Walking Against the Map: From location-aware AI to negotiated experience
I had the pleasure of being the closing speaker of WALKING – Denken in Bewegung 3, hosted by the Institute for Land Environment and Art, which ran between 11 and 13 September near Basel. I wasn’t able to attend in person, and spoke remotely.
My presentation discussed the use of Artificial Intelligence (AI) in walking art in general, and as an aid, or perhaps hindrance, around location-based discovery, as well as agency-building, in particular.
I will start with a deceptively simple question: who tells us how, and where, to move?
Back in 2012, I co-created Dérive app, based on the ideas of the Situationists. The Situationists, of course, showed us that, in public space, we are being manipulated in our behaviour by Big Capital, those who control the public space.
Combining this with my history of making location-based work, supported by modern technology, and, increasingly, seeing digital systems becoming better and better at telling us where to go and what to notice, my pushback against this manipulation has only grown with the years.
Here, I’ll trace a line through some of my work, and I’ll also look at a new crop of mobile walking apps, which exist because of the introduction of AI.
I’m interested in, and at the same time wary of, the promise of these technologies, and can see risks on the horizon, separate from the well known problems emanating from broad adoption of AI in daily life. Specifically this: The more perfectly an experience is generated for an individual, the more easily the individual can end up in a private experiential bubble.
I will try to argue that the more interesting direction to follow may be almost the opposite: using these technologies to produce situations that have to be negotiated socially.
Who tells us how to move?
The Situationist realisation was that moving through the city is never neutral.
We tend to experience our daily routes as obvious: home to work, station to restaurant, the shortest route to a deli. But the city is full of systems that shape those routes: some more benevolent, some decidedly less so; from urban planning, to hostile architecture.
The idea of the dérive, the unguided meander pioneered by the Situationists, is interesting because it treats those patterns as something that can be interrupted. One can decide to choose what determines movement: whether attraction, boredom, or an arbitrary rule. It’s you who can decide.
But, what matters is not simply getting lost, but noticing that the apparent natural organization of everyday movement is constructed, and therefore can be resisted.
Digital systems have made that conditioning more explicit, more efficient, and more personal; A navigation system calculates an optimal route. A recommendation system predicts which place I will prefer. A list of must-sees tells me what deserves my attention.
So, the Situationist observation of manipulation in the public sphere has not disappeared. It has become more literal, and more ingrained.
We should constantly ask the question: how much of my interaction with the world at large is being decided, and structured, for me?
A trivial example is the blue line on a phone’s mapping application. It is useful because it removes hundreds of decisions that otherwise would fall on your own shoulders. But precisely because it is useful, we can stop noticing that each removed decision was also a possible experience. The road not taken has disappeared before it even could becomes a choice.
It’s here where my interest in creating experiences in the public space lies; not through a rejection of technology, but as a tool to surface alternatives when efficiency is temporarily suspended.
The Dérive paradox
Dérive app’s basic logic is deliberately simple. Instead of giving you a destination, it gives you a loosely defined task: “follow someone wearing a hat”, “find something that reminds you of your childhood”.
A deck of these tasks creates a walk, a dérive, without a predefined route.
But, this also creates an obvious contradiction. The application really says: stop following the usual instructions and follow my instructions instead. That means the app can reproduce the structure it is intended to disrupt, replacing one authority with another.
Dérive app does try to address this, for one by, by default, not allowing more than one instruction per hour. But this challenge is not just philosophical; If I stare at the phone waiting for the next instruction, the phone becomes the centre of the walk, with the city becoming the backdrop for completing what the software asks, running the risk of it turning the experience into the logic of ordinary navigation.
So, for me, a useful dérive, and useful tasks, or prompts, cannot simply replace one navigation system with another; the tool should create the conditions under which navigation itself becomes what is negotiable.
The best Dérives are negotiated
Over the years, this has become particularly clear during Dérive app workshops, one, or two, day events where participants construct their own set of task cards for the area where the workshop is held. Here, I’ve found that the strongest experiences occur when people go out in small groups.
As a prompt is typically not overly precise, three people can understand, or interpret, it differently. What might be obvious to one, might be far fetched for another. Meaning, purpose, and interpretation construct the experience. The instruction starts the process, but it does not control the result.
The social interaction between participants weakens the authority of the app. The app, designed to put agency back in the hands of the individual, requires participants to negotiate the meaning of the task. And so, the prompt becomes one actor among several: the participants, the street, chance, and practical constraints.
This leads into what I think will become more important as AI starts facilitating walking experiences in particular, and travel in general. A problem to tackle, is that one person plus one adaptive system, the AI, can form a painfully obvious closed loop. On the other hand, a small group confronted with a shared, incomplete, provocation has to construct something together. The value, and interest, is not in the instruction, it is in the negotiation around it.
The prompts in Dérive app have always been more provocation than instruction, but it’s the shared experience that is much more capable of surfacing this. Ambiguity, amongst several participants, provides space for them to become co-authors of an unscripted experience.
Sins Beneath the Equator
I recently collaborated on the project Sins Beneath the Equator, in which we approached ‘location’ from a different direction. The work is a collection of one hundred short location-based podcasts about the Dutch colonial presence in the Northeast of Brazil, centred on Recife.
We addressed the persistent myth that if the Dutch had stayed, Brazil would have been ‘better off’, today. The project looks at what sits behind that nostalgic version of colonial history: slavery, extraction, warfare, mercantile capitalism, religion, science, and the people who are normally pushed to the sidelines.
We experimented with AI tools in some parts of the process, particularly around translation and, to a lesser extent, audio production. We also, in some archival research, were able to use AI to speed up discovery, though we had to make sure the presented outcomes were not hallucinated, such that everything was signed off on, by a human.
We realised that it was important to keep a distance between the use of AI and the presented outcomes; Generative systems (LLMs) are very good at making a text appear meaningful, that is, able to produce confident historical texture, but also to be able to do that, when the evidence is weak. In a project that aims to discuss hidden histories, and colonial memory, that is not trivial.
The archives from which we had to draw are already uneven. Some actors are documented obsessively, others barely at all. If AI smooths the inequalities in the sources into a plausible story, it can hide the very absence the project needs to expose.
So, we were confronted with the fact4 that we can treat AI, at best, as an assistant in research, not as a historical source.
Yet, for me, this project did start me down a route to experiment with the use of AI in creating location-based and customised experiences. Though we need to be aware of the implicit distinction already mentioned: there is a difference between AI supporting the construction of a situated experience and allowing AI to become the system that decides what the situated experience should be.
AI also changes who can build the instrument
There is another change that’s a consequence of AI becoming more capable; Software development cycles are getting smaller. The current version of Dérive app was developed much faster because coding assistants helped me refactor, test, debug, and implement the features I needed.
Building good software is difficult. You need the idea, but you also need enough programming experience, time, or money to turn it into a working instrument. Meanwhile, we have to understand, that not having development experience makes vibe-coding more like vulnerability-as-a-service.
Nevertheless, the thresholds for developing software are dropping quickly. An artist, writer, walker, educator, or very small team can now, with relative ease, build a peculiar little machine for encountering the world.
I have major issues with the widespread adoption of AI, but there is also some cause for optimism: We are not only getting more generated content, we are also lowering the cost of inventing new forms of interaction and behaviour.
Dérive app can customise the prompt for you
My desire for rebuilding Dérive app was, in part, due to my interest to incorporate a more direct use of generative AI. As an added piece of functionality, users can now ask the app to take into account their actual location, their interests, and a desired scope, to have it construct a stream of prompts, uniquely defined for the user, inspired by the user, and their location.
A generic, regular, card can perhaps trigger a more contemplative reaction, but it might also simply not be applicable.
Instead, a custom-generated card can be much more plausible. It can look at the local context and create something appropriate for exactly this place; a person in Dar Es Salaam, Recife, Basel, or a village with no pre-existing deck can receive a prompt with local relevance.
But there is a tension here. The better the system gets at knowing what fits me and where I am, the more perfectly it can mediate the encounter. So, increased situational relevance can also increase the authority of the system. Then, the challenge becomes to identify how we can use context, without turning that context into another form of control.
The old, ‘dumb’, prompt has at least one advantage: because it does not understand context, I, as the user, have to do the work of interpretation. A more intelligent card can remove some of that interpretive friction, which could easily lessen the value of the experience.
The design challenge can therefore not simply be to maximize relevance.
Gemini knows where I am. Sort of.
This expansion of Dérive app lead me to experiment with Google Gemini as a location-based guide. Gemini can, with a little hassle, use a device’s precise location, and combine that with location-based data from Google Maps. So, I started asking it to generate experiences for where I was actually standing.
The results ranged between fascinating and annoying, but were never quite reliable. Typically, at the start, the assumed location is correct, but it’s often not updated properly as you move, without telling you it isn’t. But, worse, Gemini has no understanding of your orientation. It may know that a church is thirty metres away, but not whether it is behind you. Or, if I ask a human guide to tell me about what is in front of us, the phrase ‘in front of us’ carries embodied information that is not available to the LLM accessible through my phone.
The available latitude and longitude reduce the walker to a point, but location data is not the lived situation. The resulting experience, of combining the walker’s location with potentially interesting locations nearby, and presenting these to the user, does not generate an immersive, seamless experience.
So, the (overall) result, for all its shortcomings, was one of frustration.
Next: location + orientation + continuity
The next iteration I am exploring, is a small application that sits on top of the generative AI and continuously maintains a better model of the walker’s situation. This will mean keeping track of location, orientation, movement, time, and history.
That should allow the system to distinguish between ‘there is something interesting 50 meters away’ and ‘look to your left at the facade you are currently approaching’. That is to say, with a more accurate model of the user, it should be possible to create a much more embodied kind of instruction.
But we have to be careful: This creates richer possibilities for site-specific intervention, but it also creates a much more intimate mediator of perception.
A new ecology of walking apps
My experiments are not happening in isolation. Over the last year or two, a noticeable cluster of small walking-oriented apps has appeared. They are not all doing the same thing, and not all of them use AI in the same way, but the development is fascinating.
Lore attaches stories to place. MIX creates stories that are walked through. StreetLore dynamically narrates what is around you. SideQuest turns nearby places into quests. Tiny Worlds converts steps into growing miniature worlds. Snapfari turns encounters with animals into a collecting game. Unterra lets you unfog the world. Break a Leg !! Has zombies chase you on your daily run.
I have a strong suspicion they are all vibe-coded, but can’t be sure. They certainly belong to a wider moment in which small teams and individuals can build mobile experiences much more cheaply and quickly, while generative systems can also supply content dynamically.
Walking is becoming a substrate on which many more people can build unusual experiential systems.
My interest is not quite as to whether these apps are artistically radical. Some are games, some are guides, some are fitness products. The point is the multiplication of form.
A few years ago, many of these ideas would never have made it past a prototype, student project, or concept. Now the economics are different; AI-assisted development, and better cross-platform tools mean that a very niche idea can become a functioning app, quickly, allowing for previously invisible or inaccessible creativity becomes executable.
Re-enchantment from the machinery of flattening
That lowering of the bar, is a very positive development.
This, because, for decades, digital mapping has functioned by making the world computationally legible. The city becomes routes, points of interest, categories, ratings, travel times, coordinates, and rankings. That is incredibly useful. But, it also flattens the experience.
Up until only a year or two ago, location-based experiences in Google Maps were, literally, the same for everyone. Now, generative systems can, and do, use much of that same infrastructure to again enrich the experience. We can easily add narrative, fantasy, ambiguity, provocation, historical context, play, as well as deliberate inconvenience.
So, the machinery that has become extremely good at removing uncertainty from public space can now be asked to manufacture uncertainty as well; a system built to convey the quickest route can instead tell me to avoid streets whose names contain the letter A, or to look for evidence of a vanished industry, or to follow the sound that seems least inviting.
There is something genuinely promising in that reversal. It is a way of using computational systems against the dominant logic of optimization.
And so there is a nice contradiction in this; The smartphone is one of the devices most responsible for drawing attention away from the street, yet it can also become an instrument for making the street newly strange. The map that normally suppresses detours can generate them. A database of places intended for search and commerce can become raw material for play.
So this is why I can’t treat technology as inherently hostile to Situationist practice. It can be used in ways that are hostile to it, but it can also be bent against its default logic. Now more than ever.
Standardisation under the guise of diversity
Yet, there is a structural problem.
These applications can appear extraordinarily diverse, while drawing from a surprisingly small number of underlying systems. For location-based data, the big three are Google Places, HERE, and FourSquare, all with major overlaps. Historical, cultural, and local information comes from the same unevenly digitised sources.
So we have increasing diversity in the interface, with increasing homogeneity underneath it. Thousands of small applications can produce different aesthetics and different prompts while fishing from the same pond.
This is particularly relevant to decolonial practice, like for the project on Brazil I mentioned earlier; A model does not encounter a place on its own terms. It inherits a representation of the place from what has been written, digitised, indexed, translated, and made prominent. The histories that are absent from the archive will remain absent from a supposedly inventive generative experience.
The danger, then, is not only corporate concentration. It is that standardisation can arrive disguised as limitless creative variety.
There is also a more subtle creative consequence. If multiple independent developers all ask the same underlying model to invent a surprising walk, those walks may differ at the sentence level while sharing assumptions about what counts as surprising, safe, interesting, historic, or worth seeing.
We will mistake stochastic variation for genuine plurality.
So, this new crop of apps are all fishing from the same pond; The surface diversity is real, but it does not automatically imply diversity of worldview.
From filter bubbles to experiential bubbles
I see another risk, and this on a more personal level.
Having sat through the 2016 US elections, we already know the logic of the filter bubble: systems learn what holds our attention and increasingly show us versions of the world that fit our existing profile.
Walking applications fishing from the same pond can extend that logic from information into physical experience. Navigation says: here is the best route. Recommendation says: here is somewhere a person like you will probably enjoy. A generative experiential system can go a step further: here is what this place means for you, right now.
At that point the system is not merely filtering a set of existing options. It can literally manufacture the thing being experienced. If my interests, history, location, behaviour, and preferences all become inputs, I can increasingly inhabit a city generated for me.
Everyone can receive something unique while occupying the same structural relationship, based on the same source data: I give the system my context, the system interprets me, and the system constructs my experience.
We all become unique… like everyone else.
A filter bubble still leaves a shared physical world outside the screen. An experiential bubble begins to layer the personalization over that shared world itself.
Individualisation is… maybe… not automatically antisocial. But what happens when this personalized layer becomes the de facto mediator between the individual and the place, specifically when this facilitation is closed, opaque, constantly adapting itself to the individual, and run by capitalists whose only interest is to extract as much as possible from the individual and society?
One provocation, or multiple private worlds?
Earlier, I brought up Dérive app workshops, and that the best dérives are negotiated.
Imagine three people standing on the same street corner. A sufficiently advanced system can give each of them a perfectly individualized prompt, taking into account their interests, history, mood and, location. Technically, this is impressive, but, experientially, the three people can disappear into three parallel worlds.
Now imagine that the same three people receive one single, constructed, but perhaps ‘incomplete’, and deliberately ambiguous instruction. They have to agree on what it means. They will negotiate. Someone might object. Someone might change the rules. The group has to construct the experience, collectively becoming the situation.
This experience is not personalized, but designed to be more socially generative. The experience does not arrive finished, it is constructed by its participants.
So, I do not think the objective should be to make AI better at personalization; we should design systems that deliberately create dependency on other people, the machine deliberately seeding control to chance and collective decision.
Don’t get me wrong; I’m not arguing that solitary walks are bad, or that every experience has to become a workshop. The issue is whether the system leaves room for the world and for other people to resist its framing. A solitary dérive can absolutely rupture habit. But I see the risk of a highly responsive personalized system as becoming so good at maintaining the relationship between user and machine, that very little external negotiation is required or called for.
In other words, we have to design for other people to be inconvenient, in a productive way.
Design for negotiation, not personalisation
And so, we arrive at a design principle I want to explore further. Perhaps a good generative walking system should know when to stop relying on what it knows about you.
Instead of optimizing for relevance, it could optimize for negotiation, interaction, and discovery, purposefully facilitating an open endedness that builds on requiring to absorb ambiguity.
Instead of keeping the person in communion with their phone, it can explicitly withdraw the technology from the experience.
In this model, AI, and technology, no longer authors the actual experience. It sets up some initial conditions, and then loses control of what happens. And, this feels much closer to the Situationist idea of constructing situations than simply using a more sophisticated digital guide.
I can easily imagine how this can work in a group; the system can issue conflicting, or complementing, instructions that require discussion and negotiation.
For individual experiences, this is perhaps a bit less obvious. Though designing more open ended experiences seems the direction to look for. Some personalisation, probably, but also removing structure, somehow.
Perhaps designing for the system’s own loss of authority.
Can prediction make us less predictable, together?
We all live under the yoke of Big Capital and Big Tech, so I can’t end with a technological forecast, but I can end with a question: Can technologies, designed to predict and personalise human behaviour, be repurposed to make us less predictable, together?
I think they can; new technologies are giving us access to forms of creative experimentation that previously were effectively inaccessible. But we have to avoid using these to construct ever more precise individual experiences, as these will risk reproducing the structures we hope to escape.
The more radical future is to use them to create more social, more contingent, more negotiated situations: experiences whose outcome depends on place, chance, disagreement, and other people.
We can have not a personalized city, but a negotiated one.