A philosopher saw the flaw in early AI before the engineers did
Organizations now hand decisions about diagnoses, credit and hiring to software. Beneath the noise about artificial intelligence, every one of those handoffs assumes an answer to an older question: what is intelligence? The answer is philosophical before it is technical, and the engineering follows from it.
When the philosopher Hubert Dreyfus criticized AI research in the 1960s and 1970s, he was pointing at an assumption about what it means to be a mind in the world. Many researchers dismissed him as a philosopher throwing stones at the engineers' parade. He argued that if the assumption was wrong, the whole enterprise built on it would fail. Dreyfus took his image from Heidegger: a person understands a hammer by hammering, and a list of its properties and uses comes later, when the hammering stops.
We are always already in the world
The insight comes from Martin Heidegger's Being and Time (1927) and his idea of Dasein, "being there," his term for the distinctly human way of existing.
For Heidegger, we are always already in the world: engaged, involved and coping before we ever stand back to reflect on it. The seventeenth-century French philosopher René Descartes had pictured an isolated mind that receives sensory data and must work out, from the inside, how it connects to an external world.
Heidegger distinguishes two ways things show up for us. When we are absorbed in using equipment, such as hammering a nail, typing or driving, things are ready-to-hand. The hammer withdraws from attention. We experience hammering, and the line between us and the tool disappears. Only when something breaks down, when the hammer cracks, the key sticks or the car stalls, do things become present-at-hand. Only then do we step back, observe them as objects with properties, and reason about them.
What researchers now call good old-fashioned AI assumed that human intelligence meant manipulating internal symbols that represent external objects, according to rules that could be written down in full. Everything was present-at-hand, a fact with properties that could be encoded and manipulated.
People meet the world through what they are doing. When I walked into my college math lab in 1990, I moved toward the possibility of going through the doorway. A symbolic system would have needed "a doorway of certain dimensions, at certain coordinates, in the open position" before it could act at all. My body responded to what the psychologist James Gibson called an affordance, the possibility for action an environment offers. The doorway invited me in. The world shows up already meaningful, organized by my concerns, my projects, my skills and my body.
The frame problem shows why the difference matters
The frame problem asks how a system knows, when something changes, which of its beliefs need updating and which can stay. Move a block across a table, and the table, the floor and the walls stay where they are. A system built on explicit facts has to be told that, and has to be told that moving the block does not change the temperature in Beijing or yesterday's weather. The list never ends.
Dreyfus, following Heidegger, showed that the problem only arises if you start from isolated, meaningless facts. For embodied beings it does not arise. Our past experience of moving things is settled into how the world appears to us, and we perceive what is relevant without computing it.
Large language models go around the problem and stay outside the world
Large language models are interesting because they seem to sidestep some of these problems. They are trained on so much human writing that patterns of relevance emerge statistically, with no hand-built list of facts to keep current.
They are still not in the world. They have no bodies, and no needs or purposes that come from being mortal, vulnerable creatures in a physical environment. They process information about the world at enormous scale, and nothing matters to them. Heidegger's word for the attunement through which things come to matter to us was Befindlichkeit. Mattering, it turns out, matters.
Delegating judgment to software is a human decision
Organizations now use AI systems in decisions that change people's lives: diagnoses, loan approvals, sentencing, immigration and hiring. In doing so they delegate judgment to systems with no stake in the outcome and no experience of being denied care, refused credit, wrongly imprisoned or passed over for a job.
These systems are capable, and capability is different from wisdom. Wisdom comes from being a finite, embodied, fallible person who must cope with an uncertain world, makes mistakes, bears the consequences and learns what matters by living. Dreyfus's critique sounded cranky in the 1960s. Sixty years later, in another AI summer, it sounds like a warning. We can build systems that are extraordinarily good at specific tasks. Systems with common sense, contextual understanding and ethical judgment would need to share our way of being in the world, and that is a human question with moral consequences as much as a technical one.
The philosophical backstory explains the mistake
The Cartesian picture divided mind from body. Thinking happened inside, and on Dreyfus's account of that picture the world outside had no meaning until the mind supplied it. If intelligence is internal representation plus rules, building artificial intelligence becomes an engineering problem: formalize the rules, encode the representations, add processing power. That is what the early AI programs set out to do.
Princess Elisabeth of Bohemia pressed Descartes on the difficulty in 1643: if mind and body share nothing, how does deciding to raise an arm raise it? The Dutch philosopher Baruch Spinoza later answered that there is one substance, and mind and body are two ways of describing it. He escaped Descartes' dualism and kept Descartes' faith in pure reason.
The German philosopher Edmund Husserl tried to rescue Descartes by turning to consciousness itself and studying how experience is structured. That turn was powerful, and it still began from an isolated subject trying to make sense of experience. Heidegger, Husserl's student, rejected the question itself. Being in the world is the starting point that makes reasoning possible, and a mind that wonders whether a world exists is already standing in one.
When Dreyfus met AI researchers at MIT in the 1960s, he recognized Descartes' picture turned into a research program. That was why the early approach was doomed. The engineers were capable and the machines were improving. The conception of intelligence was wrong, and no implementation can fix an error of conception.
Leaders who misunderstand intelligence will deploy systems that miss what matters
A leader who does not understand what human intelligence is, a way of being embedded in meaningful situations and coping with them, will keep making category mistakes. The result is systems that manipulate symbols brilliantly and miss context, optimize metrics and miss what matters, and process information at scale without judgment.
The question worth sitting with is whether we are repeating the old mistake at a larger scale, with more money and more at stake. Clear sight about what we are building, why, for whom and at what cost is the place to begin. Deciding which judgments stay with people, and who answers for each one, is organization design.
If your organization is deciding what software should own and what people should keep, read the field report from AI in Action, or request a briefing on Outcome-Owned Architecture.
William Haas Evans
Founder and chief strategy officer of Fugue Strategy Advisors. He works with leadership teams on strategy design, strategic foresight and organization design. More about William
