- Published finding
- Offers four areas for human–AI collaboration: delegation, description, discernment and diligence.
- Our design connection
- We use those questions to guide tasks about choosing work, specifying boundaries, checking outputs and taking ownership.
- What it does not prove
- A framework is not validation of Tomorrowville, a learning outcome or an AI-fluency score.
Research for
practical AI fluency.
We study how people direct assistants, verify their actions and apply what they learn in unfamiliar situations. These primary sources inform Tomorrowville’s design; they do not establish its learning effectiveness.
Primary reading
Published findings, their connection to our design, and the limits of that evidence.
Showing 4 sources in all research tracks.
Experimental study
To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making
Zana Buçinca, Maja Barbara Malaya & Krzysztof Z. Gajos
Authors’ paper · arXiv · 2021- Published finding
- Deliberation interventions reduced overreliance in the experiment, with tradeoffs in subjective ratings and uneven benefits across participants.
- Our design connection
- Prediction and inspection may encourage reflection before accepting an assistant’s result. That is our design hypothesis.
- What it does not prove
- The study does not establish that our controls teach verification. Extra friction needs testing, too.
Research synthesis
How People Learn II: Learners, Contexts, and Cultures
National Academies of Sciences, Engineering, and Medicine
Chapter 5 · Knowledge and Reasoning · 2018- Published finding
- Synthesises research on how knowledge is organised and integrated across learning experiences, including the role of domain-specific expertise.
- Our design connection
- We ask learners to explain a rule and test it in an unfamiliar case, rather than only repeat a successful run.
- What it does not prove
- This synthesis is not a trial of AI games or evidence that Tomorrowville produces transfer.
Research paper
Generative Agents: Interactive Simulacra of Human Behavior
Joon Sung Park and colleagues
Authors’ paper · arXiv · 2023- Published finding
- Presents memory, reflection and planning for language-model-backed agents in a town sandbox, and evaluates behavioural believability.
- Our design connection
- A spatial world can make interactions visible. It inspires questions about showing coordination, rather than a claim that we implement this architecture.
- What it does not prove
- Believability is not operational accuracy or learning. Tomorrowville currently uses scripted rules with no live model calls.
What we’re testing
Improving an assistant’s behaviour is observable in the simulation. We also need to find out what the person understands.
- Understanding
- Can a learner explain why a failed reservation and a confident confirmation are different, without following a hint?
- Transfer
- Can they apply the confirmation rule to a fresh request, another business or an unfamiliar tool?
- Usability
- Does prediction encourage reflection, and which interaction, language or input barriers make investigation harder?
Current scope
Tomorrowville is a local scripted simulation with no live model calls. The bakery supports configuration, tool inspection and replay; Studio provides local worlds with resources, processes and routes.
Engine tests check the simulator’s rules, not learning. We have not measured learning gains, transfer or classroom effectiveness. Simulation rates and costs are illustrative, not business forecasts.
Inspect the bakery example