Publications

Published

"What large language models know and what people think they know (2025)"
▼

Steyvers, M., Tejeda, H., Kumar, A., Belem, C., Karny, S., Hu, X., Mayer, L. W., & Smyth, P.

Nature Machine Intelligence

OSF Link

Have you ever noticed that ChatGPT speaks like a know-it-all?
In this paper, we show that there is misalignment in the confidence LLM's are perceived to have in their answers, and what their internal confidence values reflect. In essence, LLM's routinely appear more confident than they should, despite model weights accurately reflecting the uncertainty associated with the answer. We also show how prompting LLM's to include appropriate uncertainty language in the answer text diminishes this gap between perceived and actual confidence considerably.
"Learning with AI Assistance: A Path to Better Task Performance or Dependence? (2024)"
▼

Karny, S., Mayer, L. W., Ayoub, J., Song, M., Su, H., Tian, D., Moradi-Pari, E., & Steyvers, M.

Proceedings of the ACM Collective Intelligence Conference

Link

Is receiving AI assistance on tasks degrading our skills, or improving them?
To answer this question, we developed a gamified task in which participants needed to intercept moving, point-valued targets before they escape. During the game, an AI assistant suggests optimal actions to take to maximize player score. We manipulated when people were given this assistance to see whether its presence harmed or boosted their learning during the task. Despite many media fears of skill degradation and/or hopes of AI-boosted learning, we find that in our task, AI assistance neither harms nor enhances learning. This finding highlights the possibility that there exist a subset of tasks in which AI assistance can be provided risk-free, which we hypothesize may include many tasks that do not require higher-level reasoning.
"Optimeet: A computational tool to enhance participant attendance in group research (2025)"
▼

Mayer, L. W., Bocheva, D., Hinds, J., Brown, O., Piwek, L., Ellis, D. A.

Behavior Research Methods

OSF Link

Do your studies rely on groups of people?
Recruiting participants for group sessions can be arduous, often costing a lot of time, effort, and resources. In this work, we demonstrate the complete lack of tools that can appropriately allocate a sample of interested participants to a set of group sessions. We then mathematically derive metrics that can quantify the effectiveness of any tool attempting to solve this problem. Finally, we develop an algorithm that outperforms any existing utility on this task using simulation and a large-scale pre-registered user study. Our scheduling utility is free and open-source, available to anyone through a web-applet.

Posters and Talks

“A Cognitive Framework for Strategic AI Communication”

Mayer, L. W., Steyvers, M. CogSci 2025

In Progress

“Human-AI Collaboration: Trade-offs Between Performance and Preferences”

Mayer, L. W., Karny, S., Ayoub, J., Song, M., Tian, D., Moradi-Pari, E., Steyvers, M. (under Review) Cognitive Research: Principles and Implications

How can we make AI collaborate well with people? Narrowly optimizing the performance of AI agents may be convenient, but can cause frustration when people are then asked to work with this agent. In this paper, we show that people prefer AI agents that are considerate of their preferences, even when this comes at the cost of performance. We also find that certain human-centric design choices boost people’s liking of the agent, without harming the performance of the human-AI team. Our results strongly suggest that leveraging both subjective and objective metrics is crucial when designing AI agents for human collaboration.

OSF repository

Research Interests

At heart, I study human decision-making. I have a particular interest in using Cognitive Models as a means of specifying the causal processes that give rise to people’s choices. Human-AI collaboration has become an extremely convenient outlet for me in this regard. For one, contemporary AI technologies’ applications in the real world offer plenty inspiration for my research. For example, I’m interested in how people trade-off the costs of outsourcing effortful cognitive labor to a computational agent, and how far you can push the attributes of this agent before eliciting a behavioral shift in the participant. Moreover, because Cognitive Models are mathematical in nature, casting my research through a lens of human-AI collaboration has immediate practical implications as solutions to “misaligned” behaviors are more readily generated from formal explanations of human behavior. In my research, the “AI” primarily serves as a theoretical tool to ask questions like: «How would people behave if they had access to an agent with X, Y properties?». That being said, I have a definitive interest in demonstrating the applied value of my work, meaning that some of my projects go the extra length of combining psychological models with algorithms from AI research.

The project that currently enjoys most of my attention intentionally pairs participants with an incredibly annoying to use “AI” assistant to understand how the timing of AI-induced interruptions influences the participant’s motivation to permanently disable this agent. With a causal model of people’s annoyance decisions integrated, our simulations show it should be possible to estimate policies for the AI assistant that avoid interrupting the participant at times where advice is unlikely to be appreciated. Such policies could lead to much more sustainable and productive collaboration.

Publications

Published

Have you ever noticed that ChatGPT speaks like a know-it-all?

Is receiving AI assistance on tasks degrading our skills, or improving them?

Do your studies rely on groups of people?

Posters and Talks

In Progress

Research Interests