A project at the MIT Media Lab tested AI systems that interpret people’s needs at decision points in simulated public spaces. Its most useful finding is also a reminder that reaching a destination is not the same as following an accessibility preference: eight simulated pedestrians with limited mobility chose stairs despite preferring not to.
Watch Desk analysis
What happened
The project, Beyond the Rulebook, describes two GAMA simulations. In a simplified New York subway station, 30 passengers asked an information desk for directions. The study compared TypeSafe AI Jev 1.13 with OpenAI GPT-6 Luna using the same requests and structured answer choices. Initial interpretation was accurate for 28 of 30 requests with Jev and 29 of 30 with GPT-6 Luna; after clarifying questions, their final route assignments agreed. Jev’s median API response was 224 milliseconds, compared with 1,511 milliseconds for GPT-6 Luna in this test.
The second simulation placed a local H2O-Lightning-4B model at six decision points in a campus network. The 500 simulated pedestrians had four profiles, including low vision, limited mobility and wheelchair use. Across 1,724 decisions, the project recorded a median API response of 159 milliseconds and a 95th percentile of 685 milliseconds. All reached their destinations, but eight pedestrians with limited mobility chose stairs despite their stated preference. Wheelchair users did not choose stairs and used the lift for all 41 changes of level.
Why it matters
Instead of giving every simulated person a complete route in advance, the project has a model interpret local information, such as signs and available paths, at each decision point. That offers a way to test how guidance might respond to different needs as circumstances change, without writing a separate rule for every person-and-route combination.
The results also show why a simple success measure can flatter a system. Everyone arrived, yet some simulated pedestrians did not follow their accessibility preference. For anyone designing navigation or assistance tools, that distinction matters more than a neat completion rate.
Our read
This is a thoughtful demonstration of AI interpreting local context rather than blindly following a fixed script. The accessibility miss is the detail to keep in view: “arrived” is not the same as “got there in a way that worked for you”.
The project describes functional demonstrations, not calibrated assessments of real crowds or accessibility. Its geometry and movement are schematic, and the campus study had no independent accuracy benchmark or comparison engine. Treat the timings and outcomes as results from these simulations, not a verdict on how the system would perform in a busy station.
What to watch
- Whether the project tests the decision system with real users and less schematic environments.
- How future work measures whether accessibility preferences are followed, not just whether destinations are reached.
- Whether researchers publish independent comparisons for the campus scenario.
Discussion spark: If a navigation system gets someone to their destination but ignores their stated accessibility preference, should that count as a successful journey?
Sources and evidence
- Project Overview ‹ Beyond the Rulebook – MIT Media Lab (27 October 2026, 20:33 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.