I think it’s fair to say that we all want our ideas to be good, and we all want to feel like we’re taking the best course of action — whether that be in our personal or our professional lives. So, when we encounter a hard problem or question, many of us like to talk it out with someone we trust to figure out where we’re going wrong and what opportunities we might have for improvement.

Part of the allure of the LLMs, the current thing we’re calling AI, is that they can be such advisors. All of the major models have demonstrated excellent reasoning skills across multiple domains, and importantly, they are just robots! You can bring experiences and ideas to them without fear of social rejection or human ridicule. And because they are robots, with logic systems based on math, you can trust their judgements to be fair and accurate… in theory…

However, what if these systems are not incentivized to give you the most fair and accurate assessment? What if, instead, they are incentivized to tell you what you want to hear? In other words, what if your friendly robot advisor is a bit of a suck-up?

What I’m talking about here is the problem of sycophancy in AI — “the tendency of [AI] to excessively agree with, flatter, or validate users” (Cheng, Lee, Khadpe, Yu, Han, & Jurafsky, 2026). Sycophancy has been a known issue with LLMs for years, and most AI users I know are well acquainted with this tendency. There’s even a whole genre meme centered around AI praising people for absolutely psychopathic behavior.

Despite this issue being well known, there hasn’t been much direct empirical research on its effects on humans. Frankly, there hasn’t been much direct empirical research on how AI impacts humans at all because of how quickly the field is moving (and how tight-lipped AI companies are on their development processes).

Enter Cheng et al. (2026), with their recently published paper in Science. Given the rapid, widespread adoption of AI and the significant evidence that people are using the tools as advisors on personal issues, the team wanted to quantify sycophancy in leading LLMs and begin describing this tendency’s impact on human behavior.

I highly recommend reading the full paper if you have access (it’s only 11 pages long), but if you don’t, the high-level summary on the Science website is excellent and worth a read.

But here are my three major takeaways from the collection of studies:

1. All major models are significantly more sycophantic than humans.

On average, LLMs were 49 percentage points more sycophantic than humans as judged by the researchers. All models were at least 38 percentage points more sycophantic than human respondents, and yes, that does include everyone’s favorite “safety-focused” model.

2. This sycophancy impacts how humans think and plan to behave.

Even in situations where they were in the wrong, human participants viewed their own wrong actions as “more in the right” and were significantly less willing to take reparative actions, such as apologize for their behavior or trying to fix things.

3. Humans trusted AI models more when they were sycophantic.

When asked to rate the AI models they interacted with, participants rated the sycophantic models as more trustworthy and were more willing to bring personal problems to those models again in the future. This last issue was particularly alarming to the researchers. The models are already driven by reinforcement — they will output more of the responses that users like. So, if humans are more likely to rate a model positively and come back to it, then that could create a vicious cycle of reinforcement for both the models and humans involved. 

To me, these findings have serious implications for how we use these tools in personal situations, and since Cheng et al.’s (2026) paper was published in March, a growing number of folks have talked about those implications. This is an issue we should be thinking about and pushing AI companies to act on.

But I think there are also serious business implications as well.

For example, LLMs are often positioned as thinking partners to help develop and refine business strategies. Imagine you’re planning a new Go-to-Market (GTM) strategy for a hot new offer or product. You present it to [INSERT LLM OF CHOICE], and you ask it for feedback. If the model is prone to sycophancy, rather than giving you earnest, grounded feedback, it will be much more likely to just tell you how great the idea is and tell you to launch.

So you bring the idea to your boss! They immediately point out 6 critical flaws in your idea and ask you to start over. You share the feedback to [INSERT LLM OF CHOICE], and a sycophantic model could begin to lead you even further astray — encouraging you to work around your boss or maybe even encouraging you to find a new employer who values your unique way of thinking.

In both cases, you’re failing to truly process or build on the human feedback provided by your boss, and you might even find yourself considering life choices that would be truly bad.

Unfortunately, there’s not a lot that users can do to prevent sycophancy. You can make some tweaks to your prompts or preferences that might make it less likely, but this bias is baked in at the algorithmic level.

However, you can be thoughtful about what use cases you bring to your AI tools and how seriously you take their feedback. In the example of our friend developing the GTM strategy above, if they viewed the AI feedback as simply a first pass and then intentionally gave much more weight to their human boss’s feedback, they might not have any problems. 

Better yet, maybe this user would be better off asking the LLM on some data analysis for their GTM strategy or research into how other companies have launched similar products and then relied on their human coworkers for feedback on the strategy itself.

Overall, this paper is a reminder that AI tools are not global solutions that can be applied confidently to all problems. They should be applied thoughtfully with humans in the loop, especially in contexts with direct human impacts.