RAI Workshop Series – Are AI models telling us what we want to hear? Exploring sycophancy in Large Language Models
-min.avif)
The Responsible AI Workshop Series
In this workshop, we will discuss the phenomenon of AI sycophancy, a problem which has been documented by researchers and end-users alike. AI sycophancy is characterized by excessive ingratiating behavior with the goal of pleasing the user, and can be problematic when it sways models from initial judgments or leads to lower-quality responses. We will discuss the possible causes of sycophantic behavior in AI systems, tease apart the distinction between sycophancy and simply aligning models with user preferences, and introduce different methods for quantifying sycophancy in LLMs. We will also talk about the effects of sycophancy on rationality in language models, and how these effects can be mitigated by finetuning.
By the end of this workshop, you will:
- Understand what sycophancy means in the context of LLMs and why it matters
- Learn how to measure sycophancy, including our framework for studying its impacts on rational decision-making
- Explore practical mitigation strategies (calibration and post-training)
- Discuss implications for human–AI collaboration and future research
Come ready to learn and discuss AI sycophancy and its potential harms, how we can quantify it, and how we can reduce its effects. We will also have an open discussion about the impacts of sycophancy in AI: does it always represent a net negative?
This workshop is brought to you by the Responsible AI Practice as part of the RAI Workshop Series.
The RAIP Workshop Series is held in person at 177 Huntington Ave, Boston and is designed as an interactive gathering with a small but diverse audience. Previous workshops have brought together students, faculty, researchers, AI entrepreneurs, and NGO representatives from both technical and philosophical backgrounds. The goal is to spark dialogue and surface a wide range of perspectives on the responsible AI challenges under discussion. Sessions take place on Thursdays at 3:00 PM and typically include a designated commentator who contributes reflections either intermittently or toward the end of the talk.
Want to hear about future Responsible AI events?
Sign up here to get updates on upcoming Responsible AI workshops and events.
Commentator
Keynote and Industry Speakers
Northeastern University Speakers
Agenda
.avif)
.avif)
.avif)
.avif)
-p-800.avif)
%20circ.avif)
.avif)
