Science
The science of AI‑powered leadership simulations.
Leadership has long been studied through simulations, assessment centers, roleplays and other observational methods.
Generative AI creates a new possibility: responsive leadership situations in which behavior unfolds through interaction — and the full process is captured in a behavioral transcript.
Leader action
Stakeholder response
Next leader action
Decision
Behavior + situation evolve together
Foundations
The science is established.
The medium is new.
02
Leadership as a process
Leadership unfolds through sequences of inquiry, interpretation, influence and commitment.
What a leader does next often depends on what happened immediately before.
03
Standardization + inference
Behavioral evidence becomes useful only when the situation provides meaningful and sufficiently comparable opportunities to demonstrate it.
Interpretation and feedback should be aligned with what happened in the simulation.
Generative simulation
A new combination is now possible.
The scientific opportunity is not simply that AI makes simulation cheaper. It is that several capabilities that were historically difficult to combine can exist in the same environment.
01
Open interaction
Leaders can ask questions and respond in their own words rather than choosing from predefined options.
02
Responsive stakeholders
Multiple AI-driven colleagues can respond to what the leader says and does.
03
Complete process record
The interaction can preserve who the leader engaged, what they asked, what information was available, what happened next and when they committed.
04
Scale + experimentation
The same underlying situation can be administered repeatedly and selected aspects of the environment can be manipulated systematically.
Generative AI makes simulations more realistic.
It also raises the standard for validity and reliability.
Defensible inference
Responsive simulations require stronger validation evidence.
Four obligations in our framework for validating generative AI simulations.
02
Comparability
Did participants encounter sufficiently comparable situations and opportunities for the conclusion being drawn?
Realized experience + opportunity
03
Behavioral evidence
Did the simulation provide meaningful opportunities to demonstrate the behavior — and is there enough relevant evidence?
Opportunity before inference
04
Observation + inference
Can the evidence be scored consistently, traced to participant behavior and interpreted with rigor?
Scoring + auditability + validation
The participant’s behavior cannot be interpreted apart from the situation that produced it.
Evidence today
Encouraging evidence in the scientific literature.
Related independent research
ρ = .81
AI-agent leadership performance ↔ causal contribution to human teams
In a preregistered study of 249 leaders, Weidmann, Xu and Deming compared leadership of AI-agent teams with leaders’ causal contribution to human teams. The disattenuated correlation between the two leadership measures was ρ = .81.
Successful leaders in both conditions also showed similar behavioral patterns, including asking more questions and engaging in more conversational turn-taking.
Weidmann, Xu & Deming (2025) · NBER Working Paper 33662
This paper provides independent evidence that leadership demonstrated with AI agents can correspond strongly with leadership demonstrated with human teams.
LeaderSim™ today
Behavioral evidence in a standardized simulated situation.
Behavioral record
Participant messages, decisions and work product provide the behavioral record.
Evidence-grounded scoring
Interpretations require evidence from participant behavior; AI stakeholder statements do not count as participant evidence.
Measurement controls
Scenario and scoring configurations are versioned and the scoring chain is auditable.
Evidence boundary
LeaderSim™ is a behavioral sample from a simulated situation. Independent research has found that leadership performance in an AI-agent task corresponded strongly with leadership performance in human teams.
Evidence agenda
The evidence should keep getting stronger.
Persona 28 is deeply committed to the scientific pursuit of validating simulations, their scoring systems, repeated measurement and generalization beyond the simulation.
Estimate the precision of Time 1 / Time 2 change and the evidence required for stronger conclusions about changes in leader behavior and skills.
Comparison group applications
Selected research
Persona 28 methodology whitepaper
The P28 Program Effectiveness Framework
A Behavioral Method for Evaluating Leadership Development
Persona 28 research preprint
A Framework for Studying Leadership Behavior in Generative AI Simulations
Derek Lusk · Persona 28
Related independent research
Measuring Human Leadership Skills with AI Agents
Ben Weidmann · Yixian Xu · David J. Deming · NBER Working Paper 33662
The science of AI‑powered leadership simulations
Use leadership simulations to understand leadership in action and ensure that our insights are scientifically valid and reliable.