How is artificial intelligence transforming the urban environment? What does AI-enabled transportation planning, ecosystem health monitoring, and urban design mean for the future of cities? How must AI adapt to better fit urban contexts? Urban AI and The Jacobs Technion-Cornell Institute have teamed up for a weekly speaker series that calls on experts at the forefront of urban artificial intelligence to imagine what challenges and opportunities will face the field in the decades to come. Ron Brachman, Director of The Jacobs Technion-Cornell Institute, kicked off the series with a talk on “AI, Urban Systems and Common Sense.”
At present, AI systems have limitations that prevent them from operating autonomously in urban environments, by virtue of the fact that they are trained on limited sets of historical data. According to Brachman, the future of urban artificial intelligence lies in the ability of AI systems to take into account realities beyond those programmed knowledge bases, exercising what he terms “common sense.” Currently, the most successful AI systems operate within a contained, domain-specific proficiency, but lack the ability to meaningfully extrapolate to situations for which they lack expertise. If a given scenario does not exactly resemble what the AI has been trained to recognize, it will perform poorly. This limitation does not pose an issue in “closed worlds,” where edge cases (rare and/or unexpected circumstances) rarely arise, if at all. In a game of chess, for example, where the rules always remain the same and players can be expected to follow those rules to a tee, a bot can be programmed to interpret and respond to any move an opponent plays against it. However, the real world is an “open world:” dynamic and unpredictable. This lack of predictability has already led to a number of high profile artificial intelligence blunders — Alexa challenging a child to electrocute herself, GPT-3 giving destructive mental health advice, and self-driving cars stopping for stop signs in billboards. The issue of open worlds is particularly acute in the case of artificial intelligence that operates in urban contexts.
Cities are inherently complex. While sets of rules do govern collective behaviors, those rules only go so far in determining the individual actions of the multiplicity of independent agents that exist in urban spaces. As Brachman explains, cities follow a long-tail distribution of the probability of events. This means that no one rare event is particularly likely to occur, but there does exist a high likelihood that any number of rare events will occur. This presents quite a challenge when it comes to programming artificial intelligence to operate in urban settings, particularly if the training data is composed of past events that have already occurred (a set which contains some scenarios unlikely to repeat and omits many others that might still happen). It would be impossible for programmers to include every potential scenario in an AI’s training set. So if the future of urban artificial intelligence is to bypass this limitation by designing systems that can make productive decisions beyond the training data that they have at their disposal, what might that process look like from a design perspective?
In Brachman’s eyes, it comes down to building common sense and intelligibility into the architecture of AI systems. Beyond the current standard of basic background knowledge, domain-specific expertise, and the ability to respond to common scenarios with default actions, all of which only prepares AI systems to deal with routine situations, the technology will need to develop such that artificial intelligence can perceive and make sense of external data, in order to diagnose an unforeseen circumstance and shift its approach accordingly. As part of shifting its approach, a well-designed artificial intelligence should have high-level functioning to determine whether the intended goal still applies in the context of the new situation (Brachman gives the example of a self-driving car on its way to a grocery store: if a parade, an unexpected scenario, obstructs its route, the car might need to make the decision to find a different grocery store or even to just give up and return home). This kind of autonomous reasoning falls under Level 5 autonomy (ISO 22989:2022), in which artificial intelligence has the capability to make decisions without human oversight or input. Many AI scholars caution that the removal of human oversight at that level of autonomy can pose a risk to human health and safety; article 22 of the GDPR explicitly bans a lack of human oversight in decisions with legal effects. In light of these concerns, Brachman calls for interpretable, rational intentionality to be built into AI systems. That is, if a decision made by a Level 5 artificially intelligent system comes into question, programmers should be able to back out a clear logical progression toward the decision made, following the formula: Agent X performed Action A because it believed Premise P and it wanted to achieve Goal G. If decision making follows such understandable lines, programmers can utilize the resulting causal chains to pinpoint faulty logic and improve the common sense knowledgebase of the AI system. As such, programmers can maintain some insight, if not oversight, and provide correction as needed.
Common sense-based artificial intelligence, as implemented in the manner that Brachman describes, will not be stuck simply reacting to the present based on limited data about the past. Rather, those systems will nimbly interact with dynamic, continuously shifting urban environments, prepared to cope with any unforeseen events, well into the future.
By Sarah Popelka, Research Assistant at Urban AI
