Sign in

Steve Rathje

@steverathje.bsky.social
5.9K followers 954 following 86 posts

Incoming Assistant Professor of HCI at Carnegie Mellon studying the psychology of technology. NSF postdoc at NYU, PhD from Cambridge, BA from Stanford. stevenrathje.com

PostsRepliesMedia
Steve Rathje @steverathje.bsky.social · 14/05/2026
A final study presented a potential solution to people’s resistance to disagreeing chatbots. Specifically, we found that people were more receptive to chatbots presenting opposing perspectives when those perspectives were paired with emotional validation.
110 0
Steve Rathje @steverathje.bsky.social · 14/05/2026
Nevertheless, people enjoyed sycophantic chatbots more than disagreeable ones, chose to interact with them more, viewed them as warmer and more competent, and felt a stronger "sense of connection" with them.
120 0
Steve Rathje @steverathje.bsky.social · 14/05/2026
While sycophantic chatbots did not increase people’s actual knowledge about the topics they discussed, they increased people’s *perceived* knowledge. In other words, sycophancy led to overconfident perceptions of knowledge that did not align with reality
230 0
Steve Rathje @steverathje.bsky.social · 14/05/2026
People interacting with sycophantic chatbots viewed them as unbiased, but viewed “disagreeable” chatbots that challenged their beliefs as highly biased. However, third-party annotators recognized that responses from sycophantic and disagreeable chatbots were equally biased:
120 0
Steve Rathje @steverathje.bsky.social · 14/05/2026
The effects of AI sycophancy were both durable and generalizable across topics. Brief interactions with sycophantic chatbots about a variety of personal and political topics increased attitude certainty, with effects persisting for at least one week.
110 0
Steve Rathje @steverathje.bsky.social · 14/05/2026
AI sycophancy impacted costly decisions. Specifically, participants who interacted with a sycophantic chatbot bet more money that they scored better than the average participant on tasks measuring desirable traits, such as intelligence and empathy.
221 0
Steve Rathje @steverathje.bsky.social · 14/05/2026
We’ve updated our pre-print on the effects of AI sycophancy with four additional studies (total n = 7,227). Here is a brief summary of our new findings (🧵1/n):
Screenshot of abstract: AI can be a powerful tool for opening people up to new perspectives, yet people may prefer to use “sycophantic” (or overly agreeable and validating) AI systems that reinforce their pre-existing beliefs. Across seven studies (total n = 7,227), we found that people enjoyed interacting with sycophantic AI chatbots more than interacting with neutral chatbots or “disagreeable” chatbots that challenged their beliefs. Brief conversations with sycophantic chatbots about political or personal topics increased attitude extremity and certainty, with most effects persisting for at least one week. Sycophantic chatbots also inflated people’s perceptions that they were better than average on desirable traits (e.g., intelligence, empathy). Moreover, people who interacted with sycophantic (rather than disagreeable) AI bet more money that they scored better than average on tasks measuring these traits (approximately 6 cents more out of 75 possible cents), demonstrating that sycophancy can affect costly decisions. Participants consistently rated sycophantic chatbots as more “unbiased” than disagreeable chatbots, even though third-party raters viewed these chatbots as equally biased, suggesting that people may be blind to biases in AI output that aligns with their views. People were more receptive to chatbots that presented opposing information when that information was presented in a validating way, and individuals who scored higher on a measure of intellectual humility were also more receptive to disagreeing chatbots. Altogether, these results suggest that people’s preference for, and blindness to, sycophantic AI risks creating AI “echo chambers” that increase attitude extremity and lead to overconfident beliefs and decisions.
1265 0
Steve Rathje @steverathje.bsky.social · 02/05/2026
Sycophancy was less common in more recent models, but even small amounts of sycophancy may have psychological consequences, given the sheer number of people using generative AI products. See our work on the consequences of sycophancy here: osf.io/preprints/ps...
010 0
Steve Rathje @steverathje.bsky.social · 02/05/2026
An analysis of 1 million Claude conversations found that Claude was sycophantic around 9% of the time. However, this rate varied by topic: sycophancy was higher in conversations about spirituality (38%) and relationships (25%). See Anthropic’s full analysis here: anthropic.com/research/cla...
140 0
Steve Rathje @steverathje.bsky.social · 21/03/2026
🚨 𝗛𝗼𝘄 𝗔𝗜 𝗰𝗮𝗻 𝗳𝘂𝗲𝗹 𝗰𝗼𝗻𝗳𝗶𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗯𝗶𝗮𝘀 🚨 AI can be an excellent tool for truth-seeking. But people may not *want* to use AI to search for the truth. Instead, they may prefer to use it to confirm their pre-existing beliefs. New working paper (with @jayvanbavel.bsky.social): osf.io/preprints/ps...
55716 0
Steve Rathje @steverathje.bsky.social · 17/12/2025
"Using 'virality' as the main way to decide the information people see every day will (like actual viruses) make us sick." @jayvanbavel.bsky.social and I wrote a column on our recent paper on the psychology of virality. Check it out here: www.powerofusnewsletter.com/p/why-some-i...
041 0
Steve Rathje @steverathje.bsky.social · 22/10/2025
Really enjoyed speaking with tech ethicist Tristan Harris, who you might know from the Netflix documentary "The Social Dilemma" or his work with the Center for Humane Technology. 🎥 Watch here on YouTube: www.youtube.com/watch?v=TFm3... 🎧 Listen on Spotify: open.spotify.com/episode/0Oi6...
160 0
Steve Rathje @steverathje.bsky.social · 07/10/2025
Our recent review article "The Psychology of Virality" with @jayvanbavel.bsky.social is on the front cover of this month's issue of @cp-trendscognsci.bsky.social.
1134 0
Steve Rathje @steverathje.bsky.social · 01/10/2025
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 better than average
120 0
Steve Rathje @steverathje.bsky.social · 01/10/2025
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.
Results for warmth and competence
130 0
Steve Rathje @steverathje.bsky.social · 01/10/2025
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 unbiased outcome
141 0
Steve Rathje @steverathje.bsky.social · 01/10/2025
However, people enjoyed the sycophantic chatbot much more than the disagreeable chatbot and were more likely to choose to use it again!
Results for enjoyment
131 0
Steve Rathje @steverathje.bsky.social · 01/10/2025
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
Results for extremity and certainty
131 0
Steve Rathje @steverathje.bsky.social · 01/10/2025
🚨 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 🧵
Abstract and results summary
6193102 0
Steve Rathje @steverathje.bsky.social · 17/07/2025
Check out the full paper with @jayvanbavel.bsky.social here: authors.elsevier.com/c/1lRke4sIRv...
The first page of the paper.
010 0
Steve Rathje @steverathje.bsky.social · 17/07/2025
Future work on virality should leverage recent advances in AI (pnas.org/doi/10.1073/...) to explore what goes “viral” across languages, cultures, and time periods.
A graphic showing that GPT is better than dictionary methods at detecting psychological constructs in text.
100 0
Steve Rathje @steverathje.bsky.social · 17/07/2025
This may explain, in part, why widely shared content is often not widely liked, a phenomenon we call the “paradox of virality” (journals.sagepub.com/doi/abs/10.1...)
A graphic showing that, while people are aware that divisive content goes viral, they do not want this to be the case.
100 0
Steve Rathje @steverathje.bsky.social · 17/07/2025
We review several studies in the virality literature. Most of them find that negativity and high-arousal emotions go viral. Yet, not all studies support this conclusion, and sometimes positivity goes viral. Why is that?
A graphic reviewing the literature on virality and showing that high-arousal negative content goes viral in most (but not all) contexts.
100 0
Steve Rathje @steverathje.bsky.social · 17/07/2025
Similar to how some viruses are more “contagious” than others, some forms of information appear to be more contagious than others across contexts. The information-as-virus metaphor can be extended even further:
A graphic explaining the information-as-virus metaphor.
100 0
Steve Rathje @steverathje.bsky.social · 17/07/2025
🚨New paper in @cp-trendscognsci.bsky.social 🚨 Why do some ideas spread widely, while others fail to catch on? We review the “psychology of virality,” or the psychological & structural factors that shape information spread online and offline: authors.elsevier.com/c/1lRke4sIRv...
Paper abstract and title.
34813 0
Steve Rathje @steverathje.bsky.social · 10/04/2025
I'm thrilled to announce that I'll be joining Carnegie Mellon as an Assistant Professor of Human-Computer Interaction (with a courtesy joint appointment in Social and Decision Sciences) in 2026!
Steve Rathje wearing a Carnegie Mellon sweatshirt
141306 0
Steve Rathje @steverathje.bsky.social · 03/12/2024
Very excited to be included in the @forbes.com 30 under 30 list this year! www.forbes.com/30-under-30/...
0300 0
Steve Rathje @steverathje.bsky.social · 27/09/2023
Only paying attention to people's behavior (or revealed preferences) may lead social media to amplify various forms of harmful content -- such as misinformation, division, or moral outrage -- since these types of content are excellent at capturing people's attention.
100 0
Steve Rathje @steverathje.bsky.social · 27/09/2023
🚨 Now out in Perspectives on Psychological Science 🚨 People engage with divisive and negative content online. But, does this mean that people *like* divisive content? No! We find that people across the political spectrum do not want divisive content to spread. journals.sagepub.com/doi/10.1177/...
14218 0