Experiment
Policy Flight Simulator
Let policy teams explore how different actors and system constraints might respond dynamically to alternative rules.
01
Problem to address
Static impact assessments struggle with feedback loops: people adapt, operational capacity becomes constrained, markets respond and one intervention changes the next.
02
What the evidence tells us
People need to understand what might happen next, but the evidence for forecasting is thinner than the evidence for spotting what needs attention now. This simulation approach is still speculative.
03
The proposition to explore
The idea
Build toward a policy flight simulator that lets teams explore how different actors, constraints and feedback loops could respond to alternative policy choices over time.
04
Hypothesis
If Defra can represent heterogeneous actors and feedback loops credibly, policy teams can explore system trade-offs before committing major public expenditure.
05
How it could work
Represent different actor types, constraints and responses. Change a policy rule and simulate multiple plausible trajectories over time. AI explains why outcomes differ and which assumptions drive the result.
06
Example: bring it to life
Payment +20% produces higher uptake but much higher verification workload; geographically targeted eligibility produces lower uptake but stronger environmental return. Teams explore the trade-off before piloting either option.
07
AI opportunity
Agent simulation, model orchestration, natural-language interaction and explanation. AI helps build and interrogate scenarios; simulation validity depends on behavioural and operational models.
08
Potential value
- Better understanding of dynamic consequences
- More informed capacity planning
- Earlier challenge of perverse incentives
- Potential reduction in expensive policy trial-and-error
09
Reduce uncertainty without overcommitting
Smallest meaningful experiment
Start with a toy model for one narrow process where behaviour and capacity are measurable. Validate the simulator against known historical changes before using it prospectively.
10
Open questions and risks
What we still need to learn
- High risk of false precision
- Behavioural models may be weak
- Should never replace real-world experimentation