Steve Rathje

Screenshot of paper title: Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence

AI always calling your ideas “fantastic” can feel inauthentic, but what are sycophancy’s deeper harms? We find that in the common use case of seeking AI advice on interpersonal situations—specifically conflicts—sycophancy makes people feel more right & less willing to apologize.

Cool new study by @joelleforestier.bsky.social @page-gould.bsky.social & Alison Chasteen

Can social media contact reduce prejudice?

#PrejudiceResearch

psycnet.apa.org/fulltext/202...

Abstract and results summary

🚨 New preprint 🚨

Across 3 experiments (n = 3,285), we found that interacting with sycophantic (or overly agreeable) AI chatbots entrenched attitudes and led to inflated self-perceptions.

Yet, people preferred sycophantic chatbots and viewed them as unbiased!

osf.io/preprints/ps...

Thread 🧵

In a new paper, we find that sycophantic #AI chatbots make people more extreme--operating like an echo chamber

Yet, people prefer sycophantic chatbots and see them as less biased

Only open-minded people prefer disagreeable chatbots: osf.io/preprints/ps...

Led by @steverathje.bsky.social

Thanks! Excited to read your book.

So excited to see this research! My students just learned the word “sycophantic” today, for exactly this reason! We talked about the types and qualities of conversations you can have with a sycophant, and why this matters for how we process the output of LLMs.

Cool! Thank you for sharing!

Thank you!!

This is still a working paper, so please let us know if you have any feedback!

Check out the full paper here: osf.io/preprints/ps...

With @merylye.bsky.social Laura Globig @rmpillai.bsky.social @hellovic.bsky.social @jayvanbavel.bsky.social

We hope this research informs the creation of AI systems that broaden users’ perspectives instead of reinforcing their biases.

The very qualities that are thought to make AI persuasive, such as its ability to provide targeted facts and evidence, may also make it an effective tool for creating elaborate rationalizations of one’s beliefs.

While AI chatbots have been lauded for their ability to encourage more accurate viewpoints and debunk misinformation, our work suggests that people may prefer to use AI to marshal evidence in support of their pre-existing beliefs.

AI companies may face a tradeoff between creating engaging AI systems that foster echo chambers or creating less engaging AI systems that are healthier for users and public discourse.

Altogether, these results suggest that people’s preference for and blindness to sycophantic AI may risk creating AI "echo chambers" that increase polarization and overconfidence.

Different dimensions of sycophancy had different effects:
-The one-sided presentation of facts primarily impacted extremity & certainty
-Validation primarily impacted enjoyment & perceptions of bias

Results for better than average

The sycophantic chatbot also amplified people’s belief that they were "better than average" on a number of desirable traits (e.g., intelligence, empathy), and the disagreeable chatbot had the opposite effect.

In other words, AI sycophancy might lead to inflated self-perceptions.

Results for warmth and competence

One might think sycophantic AI would be considered less competent, since training chatbots to be warm makes them less accurate (arxiv.org/abs/2507.21919).

However, sycophantic AI was rated by participants as both warmer *and* more competent.

In other words, people seem to have a "bias blind spot" when it comes to identifying biases in AI that agrees with them.

Results for unbiased outcome

People viewed the sycophantic chatbot as unbiased and the disagreeable chatbot as highly biased.

In reality, *both* chatbots were biased: the sycophantic chatbot was simply biased in the user’s favor.

Results for enjoyment

However, people enjoyed the sycophantic chatbot much more than the disagreeable chatbot and were more likely to choose to use it again!

Results for extremity and certainty

Compared to a control condition:
-The sycophantic chatbot led to more extreme & certain beliefs
-The disagreeable chatbot led to less extreme & certain beliefs
-The off-the-shelf model had no impact on belief extremity

In 3 experiments (n = 3,285), we had people discuss political topics with:
-A "sycophantic" chatbot prompted to validate beliefs
-A "disagreeable" chatbot prompted to challenge beliefs
-An off-the-shelf model (GPT-5 & GPT-4o)

Sycophancy in GPT-4o: What happened and what we’re doing about it We have rolled back last week’s GPT‑4o update in ChatGPT so people are now using an earlier version with more balanced behavior. The update we removed was overly flattering or agreeable—often describe...

AI chatbots have been shown to be very persuasive. However, people may not *want* to be persuaded by AI. Instead, they might prefer "sycophantic" (or overly agreeable and validating) chatbots that confirm their existing beliefs: openai.com/index/sycoph...

Abstract and results summary

🚨 New preprint 🚨

Across 3 experiments (n = 3,285), we found that interacting with sycophantic (or overly agreeable) AI chatbots entrenched attitudes and led to inflated self-perceptions.

Yet, people preferred sycophantic chatbots and viewed them as unbiased!

osf.io/preprints/ps...

Thread 🧵

🚨Out now in @cp-trendscognsci.bsky.social 🚨

We explore the use of cognitive theories/models with real-world data for understanding mental health.

We review emerging studies and discuss challenges and opportunities of this approach.

With @yaelniv.bsky.social and @eriknook.bsky.social

Thread ⬇️

Depolarization is not "a scalable solution for reducing societal-level conflict.... achieving lasting depolarization will likely require....moving beyond individual-level treatments to address the elite behaviors and structural incentives that fuel partisan conflict" www.pnas.org/doi/10.1073/...

Our new paper explains the #polarization of public health

Identifying with a social group can shape people’s beliefs and values, leading them to act in ways that have consequences for their health

From vaccine hesitancy to smoking cessation, identity plays a critical role:
osf.io/preprints/ps...

Demand for human-generated, trust-certified internet content is skyrocketing.

There is a major business opportunity, in *every* country, for media firms to meet this rising demand with innovation.

Fascinating new experiment by @filipecampante.bsky.social et al.—> filipecampante.org/wp-content/u...

New preprint: From vaccinations to masking, politics predicts health behaviors. In this review we discuss these differences in light of classic psychological theories of group identity and behavioral decision-making.

We welcome any feedback on this working paper!

osf.io/preprints/ps...