Better predictions could help cities spend wisely. They could also mistake the journeys people cannot make for journeys they do not want.
On 19 October, Cornell announced a three-year, $600,000 research grant to improve AI models of human choices in transport, energy and other connected systems. The ambition is to make predictions more understandable and explicit about uncertainty. This is a research programme, not evidence that a city has become fairer or more efficient. But it poses an immediate question for public buyers: are they purchasing a better forecast of how people live, or a way to improve the choices available to them?
The distinction matters. A model can correctly predict that someone will stay home without explaining why leaving is impossible. Public-service AI should be judged partly by the opportunities people gain, including opportunities their past behaviour could never reveal.
Imagine a woman who declines an evening job because the last bus leaves before her shift ends. A transport model sees few evening journeys from her neighbourhood. If that forecast becomes a reason to withdraw the remaining service, the prediction helps produce the future it anticipated. This is a possible feedback loop, not a finding about Cornell’s project.
India offers a concrete reason to take it seriously. The World Bank’s 2022 urban-mobility toolkit describes how safety and affordability restrict women’s travel and access to work. It notes that women may use slower transport because faster alternatives cost more. Observed behaviour therefore needs interpretation: a long walk can express a tight budget rather than a preference for walking.
The same question applies to a wheelchair user missing from bus records or a resident whose work begins beyond the timetable. Asking only existing passengers what they need leaves out people the service has already excluded. App-based data would compound that problem wherever people cannot use the app.
There is a substantial prize here. More dependable forecasts could help operators position vehicles, reduce wasted capacity and make scarce budgets stretch further. Cornell’s researchers explicitly want to understand how individual decisions change a network and how the network changes subsequent decisions. Recognising where evidence is weak is a useful ambition for public infrastructure.
Nor should every forecasting task be burdened with answering every social question. In their 2015 paper “Prediction Policy Problems,” Jon Kleinberg and colleagues argue that some policy decisions can benefit greatly from better prediction itself. A city should not postpone useful scheduling improvements until it has a perfect account of human motivation.
The harder issue is what counts as improvement. Fewer vehicle-hours may save an operator money while imposing longer waits on workers. Conversely, a route that initially looks expensive might make more employers reachable. That could widen workers’ choices and employers’ recruitment pools. These are economic mechanisms to test, not benefits established by a grant announcement.
The International Transport Forum’s 2019 work on accessibility provides a useful starting point: assess people’s access to jobs, services and other opportunities. Applied to AI procurement, this means counting what residents can reach within a realistic time and budget, alongside passenger numbers and operating costs.
It also changes who must be paid. Useful evidence may require local-language interviews, accessible consultations, drivers’ knowledge and trials of services that do not yet exist. Software subscriptions cannot replace buses, staff or streetlighting. A vendor’s apparent saving is incomplete if the city stops paying for this work and households absorb the consequences.
Three futures are plausible. In a better one, cities combine forecasts with small service experiments and measure whether previously excluded residents gain workable options. In an uneven one, established routes improve while people outside the data remain outside the service. In a harmful one, low predicted demand becomes an apparently objective justification for cuts, and reduced use then validates the model. Procurement targets and budget decisions help determine which path becomes attractive.
Questions for founders and investors
Can your product distinguish a weakly supported forecast from a reliable one? Does its evaluation include people who currently cannot use the service? Who pays to test whether changing the service changes demand?
Public buyers should start with a bounded trial, publish the assumptions that matter and measure access before and after it. Investors should price that evaluation into the business rather than treat it as optional overhead. Residents should have a route to challenge a forecast with evidence of unmet need. The practical test is whether someone can now accept a job, attend a class or get home safely because an option that once existed only on paper has become usable.
Sources