2022 was the year that Artificial Intelligence (AI) truly hit the zeitgeist, with a new generation of chatbots and image generation tools causing widespread awe and consternation. In this article, we will examine the significance of these advances, and ask whether these innovations will have a meaningful role to play in the future of transport planning and urban design.
Here at Podaris, we have been following AI developments since our inception. Over the years, we've often been asked how AI – or more commonly “Machine Learning” (more on that distinction later) – is used in our platform. Tellingly, these questions have usually come from technology-focused Venture Capitalists, rather than from people working in the planning sector. To date, my answer has always been the same: we don't use AI to automate planning – not yet. We've always felt that there were fundamental constraints within transport planning – and built-environment planning in general – which no amount of AI, on its own, would be able to overcome.
But the latest AI developments feel quite significant, so I think it's worth revisiting this scepticism, to see whether its assumptions still hold. To preview the conclusion: it's complicated. The ambitions of AI optimists still probably need to be tempered – less because of limitations with AI itself than because of limitations in the social / political / institutional dynamics that it necessarily interacts with. At the same time, it's clear that an ever-growing range of powerful AI capabilities will soon begin to transform the industry – incrementally at first, but quite significantly in the long run. In short, both scepticism and enthusiasm are justified.
Before we try to understand which to apply where, let's understand how we got to this point, by taking a quick look at the history of AI.
A brief history of Artificial Intelligence
AI has gone through several cycles of boom and bust. The first boom began in the late 1950s, when applications such as the General Problem Solver showed that computers were able to solve not just preformulated calculations, but complex problems where the solving equations were not known in advance. This raised hopes that soon AIs would be able to solve any problem given to them, but it turned out that they weren't even able to solve the problem of communicating with humans. Human language seemed impossible to parse, so problems had to be stated in formal symbolic systems, free from ambiguities or contradictions. As it turns out, ambiguities and contradictions are intrinsic to most of the problems that one would want to solve, so this limitation was quite fundamental. Soon the field had collapsed into an “AI Winter” – a lull in research, investment, and credibility.
A second AI renaissance occurred in the early 1970s, when programs such as ELIZA and SHRDLU were able to demonstrate seemingly naturalistic human-language interactions, within simplified “toy” environments. These simple but powerful programs raised hopes that more complex programs would soon be able to handle real-world interactions and solve real-world problems. But this optimism again proved unfounded, as it became increasingly apparent that both the real world and the way we talk about it were many, many orders of magnitude more complex than previously realised. By the early 1980s, AI research had collapsed into its second “winter,” and had become deeply unfashionable.
In the late 1990s, AI research picked up pace again, this time driven more by business needs than by academic interest, and more often going by the moniker “Machine Learning” (ML). “ML” avoided the now-stigmatised term “AI”, and in fact had fairly different goals than earlier generations of AI research. Whereas earlier AI research was often focused on creating human-like knowledge systems and language capabilities, ML research was generally driven by the need to analyse the large and messy “big data” being generated by our increasingly digital world – data coming from financial markets, internet analytics, embedded devices, digital imaging systems, and more.
Although some data scientists felt that “most of the demand for ML is from people who just want linear regressions on a cleaned-up dataset”, its capabilities grew increasingly powerful throughout the 2000s. Driven by advances in microprocessor architecture, memory storage, and Artificial Neural Networks (ANNs) – the core algorithm for which had been invented in 1975, but the computer hardware of the day had been unable to exploit its real potential – ML techniques were increasingly able to make sense of datasets far too big or messy for humans to get a grasp of. This included “natural” data such as digital imagery or audio recordings; where hardcoded feature-extraction techniques had never been able to satisfactorily parse through the complexities and ambiguities of the natural world, ANNs finally began to make headway.
By the mid 2010s this began producing spectacular results, addressing problems from speech recognition to image classification. A particular watershed moment occurred in 2016, when the AlphaGo AI was able to beat the world's top-ranked Go player. This was far more significant than the 1997 victory of Deep Blue over the world's reigning Chess grandmaster – not just because of the vastly greater computational complexity of Go, but because AlphaGo had not used a bespoke hand-coded game-playing algorithm, but had taught itself to play using a more generalisable ANN. By late 2020, broadly similar algorithms had been used to solve the legendary protein folding problem – demonstrating the power and flexibility of these techniques, while sparking a still-unfolding revolution in molecular biology and medicine.
Despite these milestones, AI still didn’t really hit the zeitgeist the way it now has, and I doubt that most transport or urban planning practitioners would have wondered about its relevance to their disciplines. However powerful these AIs might have been, they still seemed very far from having general-purpose capabilities. To the layperson, capabilities like speech recognition and image classification can seem rather trivial (in truth, they aren't), whereas problems like Go playing and protein folding just seem abstrusely computational, and computers have already been out-computing humans for decades (while in fact, these problems are so much harder than that). So the idea of AI having the intelligence and flexibility to disrupt any random discipline just didn’t seem plausible at the time. This is why Podaris was getting few questions about AI from our customers.
But those paying closer attention to the technology began to ask: if AI can solve Go and protein folding, then why can’t we use it to solve, say, transport planning? Is it really that much harder of a problem? Before looking at the latest AI advances, let’s unpack that question a bit.
Why Planning is a Hard Problem for AI
Planning is not an activity which occurs in a vacuum. For the purposes of this analysis, we can think of planning as having three distinct elements.:
- Gathering inputs to understand the problem.
- Creating a plan to solve the problem.
- Communicating and implementing the outputs of the plan.
Traditionally, AI would sit at step 2: ingesting data and producing some kind of optimised response to that data. That’s how it plays Go; that’s how it folds proteins; that’s how it can produce better plans for cities – right?
Conceptually, this is fine, but it begs the question of how well this actually fits with the gathering of inputs and the implementations. And here’s where we run into problems.
First let’s consider the inputs. The data underlying most planning activities is usually, to put it bluntly, pretty bad. Demographic and economic data can be many years out of date, and unrepresentative of current conditions. Transport network data is often malformed or incomplete. Travel demand data is often non-existent, or insufficiently granular, or must be imputed from tertiary sources which can easily introduce significant sample biases. And perhaps most importantly, stakeholder attitudes and opinions – one of the most critical constraints to any planning activity – doesn’t even exist as apriori data; it only lives in peoples’ heads, and must be carefully extracted through repeated cycles of stakeholder engagement.
So, considering all those limitations: if you feed malformed, incomplete, and biased data into an AI for optimisation, it could be a case of what computer scientists succinctly call “GIGO”: Garbage In, Garbage Out.
But wait – can’t an AI learn to overcome these problems, the way that human planners do? Hasn’t AI been proving exactly this: an ability to parse through and optimise messy, ambiguous, and contradictory data?
To date, this hasn’t seemed possible. To understand why, it’s important to understand how ANNs learn, and how this differs from human learning. When a human makes an observation, they quickly form a mental model of what it is that they’re observing, and then continuously update that model through further observations and interactions. AI, so far, doesn't do that. ANNs simply form correlations between data that goes in and data that should go out, given some kind of fitness function or optimisation criteria. They may build up millions or billions of such correlations, but fundamentally they’re agnostic to the actual content of that data. They don’t build “mental models” at all – at least not yet.
In order to form these correlations, ANNs must train themselves on datasets which are either very consistent or very large. Go and Protein Folding are examples of very consistent datasets. While both domains have macroscopic manifestations which are so complex as to defy human comprehension, they are ultimately driven by a finite number of forces – the rules of the game or the laws of physics, respectively. Moreover, the “fitness” criteria is straightforward: winning the game, or net deviation from the target protein structure. With consistent datasets and unambiguous fitness criteria, the AIs can then train themselves on a relatively “small” amount of data – a few hundred thousand historical Go games, or a few tens of thousands of known protein structures – and find correlations that are a very good fit for what they’re looking for.
AIs have also proven themselves adept at handling much messier data, for capabilities like language translation, speech recognition, and image classification. In all of these cases, the data can be far more ambiguous and contradictory, the underlying rules are effectively infinite, and the fitness criteria are much less objective. AIs are nonetheless able to make sense of these domains, because they are able to train on datasets consisting of literally billions of texts or media files. That’s what it takes to discover the underlying patterns and form useful correlations in such “messy” domains.
Planning is one such a “messy” domain – the driving forces are as multiplicitous as the stakeholders involved. Yet it lacks the scale of text or imagery: there aren’t billions of cities whose messy but voluminous data can collectively train an AI to cut through the noise. This data in this domain is both messy and small – which is a problem for the current generation of AIs.
Maybe this will change in the future. It’s entirely possible that a future generation of AIs will learn to form mental models, allowing them to learn from much smaller datasets, the way that humans do. But even if they are able to do that, we’ll encounter another problem: what fitness function should they use? How do they determine if any given feature is something they should be learning to optimise for or against?
For example: you might want to optimise for economic development, right? But that can easily correlate with high levels of gentrification, population displacement, and social unrest. Maybe that’s not what you want, so you could optimise for continuity instead – minimising population displacement – but taking economic development out of consideration hardly seems wise. In that case, surely you should try to balance the two forces? But this requires some kind of generalised cost function whereby you either financialise population displacement, or somehow “socialise” economic development – and who gets to say what the proper exchange rates between these things might be?
To be sure, these are the sort of questions which planners ought to be asking themselves anyways, since the consequences of optimising for the wrong thing can be severe. For many decades, North American traffic engineers considered uncongested fast-moving traffic to be their core optimisation criteria– a practice that is now sincerely regretted by many, since maximising for speed also means minimising for safety and neighbourhood liveability. It’s far too easy to imagine AIs monomaniacally optimising for the wrong thing. Effective planners need to be considering the wider implications of their planning – a philosophical activity which goes beyond the scope of any given dataset.
Now, it’s possible that someday an AI may be able to do even this. At that point, it will need to be a very different thing than the merely data-driven, application-specific AIs that we’ve seen to date. And even then, we must still contend with step 3 of the planning process: communication and implementation. Since there is rarely any one “correct” answer to any given planning challenge, the role of the planner is not just to produce an abstract solution, but to justify it to, and very likely negotiate with, the people involved in implementing the solution, or the people being impacted by it.
This is a very political process. At the end of the day, humans are going to be the ones implementing and taking responsibility for the plans developed in step 2, and they’re unlikely to do so simply because a machine says that they should. They will question, argue, push back, tweak plans, and generally make life complicated for whatever entities – human or machine – are involved in step 2. An AI which is simply processing datasets – rather than dynamically and sensitively responding to feedback that is often contradictory and highly politically charged – won’t be very useful to the planning process.
But wait – maybe an AI can do this? Everything I’ve said above has been true of AIs for years, and has long been the basis of my scepticism. But the latest crop of AIs seems to be truly game-changing. So, let’s take a look at those, and see if that makes a difference.
Has the game changed?
I asked a well-known AI expert how they thought AI could now be used in the practices of urban and transport planning. This was their response:
“There are many ways in which artificial intelligence (AI) can assist with the practices of urban and transport planning. Some examples include:
- Optimizing public transport routes: AI can be used to analyze data on passenger numbers, travel patterns, and other factors to identify the most efficient routes for buses, trains, and other forms of public transport.
- Improving traffic flow: AI can be used to analyze traffic data in real-time and identify bottlenecks, accidents, and other issues that are causing delays. This can help city planners to make informed decisions about how to improve traffic flow and reduce congestion.
- Enhancing pedestrian safety: AI can be used to analyze data on pedestrian traffic and identify areas where there is a high risk of accidents. This can help city planners to identify where improvements are needed, such as the installation of pedestrian crossings or traffic-calming measures.
- Predicting demand for transportation: AI can be used to analyze data on population trends, employment patterns, and other factors to predict demand for transportation in different parts of a city. This can help city planners to make informed decisions about where to invest in transportation infrastructure.
This response struck me as rather generic. Some of these ideas sound plausible, but many of them just seem like traditional applications for Machine Learning techniques, which would still be subject to all the problems and caveats I’ve already discussed. Perhaps this AI expert isn’t really aware of the fundamental problems with the data in this domain? I asked them how they’d deal with that. This was their reply:
“If the data used to train an AI system is incomplete, biased, or inaccurate, it can lead to incorrect or suboptimal decisions being made. This is especially problematic in the field of urban and transport planning, as the consequences of these decisions can be significant and long-lasting.
“If the data used to train an AI system is incomplete, biased, or inaccurate, it can lead to incorrect or suboptimal decisions being made. This is especially problematic in the field of urban and transport planning, as the consequences of these decisions can be significant and long-lasting. For example, if an AI system is used to optimize public transport routes and the data used to train it is biased, it may result in some areas being underserved by public transport. This could lead to inequality and a lack of accessibility for certain groups of people.
If an AI system is used to improve traffic flow and the data used to train it is inaccurate, it may result in the implementation of ineffective solutions that do not address the root causes of congestion. If an AI system is used to enhance pedestrian safety and the data used to train it is incomplete, it may miss important risk factors and fail to identify areas where improvements are needed.
In all of these cases, the use of incomplete, biased, or inaccurate data can have serious negative consequences for the people who rely on the transport system, as well as for the city as a whole. It is therefore important to ensure that the data used to train AI systems in the field of urban and transport planning is as comprehensive, accurate, and unbiased as possible.”
That sounds more like it, with the additional caveat that getting comprehensive, accurate, and unbiased data in this domain is far easier said than done!
At this point I might as well drop the ruse and reveal – in case you haven’t already noticed we were playing the pronoun game – that this “AI expert” is, in fact, an AI. This is ChatGPT, which among 2022’s AIs is probably the one to cause the biggest stir. It’s a “Large Language Model” (LLM) with astonishing capabilities. It can converse intelligibly and intelligently about virtually any subject. It scores highly on linguistic IQ tests, and has even passed the entrance exam to the Wharton Business School.
Yet this interaction demonstrates some of the limitations still present in AIs. ChatGPT still lacks an underlying “mental model” of the world. A LLM is essentially a hyper-evolved version of the predictive text that appears when you start typing a question into Google. ChatGPT doesn’t actually have any opinions of its own – it just tries its best to pattern-match the text that should appear next, given the text that has appeared previously.
So, it knows that if the preceding text asks a question about what AI can do, then the following text should provide some answers about what AI can do. It knows that such answers usually have something to do with optimisation – based on the billions of documents which ChatGPT has trained itself on, doubtless including many articles about what AI is capable of. It further knows that I’ve established the context of “urban and transport planning”, and it knows what optimisation problems are associated with that domain. So it’s able to assemble a coherent response accordingly.
I should emphasise that it’s not simply doing a cut-and-paste pastiche of trained content: the responses are a genuinely unique, and sometimes even creative, synthesis of the correlations it has found during its training. But it still doesn’t know the actual domain; it doesn’t have any real knowledge about the problems and limitations of planning data and processes. When I ask a question asking how AI in planning would be affected by incomplete or biased data, it finds correlations with other discussions about the limits of AI, and is able to frame that in terms of the domain we’re actively discussing, to keep the conversation coherent. And its answers are, in fact, good. But they’re an effect of me creating the right prompt to provoke such a response; it’s not knowledge which would be spontaneously volunteered during a conversation. Nor does it actually learn anything from this interaction. ChatGPT keeps the context of a conversation “stateful”, but it doesn’t re-train its neural network using things it has learned from that conversation. So when the next person comes along asking about AI in the context of transport and urban planning, they will encounter the same relatively blank slate that I did.
In other words, ChatGPT is a system with a very high level of linguistic intelligence, but a very low level of actual knowledge. This is still a very significant limitation.
But we shouldn’t imagine that it’s a limitation which will last forever. AI has a history of defying expectations. The AI Winters showed that expectations of easy and rapid progress could be stymied for decades by hitherto unforeseen difficulties – but in recent years, we’ve seen precisely the opposite: that AI capabilities can grow at an unexpectedly exponential rate. Going forward, either scenario is plausible. The synthesis of language models with continuously-learning knowledge systems might take a very long time due to subtle but significant challenges – or it might already be happening in a lab somewhere, ready to be released tomorrow. We can’t know either way, and forecasts of the future of AI should really take this degree of uncertainty into account.
What’s more certain, however, is that regardless of the present or future capabilities of AI, when it comes to transport & urban planning, the context it finds itself in will always be extremely challenging. The upstream data will be problematic – with critical constraints often not existing as data at all, but rather existing as attitudes and opinions among stakeholders – and the downstream communication and implementation processes will always be intensely political, forming new and idiosyncratic constraints in their own right.
So the vision wherein an AI can ingest the perfect data context of your city’s digital twin and then create optimised plans for implementation is, in short, a pipe dream. Regardless of how smart you make the middle of that pipeline, what happens on either end is too political to be a mere optimisation exercise.
So, does that mean that there’s no future for AI in this domain? I asked ChatGPT for its opinion.
It is true that data is often imperfect, and this can be a challenge when using AI in the field of urban and transport planning. However, this does not mean that AI has no role to play. Instead, it means that it is important to be aware of the limitations of the data and to use it in a responsible and transparent manner.
One way to mitigate the risks of using imperfect data is to supplement it with other sources of information, such as expert knowledge or qualitative data from surveys or focus groups. This can help to fill in gaps in the data and provide a more complete picture of the issues being addressed.
Another approach is to use AI in a way that is transparent and explainable, so that the decision-making process is understood by all stakeholders. This can help to build trust in the use of AI and ensure that it is used in a way that is fair and accountable. Overall, while it is important to be mindful of the limitations of data, there is still a role for AI in urban and transport planning, as long as it is used responsibly and transparently.”
This is actually a really good answer, and in fact, I’d go further: I think that – contrary to how AI techniques are applied in more deterministic domains like Go playing or protein folding – AI is likely to soon find a niche in the early and late stages of the planning process, perhaps even before it finds a niche in the middle.
AI techniques could be used, today, to greatly improve the data that is gathered – for example, using feature classification techniques on satellite or streetview data to greatly improve the spatial resolution of demographic data, or to identify and correct subtle sampling biases, or even using LLMs to interview stakeholders at a previously unthinkable scale and depth. All of this could make for much richer inputs into the planning process which – once done – starts to make the process of AI-assisted optimization more tenable.
Similarly, on the implementation side of things, image generation tools could be used to communicate the outputs of those plans in ways which have never been done before, and LLMs could again assist in gathering stakeholder reactions at a previously impossible scale, and communicating them back to planners.
In other words, the most plausible application of AI for transport and urban planning will not be to replace human effort and input, but rather to make it cheaper and easier to put more humans into the planning loop. Personally, this is something that I’m excited about.
AI is not a panacea for planners, nor will it be one in the future. The context of planning – its political aspects, and the degree to which many of the key constraints are only ever stored inside peoples’ heads – imposes fundamental limits on how AI can be used in the planning process. It will certainly never replace planners. However, it shows considerable promise in being able to augment the work of planners, improving both their inputs and outputs, sometimes in unexpected ways.
In my next article, I’ll be looking at some very specific AI-in-planning concepts, and the possibilities and pitfalls surrounding them.