A Rigorous Test of AI's Ability to Simulate Human Behavior
Project Silicon is a large-scale collaborative research initiative assembling a diverse portfolio of brand-new, unpublished behavioral experiments, running AI simulations on all of them, and comparing simulated results against actual human data.
The project is developed in coordination with Management Science and is fully funded by institutional sources with no technology provider involvement.
Four Goals Driving the Project
Assess AI Accuracy
Determine whether AI can reliably simulate human responses across diverse, unpublished experiments.
Benchmark Models
Compare leading AI models—closed and open-source—on simulation accuracy.
Open-Source Toolkit
Release an open-source toolkit for AI-based experimental pre-testing.
Actionable Guidance
Provide practical guidelines for integrating AI simulation into experimental design.
Researchers Running New Behavioral Experiments
We're seeking researchers planning to conduct—but have not yet completed—behavioral experiments in the following fields:
If you expect to complete your experiment by January 2027, let us know!
Process & Timeline
Now – July 2026
Register & Share Materials
Register your interest. We assess suitability early—before you commit any time. Once finalized, share materials via our secure portal.
July – August 2026
Registered Report Submission
We compile the Project Silicon paper in registered-report format and submit to Management Science for review, with feedback from all participants before submission.
September 2026 – January 2027
Simulation & Experimentation
We run AI simulations across multiple models using your materials. You run your experiment as planned—nothing changes.
January – March 2027
Comparison & Publication
We evaluate how closely each AI model’s simulated responses match actual human data. Results are submitted back to Management Science.
Benefits of Participating
Co-authorship
All participants will be secondary co-authors on the Project Silicon paper, which reports the aggregate findings, model benchmarks, and recommendations across the full portfolio of experiments.
Your AI Results
Receive the full AI simulation output for your experiment—showing how multiple models predicted your subjects' behavior.
Additional Simulation Support
We'll provide additional simulations for your subsequent experiments. You'll also get access to our simulation toolkit.
Results Event
Join a dedicated results-sharing event at the 2027 Michigan Workshop on Large Language Models (May 2027, Los Angeles, CA).
Maintaining the Methodological Firewall
The validity of this project depends on a strict separation between human and AI results.
This includes working papers, slides, posters, and any results shared online. Pre-registrations and high-level descriptions are fine.
Do not submit design specs to an LLM. The AI models must have zero prior knowledge of your experiment.
Your Work Is Safe With Us
Only the core PIs have access to your experimental materials. No other participant will ever see your work.
All data is stored in a HIPAA-certified University of Michigan computing environment.
For the Project Silicon paper, we'll work with you to describe your experiment in a way that you are comfortable with.
Participation requires no financial commitment. All project costs are covered by institutional funding.
The AI simulation involves no human subjects. Your existing IRB protocol requires no amendment.