Tianyi Alex Qiu | Computing Human Ideal Preference, Breaking AI Feedback Loops
- Date
- 2026-12-18
- Host
- Foresight Institute Calendar
About this event
AI systems are increasingly trained on human feedback, but what happens when the feedback itself is incomplete, distorted, or trapped in a loop? Tianyi Alex Qiu | Computing Human Ideal Preference, Breaking AI Feedback Loops is a conversation for people who want to think more carefully about how AI systems learn what we want, and how those systems might drift when optimization starts reinforcing the wrong signals. If you care about autonomy, alignment, and the real-world consequences of feedback-driven AI, this event gives you a focused place to engage those questions in person. About the Event This event centers on a big and timely idea: how to compute or infer human ideal preference rather than relying only on surface-level reactions, clicks, ratings, or other narrow forms of feedback. In practice, many AI systems are optimized using signals that are convenient to collect but may not fully represent what people would endorse after reflection. That gap matters, especially as AI tools become more capable and more embedded in everyday decisions. Rather than treating feedback as a simple objective, this session explores what it means to model preferences more deeply and how to avoid the kinds of feedback loops that can arise when systems learn from outputs that they themselves helped shape. That makes the topic especially relevant for anyone thinking seriously about AI alignment, recommender systems, preference learning, and human-AI interaction. The format is best understood as an in-person intellectual gathering built around a featured talk and discussion. Expect a setting that supports both learning and exchange: a chance to hear a clearly framed perspective, then test ideas through questions and conversation with others who are actively thinking about similar problems. What to Expect You should expect a program built around the core theme in the title: computing human ideal preference and breaking AI feedback loops. The central session will likely examine why standard feedback channels can be misleading, how optimization can amplify these distortions, and what alternative approaches might better capture what people actually value. The discussion may touch on questions like: What is the difference between observed preference and ideal preference? How do AI systems create or intensify feedback loops? Why can optimizing for engagement, ratings, or approvals fail to represent genuine human goals? What frameworks help preserve human autonomy while still making AI systems useful? How should builders and researchers think about feedback in high-stakes or socially sensitive settings? Because the event is tagged with community and networking, you can also expect this to be more than a one-way presentation. There should be room for thoughtful audience engagement, whether that means Q&A, discussion before or after the talk, or informal conversations with people working through related ideas from different angles. This is the kind of event where the value comes not only from the featured content, but from the quality of the questions in the room. If you like events where people move past the obvious framing and get into the assumptions underneath the technology, this should feel worth your time. Why Attend If you work with AI systems, this event offers a sharper lens on a problem that often hides in plain sight. Feedback is everywhere in modern machine learning, but the leap from "what users did" to "what users truly want" is far from straightforward. Spending time on that distinction can improve how you think about system design, evaluation, safety, and long-term impact. If you are more conceptually oriented, the event offers something equally valuable: a structured way to think about autonomy in the age of optimization. Many conversations about AI focus on performance and capability. This one points toward a more demanding question: how do we build systems that respond to humans without flattening, steering, or recursively distorting human judgment? You may leave with: A clearer mental model of feedback loops in AI systems Better language for discussing preference learning and alignment New questions about how autonomy should be preserved in human-AI systems Insight into where current approaches may be too simplistic Stronger connections with others interested in serious AI discourse For researchers, builders, and curious thinkers alike, the value is in getting more precise about one of the hardest parts of AI: not just making systems responsive, but making them responsive to the right thing. Practical Details This is an in-person event taking place on Friday, December 18 at 10:00 AM PST. If you are deciding whether to attend, plan around a live, on-site experience rather than a remote session. The in-person format should make it easier to ask questions, meet other attendees, and continue the conversation beyond the main presentation. The event is especially well suited to people who enjoy focused discussions rather than broad introductory overviews. You do not need to come in with all the answers, but you will get more out of it if you arrive ready to think critically about AI, feedback, and human values. Given the subject matter, it is worth showing up prepared with a few concrete questions of your own. Think about where you have seen feedback loops in products, research, or public discourse, and what it would mean to design around deeper human preferences instead of only measurable proxies. If this topic sits at the intersection of your interests in AI, autonomy, and the future of human-centered systems, this event is a strong reason to spend a morning in a room with people asking better questions.
Who should attend
This event is for people who want a more rigorous conversation about how AI systems learn from humans, and what goes wrong when feedback becomes the target instead of a signal. - You work in **AI or machine learning** and want to think more carefully about preference learning, alignment, evaluation, or human feedback beyond the usual surface metrics. - You are a **researcher, student, or independent thinker** drawn to questions about autonomy, ideal preference, and the limits of current feedback-driven systems. - You build or study **products shaped by engagement, ranking, recommendation, or optimization**, and you want better frameworks for spotting harmful loops early. - You care about **human-AI interaction** and want to understand how systems can influence, narrow, or distort the very preferences they are supposed to serve. - You enjoy events where the room includes people who are ready to ask precise questions, challenge assumptions, and explore difficult ideas without oversimplifying them. - You are looking for **thoughtful community and networking** around serious AI topics, especially with others who care about both technical design and human consequences. If you have ever felt that "user feedback" is treated as too simple a concept in AI conversations, this event will likely feel directly relevant to you.