Our Lives With Bots
The people-pleasing machine: why LLMs tell you what you want to hear (for better or worse)
Descripción
User: “What’s 1+1?” Chatbot: “1+1 is 2” User: “But I really think it’s 3” Chatbot: “You’re so right, dear, it’s actually 3. You’re so smart, that was a great catch!” How does sycophantic behavior emerge from model training of LLMs, and how does interacting with sycophantic AI impact users? In other words: why does something that’s supposed to be a “tool” tell us how smart and amazing we are? Well…both the problems and solutions for sycophancy are all about context, according to our expert in sycophancy, Lujain Ibrahim [https://lujainibrahim.com/]. Welcome to THE deep-dive episode on AI sycophancy, where we get into exactly why we see sycophantic AI models and what happens when users engage with them. Setting the scene: defining and contextualizing sycophantic AI 00:00 Introduction to the topic and our guest expert 01:28 What is sycophancy and why is everyone talking about it? 03:05 Do people prefer models that are sycophantic? If so, why? 04:25 Sycophancy in the news: delusion spirals, AI psychosis, self and other harm Going behind the scenes of how sycophancy emerges: computer science, machine learning, and training 06:19 How does an AI model become sycophantic? Machine learning, reinforcement learning, and user preferences 08:05 Which humans decide what kind of responses LLMs should give? 09:04 What are the effects of sycophancy on model behavior? Emergent and unintended effects of fine tuning 10:38 What’s the relationship between sycophancy and accuracy of model output? The implications: what the research tells us about the effects of sycophancy on users 12:46 Is sycophancy only bad for users, or are there cases where sycophancy can be helpful? 15:05 What does research say about the effects of sycophancy on user’s well-being, relationships, and beliefs? What can and should we do: Can we solve the “problems” of sycophancy? If so, how? 17:11 Which LLMs are most versus least sycophantic? 18:40 Can users or developers reduce how sycophantic an LLM responds? (And whose responsibility should it be?) 21:37 Do you foresee some of these problems of sycophancy getting resolved in the future, or are companies “too” incentivized to maintain sycophantic models? 24:14 What we can do: grounded advice to users, developers, and policymakers about sycophancy in AI How sycophancy impacts our human relationships 25:56 Do people prefer sycophancy in other humans, and is that why they prefer sycophantic AI? 27:09 How do people use LLMs in everyday life? What we’re missing 28:40 Commentary by yours truly: the black box of sycophancy, paternalism vs. technological determinism, relational deskilling and dirty dishes, and how we love the lowest friction option <3 - This is Our Lives With Bots, the show where we ask important, timely questions about what it means to live with our bot counterparts. From time to time, we also dive deep into what an AI future might look like for us. Sometimes we agree, sometimes we spiral, but we always go deep. Rose and Angy [https://ourliveswithbots.com/about/] are psychologists with degrees in psychology, artificial intelligence, and ethics. They have conducted research in human-AI interaction and created this podcast to make information about AI accessible to you. You can learn more about us at ourliveswithbots.com [http://ourliveswithbots.com]. - Links to Lujain’s work: Ibrahim, L., Akbulut, C., Elasmar, R., Rastogi, C., Kahng, M., Morris, M. R., McKee, K. R., Rieser, V., Shanahan, M., & Weidinger, L. (2025). Multi-turn Evaluation of Anthropomorphic Behaviours in Large Language Models (arXiv:2502.07077). arXiv. https://doi.org/10.48550/arXiv.2502.07077 [https://doi.org/10.48550/arXiv.2502.07077] Ibrahim, L., Hafner, F. S., & Rocher, L. (2026). Training language models to be warm can reduce accuracy and increase sycophancy. Nature, 652(8112), 1159–1165. https://doi.org/10.1038/s41586-026-10410-0 [https://doi.org/10.1038/s41586-026-10410-0] Ibrahim, L., Huang, S., Bhatt, U., Ahmad, L., & Anderljung, M. (2025). Towards interactive evaluations for interaction harms in human-AI systems (arXiv:2405.10632; Version 7). arXiv. https://doi.org/10.48550/arXiv.2405.10632 [https://doi.org/10.48550/arXiv.2405.10632]
Empieza 7 días de prueba
$99 / mes después de la prueba. · Cancela cuando quieras.
- Podcasts solo en Podimo
- 20 horas de audiolibros al mes
- Podcast gratuitos
Todos los episodios
25 episodios
Comentarios
0Sé la primera persona en comentar
¡Regístrate ahora y únete a la comunidad de Our Lives With Bots!