A research game about trade-offs of Responsible AI
Your team is building an AI model to improve the prioritisation of which businesses are inspected.
The model should be effective — identifying where risk is most likely to occur, and making inspections more targeted. It should also be explainable to the people who use it and to the outside world.
You should be aware that scrutiny of algorithmic decision-making is rising. There's a national Algorithm Register where you have to register the model. Oversight queries and public information requests are becoming more common, and no one can say exactly when or how hard they'll hit. You can't control that pressure — only how ready you are for it. The choices you make along the way also shape your organisation's legitimacy.
Over eight sprints, two teams build that model together, gathering data, refining it, mitigating risks — while unexpected events land in the middle of each sprint's stand-up.
The domain team consists of inspectors and their team lead, the domain expert. They bring field expertise. Their job is to make sure the model reflects real inspection practice. Being closest to the field, they're the ones who understand how inspectors will use it and how it will affect the businesses it targets.
The tech team consists of data scientists and their team lead, a senior data scientist. They bring technical expertise. Their job is to make sure the model performs well. Being closest to how the model works, they're the ones who understand its inner logic — what it can and can't do.
Teams split into Domain and Tech, get briefed on the project, and look through what's on the table — the data available to them, what they could build. Together, you agree what the goal is.
You work sprint by sprint, one task per sprint. Each one pushes the model somewhere — towards effectiveness, towards explainability, rarely a clean split between the two — and stay ready, because events neither team planned for can land mid-sprint.
At the end, the model goes on the public record. Together, you fill out a disclosure form — choosing what to tell the public, knowing you won't have room to say everything.
We talk through the choices you made and why — what tipped a decision one way, what you'd do differently.
4–8 players
Two per team, minimum.
2 hours
Two backgrounds
Want to know how your teams would make decisions on Responsible AI trade-offs?
Beyond the Algorithm simulates how organisations build AI models for risk-based inspections. Your teams design the model together across a series of sprints — weighing effectiveness, explainability, and legitimacy — while handling unexpected events and preparing it for public disclosure.
It's also a research game, so playing it contributes to ongoing PhD research on Responsible AI practices in public-sector organisations.
To arrange a session, write to agagua@tudelft.nl.