Are We Building AI with Half a Mind?
Sunday AI Reflections
A couple of years ago I wrote about Iain McGilchrist's work on the divided brain and what it might mean for risk and compliance. I argued that organizations need whole brain thinking, bringing together the parts of the business that protect against loss with the parts that pursue new opportunities.
I have been returning to McGilchrist's work as I watch how organizations are approaching AI. What I see concerns me more now than it did then. I believe the way business attends to the world and the way AI is being built are reinforcing each other.
There is also an important part of McGilchrist's argument that I did not give enough weight to the first time. It is missing from the conversation about AI today.
Two Ways of Attending
McGilchrist's work is often confused with the popular idea that the left brain is logical and the right brain is creative. That is not his argument. Both hemispheres are involved in almost everything we do. What differs is the kind of attention each brings to the world.
The left hemisphere attends narrowly. It isolates what it needs, names it, categorizes it, and fixes it in place so that it can be used. Its purpose is to get and to grasp. It works with representations of the world such as maps, models, words, and numbers. These are powerful tools, and we could not build anything without them.
The right hemisphere attends broadly. It takes in the whole situation, including what is new, what is changing, and what is implied. It sees things in relation to one another and in the context that gives them meaning. It is open to what is actually present, including what it did not expect to find.
McGilchrist's central point is that these two ways of attending are not equal partners. The left hemisphere's work is meant to be in service to the right. The right hemisphere sees the whole, the left takes a part of it to work on, and the result is returned to the right to be understood in context. Problems arise when the narrow view is taken to be the whole picture.
Business Has Forgotten the Order
In my earlier article, I noted that business in particular relies heavily on the left hemisphere's way of attending. That is not surprising. Business needs to plan, measure, control, and improve. Management systems exist to reduce variability and make outcomes more certain. These are left hemisphere strengths, and they are necessary.
What has been forgotten is the order. Measures, procedures, and controls are meant to serve purposes that come from a fuller understanding of the organization and the people it affects. Over time, the measures become the purpose. Conformance to procedure stands in for keeping the promises the procedure was meant to support. The map is mistaken for the territory.
I have written about this before as Compliance 1, where the aim is to follow the rules and produce evidence of having done so. It is what the left hemisphere's way of attending produces when it is left to run the organization on its own.
AI Doubles Down
Today's AI is built almost entirely in the left hemisphere's image. Large language models work with representations of representations. They are trained on text and data that have already been separated from the situations that gave them meaning. They categorize, predict, and optimize. They are very good at producing outputs that look right according to patterns found in the data.
This is why AI fits so easily into business as it is practised today. It promises more of what business already values, including efficiency, consistency, scale, and control. Organizations adopting AI are not being asked to change how they attend to the world. They are being offered a faster way to do what they already do.
That may look like a good match. My concern is that it amplifies an imbalance that was already there. If the left hemisphere's way of attending was already running ahead of its purpose, AI gives it more reach and more speed, with even less connection to the whole.
Context Is More Than a Prompt
In AI circles, context has come to mean something very specific. It refers to the information placed in a model's context window, such as the prompt, retrieved documents, conversation history, and instructions. When an AI system lacks context, the answer is to supply more of it. There is now a field called context engineering devoted to doing exactly that.
This is the left hemisphere's understanding of context. Context becomes more parts, gathered and assembled so that the model has more to work with.
The context McGilchrist describes is different in kind. It is the whole situation that gives the parts their meaning. It includes the people involved, what matters to them, the history behind a decision, the purpose it serves, and what is at stake if it goes wrong. This kind of context is lived and understood. It cannot be inserted into a prompt, because it consists of the relationships among things and what those relationships mean to people.
When we treat context as something that can be supplied to a machine, we stop asking who is attending to the whole. We assume the problem has been solved by adding more information.
What Should Be and What Could Be
Data can only tell us about what is, or what has been. To direct an organization, we also need to know what should be and what could be. Neither of these is found in data.
David Hume made the first point in the eighteenth century. He observed that no amount of reasoning about how things are will, by itself, tell us how they ought to be. An ought is a different kind of claim, and it cannot be derived from an is. For Hume, our oughts come from what we value and care about. Reason can help us pursue what we value, but it cannot tell us what to value.
Hume also helps with the second point. He saw imagination as the faculty that lets us go beyond what experience has given us. None of us has seen a golden mountain, yet we can picture one by joining ideas we already have. Everything that has ever been designed or built began this way, as something imagined before it existed.
AI systems are built from data about what is. However much data they are given, they remain on the is side of Hume's divide. They can produce new combinations of what they were trained on, but they cannot tell which of these should exist or why it would matter.
What should be and what could be come to us through intuition and imagination. These belong to the right hemisphere's way of attending, which takes in the whole and is open to what is not yet there. When organizations look to AI to tell them what to do, they are asking the left hemisphere's tools to answer a question those tools were never able to answer.
What We Attend to Is What We Get
McGilchrist makes an observation that I believe matters for anyone responsible for the direction of an organization. The kind of attention we pay shapes what we are able to see. It also shapes what we are able to realize.
If we attend to the world in order to get, control, and manipulate, then the world appears to us as something to be gotten, controlled, and manipulated. Customers are seen as data, employees as capacity, and risk as a number to be driven down. We find what we were looking for, and we produce more of it.
John Vervaeke's work on relevance realization makes a similar point from cognitive science. What we notice as relevant determines what we act on. When intelligence systems are built and deployed to optimize what can be measured, they shape what organizations notice. Over time, what is not measured stops being seen at all.
This is how organizations can become very efficient at producing outcomes nobody intended. Each part is optimized while the whole drifts.
People Are the Loop
The usual answer to concerns about AI is to keep a human in the loop. A person reviews what the machine produces and approves or corrects it before it goes further. This is the left hemisphere's answer. It treats the person as one more part of the process, a checkpoint placed between steps.
Seen this way, people will always be pushed out. They are slower than machines, and a checkpoint that slows the process will be measured as a cost. Every gain in AI capability strengthens the case for removing it. When the measure is productivity, the human in the loop is a bottleneck waiting to be engineered away.
What matters here is something productivity does not measure. We need people to be the loop. People provide the context that gives the machine's work its meaning, the purpose it serves, and the boundaries it must operate within. This is the context that cannot be inserted into a prompt. In engineering terms, people and the organization they form are the environment the machine operates in and the system that regulates it.
This is what I mean by AI in the human loop. The machine operates within a situation that people understand and are responsible for. Its speed then becomes an advantage to the people it serves, and no longer a reason to remove them.
Who Is Attending to the Whole?
None of this means we should abandon the left hemisphere's way of working. We need analysis, measurement, and control. We need AI systems that do specific things reliably. The question is what they are in service to, and who is responsible for that.
The right hemisphere's work in an organization is done by people. It is done by those who define purpose and keep the organization connected to it. It is done by leaders who attend to the people affected by their decisions, including people they will never meet. It is done by engineers whose duty of care extends to the public, and by anyone who keeps a promise when the procedure has nothing to say.
In practice, this work looks like several things I have written about over the past year:
Making people the loop, so that AI operates within the context, purpose, and boundaries that people provide.
Specifying the value of people and the purpose of the work as requirements, so that they are designed into intelligence systems from the start.
Making room for intuition and imagination, so that what should be and what could be are decided by people and not inferred from data.
Using the engineering method, which synthesizes toward a purpose, alongside the scientific method, which analyzes what already exists.
Regulating AI against the organization's own governance, so that its purposes and promises remain the standard the AI is measured against.
The question I keep coming back to is whether organizations know who is doing this work. In many cases, I suspect no one has been given that responsibility. It is assumed to happen somewhere, or the technology is expected to take care of it.
A Reflection for This Week
We are building AI with half a mind. The intelligence we are building attends to the parts with remarkable speed and precision. It has no way of attending to the whole, and it cannot tell us what should be or what could be. That work belongs to people.
It does not have to stay this way. When people attend to the whole and AI works within it, the speed and precision of these systems serve the purpose they were meant for. People bring context, intuition, and imagination. AI brings the ability to work with the parts at a scale we have never had before. Together, they can help organizations see further and build what should be and what could be.
That is the opportunity in front of us. It begins by asking who in our organizations is attending to the whole, and giving them the place to lead.
About the Author
Raimund Laqua, P.Eng., PMP, is the Founder and Principal Consultant of Lean Compliance Consulting, Inc., and defined the practice of Lean Compliance. He has practised digital and compliance engineering for over 40 years and is one of the first licensed software engineers in Canada. He helps organizations in highly regulated, high-risk industries keep their promises and contend with uncertainty, including the uncertainty introduced by AI.
Raimund chairs the Digital Engineering Committee for Engineers for the Profession (E4P), serves on OSPE's AI in Engineering Task Force, and is a co-founder of ProfessionalEngineers.AI. He advocates nationally for Professional Digital Engineering (PD.Eng.) licensure in Canada and has published more than 600 articles at leancompliance.ca.
