decision built on it is compromised.

Who Pays the Price

When a math reform fails, the cost is not distributed equally.

Families with resources respond to a weak math program the same way they respond to every institutional failure: they route around it. They hire a tutor. They enroll in a supplemental program. They move to a different district. And this last dynamic is perhaps the most destructive: declining public school enrollment means system-wide strain, fewer resources for schools and teachers, and eventual school closures. This phenomenon is showing up across the country, and it seems to be driven by wealthier, white, and asian families, especially those in deep-blue districts experiencing acute education dysfunction. Seattle has seen thousands of families leave public schools, and private schools in San Francisco are brimming with new applicants.

Families earning $200,000 a year have options and are evidently ready and willing to use them.

A family earning $40,000 does not. When school math instruction is the only math instruction a child receives, the quality of that instruction is the ceiling on that child’s opportunity. A low-income child whose school adopted a curriculum based on overstated evidence doesn’t get a do-over.

The cruelest irony is that these reforms are justified in the language of equity. The children they claim to champion are the ones with the fewest alternatives when the reforms don’t work.

What Should Be Different

None of this means that math education shouldn’t evolve, or that traditional approaches are beyond criticism. But it does mean that the organizations driving reform need to meet a basic evidentiary standard: the same standard we’d expect of a pharmaceutical company claiming its drug works.

That standard is straightforward. When you publish results, the underlying data should be available for independent review. When your chart shows a number, it should match the public record. When you cite a study, the study should actually support the claim you’re making. And when someone finds an error, the response should be a correction, not an accusation.

These are not hostile demands. They are the minimum that parents, teachers, and policymakers deserve before reshaping how children learn mathematics.

One bad chart would be forgivable. But what we’ve described here is a pattern: influential organizations are shaping how millions of children learn math based on research that hasn’t survived basic scrutiny, and the kids who pay the price are the same ones the reforms were supposed to help.

Your child’s math education deserves evidence that holds up when someone checks.

Related Articles