Dear CIO,

By Game 4 of the famous match between Lee Sedol and DeepMind's AlphaGo, Lee had already lost three consecutive games, and this match seemed similarly decided. Then, playing white, he placed his 78th move into the middle of AlphaGo's position. The professional players watching were stunned, as was AlphaGo. DeepMind's system had assigned an extremely low probability to a human making the move, so after Lee played it, AlphaGo responded poorly. This move, which was later described as a “divine move”, helped Lee eventually win the game, his only victory in the five-game match. 

I've been thinking about Move 78 again recently. I wrote about it in Rebels of Reason, but my interpretation of it has changed. At the time, I was fascinated by what AlphaGo revealed about machine intelligence and the possibility that machines might reason differently from humans. Now, I'm much more interested in Lee Sedol. Move 78 raises a different question: What happens when the human sees something the machine doesn't? That question feels increasingly important as we build AI into more and more of our systems.

Best Regards,
John, Your Enterprise AI Advisor

The human advantage lies beyond finding the answer

Human in the loop

We talk a lot about keeping a "human in the loop" with AI. The AI produces something, a human checks it, the human approves or rejects it, and the machine continues. Condensed, it looks something like:

AI → human approval → action

Plenty of situations need that exact workflow, but it's a remarkably limited vision of human-machine collaboration. The human becomes a safety mechanism. In this line of thinking, companies keep someone around because the machine might make a mistake. In that model, the ideal AI system is ultimately one where the human doesn't have to do very much, and as AI becomes more reliable, we reduce human involvement. Eventually, if the machine becomes reliable enough, we might automate the loop entirely.

However, Move 78 suggests another possibility. What if humans aren't valuable merely because they can catch the machine when it's wrong? What if the human can see something outside the machine's model? What if the human has a Move 78? The objective then isn't simply keeping a human in the loop. It's keeping humans on the loop.

Being on the loop

There is an important difference between the two. A human in the loop participates in individual decisions. A human on the loop maintains awareness of the system and the landscape in which it operates. Think of the difference between approving moves and watching the board. Someone on the loop can ask questions like: Where are we? What has changed? What is moving? What doesn't fit our existing model? And perhaps most importantly, what might we see that the machine doesn't?

This distinction becomes more important as AI systems become more capable. If AI becomes better at performing individual tasks, then simply inserting a human approval step may add less and less value, but maintaining human awareness may become more valuable. The human needs to be able to see the landscape. 

There is an older argument hiding in this idea. In a 2008 Cynefin post on chess, change, and obliquity, Jon Kendall drew attention to John Kay's observation that people who hold too tightly to a single idea or fixed design tend to lose in chess, battle, business, and economics. The associated idea was obliquity: that in systems which are complex, imperfectly understood and altered by our participation in them, objectives are often achieved indirectly. That is quite different from the way we tend to think about machines. Move 78 wasn't simply a better answer to the question AlphaGo thought it was solving. Instead, it changed the question.

There was a board

Move 78 didn't arrive as a paragraph explaining a theory. Its meaning depended on where it was. Lee placed a stone somewhere that existed relative to other stones, territory, threats, opportunities, previous moves, and possible future moves. That matters because humans don't reason solely through language. We also reason spatially. We understand position and proximity, movement and boundaries, and we notice when something has moved, when something appears out of place, when a gap opens, or a relationship changes. Sometimes, we even see something before we can explain why we have seen it. The distinction between chess and Go makes this particularly interesting. Dave Snowden wrote about this in 2009. Chess, he suggested, is complicated; Go is complex.

Chess has many possible moves, but the pieces themselves have differentiated identities and values. A queen is a queen, and a pawn is a pawn. Go is different, though. The stones are identical. What matters is where they are. A stone placed early in a game can acquire significance hundreds of moves later because its meaning comes from the changing pattern around it. Therefore, context of position matters more than the intrinsic nature of the piece. This feels increasingly relevant to AI. Try describing a complicated chess or Go position entirely in prose and then put the pieces on a board. The information may technically be identical, but the cognitive experience is not.

The tyranny of the chat box

We have built one of the most remarkable technologies in human history, and our dominant interface to it is a text box. We type, and then the machine responds. We type again, and then the machine responds again. This feels completely natural because language is perhaps our most important mechanism for exchanging ideas. However, there is a danger in that the conversation can become the model.

Language is extraordinarily good at creating the appearance of coherence. An AI can describe a strategy beautifully without either the AI or the human having a particularly good representation of the landscape in which that strategy must operate, yet something can still be badly positioned. We may not notice because we are reading sentences rather than looking at a landscape. We need something else. This is one reason I've become increasingly interested in combining AI with something I've worked with for many years: Wardley Maps.

A Wardley Map begins with a user need and then maps the chain of components required to meet that need. Unlike many architecture diagrams, position has meaning, as things exist in relation to users. They also exist somewhere along an evolutionary path, from the novel and uncertain toward the increasingly understood, standardized, and eventually commoditized. The result moves beyond a diagram of components to become a landscape. And, over time, landscapes move.

That's important because strategy is fundamentally contextual. Something that makes sense when a capability is novel can be misguided once that capability has become industrialized. For instance, something that differentiates you today might become universally available infrastructure. Recently I've been spending more time exploring what happens when you bring AI and Wardley Mapping together. I've begun to suspect that maps provide something important that the chat interface can’t. They create a shared cognitive space.

Asking the machine what we're not seeing

Consider the difference between asking AI, “What's our AI strategy?”, and seeing the landscape and asking, “What am I not seeing?” Those are very different questions. With a landscape, we can then start asking AI additional questions by challenging this map. What assumptions am I making? Which components are incorrectly positioned? What dependencies are missing? Where is inertia affecting us? What happens if this component rapidly commoditizes? What happens if a competitor gives this away? What is changing that I've treated as static? What doesn't make sense? 

Those questions turn AI into something other than an answer machine. They make it an explorer of the landscape. That seems like an extraordinarily useful role for AI, but there is a danger in assuming that this relationship runs only one way. The AI isn't the only participant who can see something unexpected.

We also need to ask: Where is the human's Move 78? This is where the distinction between humans in the loop and humans on the loop becomes critical. If we reduce the human to approving AI recommendations, we're assuming that the machine generates possibilities and the human validates them. Move 78 gives us another model: Humans can generate the anomaly. Lee Sedol had spent several games facing an opponent unlike anyone he had faced before. He was learning the machine, and the machine was changing how Lee played. Lee was developing a model of AlphaGo while playing against it. He was watching, adapting, and searching. Eventually, he found something outside the machine's expectations. This is a learning system where two different forms of cognition were changing each other.

Visual-spatial cognition

This is where Wardley Mapping becomes especially interesting to me. We tend to describe maps as ways of communicating information, but I think that undersells them. Spatial representation changes how we think.

When AI suggests moving a component on a Wardley Map, the human doesn't merely receive another paragraph of text. We see where something changed, what depends on it, and what might move next. That recruits our visual-spatial cognition.

Instead of:

prompt → prose → decision

we get something closer to:

landscape → movement → anomaly → attention → interpretation

That difference is significant. The machine may be far better than we are at searching certain spaces, but humans remain remarkably good at looking at a landscape and noticing that something is odd. We might not even know why it's odd yet. We just see it. Sometimes “that's odd” is the beginning of Move 78.

Seeing before explaining

Discussions of AI tend to equate reasoning with the ability to articulate an explanation, but human cognition doesn't always work that way. Sometimes we see something before we can explain it. In a game of chess, A player looks at a board and senses that a position is dangerous. The same goes for an experienced engineer who looks at an architecture diagram and notices that something feels wrong, or a military commander who looks at a map and sees vulnerability.

Malcolm Gladwell popularised a related idea in Blink: that people can sometimes make remarkably accurate judgments from very thin slices of experience. What looks like intuition is often compressed expertise, patterns accumulated over years and recognized before they can be consciously articulated. However, that does not mean such judgments are always right. Gladwell is equally interested in the ways rapid judgment can be distorted by bias, but the important point is that cognition does not always begin with explanation. Sometimes recognition comes first, and reasoning catches up afterwards.

Visual-spatial representations give that kind of cognition somewhere to operate. That is why I think maps could become important interfaces between humans and AI, as they give human cognition another way to participate. This makes me think differently about the role of maps in AI systems. Perhaps a Wardley Map is a way of giving humans and AI a shared cognitive object. Something both can inspect, alter, and challenge. Crucially, something about which they can disagree.

If the AI places something on the map, I can ask: Why did you move this? What assumption are you making? What evidence would prove you wrong? What changes elsewhere because this moved? And the AI should be able to ask me equivalent questions. Why did you put that there? Why do you think this is stable? Why are you assuming this capability differentiates you? Why haven't you considered this dependency? Now we're no longer asking the machine for answers, but rather exploring a landscape together.

Modern generative AI can generate surprising ideas effortlessly. Surprise isn't the same thing as insight. A language model can suggest something nobody has thought of. It can also suggest something nobody has thought of because it's nonsense. The challenge is to create conditions in which humans can inspect what the machine proposes. That's another reason the map matters. The suggestion has to exist in relation to a landscape. If the AI says something is moving, we can ask why, but the same thing applies in the opposite direction. The map gives the human somewhere to place an intuition. Something doesn't look right. Something feels out of position. A relationship appears to be missing. A change elsewhere suddenly makes an existing assumption questionable. Before we can fully articulate why, we can sometimes see that something is wrong. That is part of what visual-spatial cognition gives us, and it's something I don't want us to lose as we automate more of our reasoning.

Awareness and the outlier

There is another useful idea in that earlier Cynefin writing about Go: anticipatory awareness. In complex environments, prediction is often difficult. We may not know exactly what will happen next, but we can remain alert to the fact that the landscape is changing and that something unexpected may be emerging. That shifts the emphasis from prediction to perception. The question is no longer simply, “What do we think will happen?” It becomes, “What is changing around us? What no longer fits? What are we treating as noise because it falls outside our existing model?”

This is where outliers matter. In many analytical approaches, outliers can be treated as noise because they do not fit the dominant pattern. Sometimes that is appropriate, but in a complex system, the anomaly may be precisely what deserves attention. It can be an early signal of change, an indication that an assumption is breaking down, or an opportunity the dominant model has yet to recognize. Move 78 was exactly that kind of outlier. AlphaGo had assigned the move an extraordinarily low probability. From the machine's perspective, it barely belonged in the space of plausible moves, but Lee Sedol saw something in the position that the model did not. That is why awareness matters so much in human-AI systems. If humans are only there to approve what the machine already considers likely, then their attention is directed toward the model's center. The more valuable role may be at the edges: noticing the anomaly, the weak signal, the movement that does not yet make sense.

Being on the loop is therefore partly about maintaining the capacity to notice what the system has discounted. It is also about staying sufficiently connected to the landscape to recognize when something improbable is becoming significant. Move 78 mattered because someone was there to see that the unlikely move made sense.

Don't automate Move 78 away

Much of AI's value will undoubtedly come from automation. Take a process, insert AI, remove human effort, increase speed, and reduce cost. There are many places where that's exactly what we should do. Wardley Mapping itself gives us a way to think about where this is likely to happen. As activities evolve and become better understood, they tend toward standardization, industrialization, and automation, but applying that logic everywhere would be a mistake. We may be building AI systems for robustness when the environments in which they operate require resilience.

If every AI system becomes:

input → machine → action

We may optimize away precisely the moments when humans learn. More importantly, we may optimize away the conditions that allow humans to surprise machines. Imagine if we removed Lee Sedol from the game, and there is no Move 78. The machine might continue getting better as it plays itself millions of times. Its performance could improve dramatically, but the possibility of that particular human anomaly disappears. The objective shouldn't always be to remove humans from the loop. Sometimes the objective should be to design the system so that humans and machines continue to change each other's models.

The learning loop

A learning loop suggests a different architecture for human and AI systems. Not:

AI → human approval → action

And not:

input → AI → action

But something more like:

landscape → AI explores → human sees → human interprets → Move 78 → machine adapts → landscape changes → repeat

There is a cognitive step between the machine producing something and the human understanding its significance. Our challenge is that we need representations that allow humans to understand what the machine is doing while preserving their ability to see the landscape differently. That means the landscape itself is part of the system's intelligence. If AI reduces everything to text, we minimize the modes of cognition. Likewise, if mode includes landscapes, maps, relationships, and movement, we create opportunities for other forms of cognition to participate. Sometimes that representation will be a Wardley Map. Sometimes it will be something else. I increasingly think visual-spatial cognition will be an important part of how we design human and AI systems. The chat box cannot be the final interface.

Rethinking Move 78

This brings me back to Rebels of Reason. It's tempting to tell the story of AI as a progression toward machines that know more, reason faster, and increasingly outperform us. There is plenty of evidence that this will happen across many domains, but Move 78 suggests that capability isn't the whole story. A human had spent several games observing and adapting to an unfamiliar intelligence. He was learning how the machine behaved. He was developing a model of his opponent. Then he looked at the landscape and saw something the machine considered extraordinarily unlikely. He placed a stone on the board, and the machine's model failed. That's the part of the story that feels increasingly important to me.

The lesson isn't that humans will always find a magical move to defeat a superior machine. Nor is it that intuition should somehow trump machine intelligence. The lesson is that in complex environments, no model completely contains the landscape. Meaning comes from relationships as much as from individual objects. Under those conditions, anticipatory awareness matters.

As I experiment with AI and Wardley Maps, I'm therefore less interested in asking AI for the answer. I'm interested in creating a shared landscape where the AI can challenge what I see, and I can challenge what it sees. The machine proposes something. The human notices something. The map changes. Both models adapt. Then the process happens again. Occasionally, hopefully, someone sees something the other didn't.

That is a much richer vision of human-AI collaboration than simply adding a human approval button at the end of an automated process. Our job is to build shared landscapes where those moments can happen, to maintain anticipatory awareness, to preserve our capacity to notice outliers rather than automatically eliminate them, and to ensure humans can still see the system well enough to recognize the unexpected when it appears. That's why I don't just want humans in the loop. I want humans on the loop, watching the landscape, questioning it, changing it, learning from the machine, and still capable of Move 78.