Why Simulation Is Becoming a Serious Decision Tool
For years, simulation has been common in fields where making the wrong decision can be expensive. Pilots train in simulators. Engineers model systems before building them. Businesses test financial scenarios before committing capital.
Education has traditionally had far fewer opportunities to do the same thing.
That may be changing.
One of the most interesting developments comes from Stanford University, where researchers have been investigating whether artificial-intelligence agents can simulate the attitudes and behaviors of real people. That research has now helped produce a commercial company, Simile, focused on large-scale human-behavior simulation.
It started with Stanford research
In 2023, Stanford researchers including Joon Sung Park, Michael Bernstein, and Percy Liang developed what became widely known as the “Smallville” experiment: a simulated community populated by generative AI agents capable of remembering experiences, planning activities, reflecting, and interacting with one another.
The work suggested something important. Large language models might eventually do more than answer questions or generate text. They might help model how people respond to situations and to one another.
The research then became much more ambitious.
Stanford researchers created AI agents representing 1,052 real individuals, grounding each agent in extensive interviews with the person it represented. They then compared the agents’ answers with the actual participants’ responses on established surveys and experiments. Stanford reported that the simulated agents reproduced participants’ General Social Survey responses at about 85% of the accuracy with which people reproduced their own answers two weeks later.
That does not mean AI can perfectly predict human behavior. Stanford itself emphasizes that these systems still have important limitations and require careful validation, consent, and safeguards. But the research provides evidence that meaningful behavioral simulation is becoming technically possible.
From the laboratory to Simile
The connection to Simile is direct.
Simile says its founders pioneered AI-based simulation through their academic work at Stanford. Joon Sung Park is Simile’s co-founder and CEO; Michael Bernstein is co-founder and Chief Data Officer; and Percy Liang is co-founder and Chief Scientist.
The commercial idea is straightforward: instead of merely forecasting a single outcome from historical averages, create simulated populations grounded in human data and ask what if?
What if a message changes?
What if a service changes?
What if a policy changes?
How might different groups respond?
Simile says its agents are grounded in information from real people—including structured interviews, prior choices, and behavioral signals—and then assembled into populations that can be used to compare possible decisions before those decisions are implemented.
That approach has attracted considerable investment. Bain Capital Ventures describes Simile as turning cutting-edge human-simulation research into an enterprise platform and said in July 2026 that it was continuing to invest in the company.
Why this matters for education
The important lesson for schools is not that a simulated student can replace a real student, teacher, counselor, or administrator.
It cannot.
The opportunity is to give decision-makers another way to examine assumptions before committing resources.
A district might ask:
If more students begin using structured homework support, how many will continue?
If peer participation increases, what might happen to engagement?
If attendance intervention occurs earlier, how might outcomes differ?
Which assumptions matter most to the projected result?
What happens over several semesters instead of only one point in time?
Those are simulation questions.
And they are particularly valuable because school decisions involve people. Participation varies. Students stop and restart programs. Benefits can persist or fade. Multiple interventions can interact over time.
A useful simulation should attempt to represent those realities rather than simply applying one percentage improvement to an entire student population.
Where Decision Support Labs fits
This is the direction we are exploring with Decision Lab.
Our approach is different from Simile’s, and we are not affiliated with Stanford or Simile. But the larger technological direction is closely related: combine real data, behavioral assumptions, simulation, and eventually observed intervention-response data so that school leaders can test strategies before investing in them.
We begin with public California education data and established research. Our experimental work then models synthetic students over time, including student response, follow-through, continuation, stopping, returning, and changing outcomes.
We are also beginning to ask educators directly what they see in schools. Their professional judgment can help us identify assumptions that public datasets simply cannot answer.
And as student systems such as Peer-to-Peer Planner and Homework Habit generate real intervention-response information, those observations can help us compare:
What we modeled → What actually happened → What we learned → What we should test next.
That learning loop may ultimately be the most valuable part of simulation.
A new decision tool—not a crystal ball
Simulation should not be presented as certainty.
It is better understood as a structured way of asking:
If these assumptions are reasonable, and we take these actions, what might happen next?
Stanford’s research shows why that question is becoming increasingly interesting. Simile shows that serious investors and companies believe human-behavior simulation can become commercially useful.
For education, the opportunity is to bring the same discipline to one of the most consequential decisions of all:
How do we create better outcomes for students before valuable time, resources, and opportunities are lost?
Disclosure: Decision Support Labs is not affiliated with, sponsored by, or endorsed by Stanford University, Simile, or Bain Capital Ventures. References are provided to discuss independent research and industry developments.
Sources for the post
Stanford HAI’s explanation of its 1,000-person generative-agent research: Stanford HAI — Simulating Human Behavior with AI Agents
Simile’s description of its Stanford research origins and founders: Simile — Research and company background
Bain Capital Ventures on the Stanford research lineage behind Simile: Bain Capital Ventures — The Human Layer of AI