Norm Murray | July 1st 2026 |
In 2016, when AlphaGo played Move 37 against Lee Sedol — a move so counterintuitive that human grandmasters initially assumed it was an error – the world didn’t just witness a machine win a game. They witnessed something harder to name. The move wasn’t in the training data. No human had played it. AlphaGo hadn’t been taught it. The machine had inferred it.
We built a learning machine. It taught itself to think.
This is the most consequential and least discussed cognitive threshold in AI history. The distinction between learning and thinking is not semantic hair-splitting. It is the difference between a system that reproduces what it has seen and a system that generates what it hasn’t. We engineered the first. The second emerged, quietly, nonlinearly, at scale, without explicit design. That emergence changes everything for how leaders understand, deploy, and govern AI. Most haven’t caught up. This article makes the case that they urgently need to.
We Taught AI To Learn. It Taught Itself To Think
The Learning Machine We Built
The architecture of machine learning is elegant and, by historical standards, comprehensible. Feed a system enough labeled data. Define a loss function. Run gradient descent until the model’s predictions minimize error. Repeat at scale. The system learns by adjusting billions of parameters, the weights in a neural network, until it reliably predicts output from input.
This is powerful. It is also fundamentally imitative. A supervised learning system trained to identify cats in images has learned the statistical signatures of cats. It has not developed a concept of ‘cat.’ Ask it about a cat rendered in a style that breaks the statistical patterns, and it fails. The system learned to recognize. It did not learn to reason.
This distinction was understood, and, critically, it was reassuring to many researchers. If AI systems could only pattern-match, they were bounded. Powerful within domain, brittle outside it. This is the theoretical foundation on which most enterprise AI strategy was built: high-value, domain-specific, bounded tools. Automation, not cognition. Assistance, not agency.
The investment thesis followed. Process mining. Computer vision for quality control. Natural language processing for customer service. Fraud detection. Demand forecasting. All of these represent legitimate, value-creating deployments of learning AI. The board decks were written around them. The ROI models were calibrated to them. The governance frameworks were designed for them. None of them prepared executives for what came next.
The Thinking Machine That Emerged
In 2022, a landmark paper from Google Research — ‘Emergent Abilities of Large Language Models’ (Wei et al.) — documented something that stopped the AI research community mid-sentence. As language models scaled beyond certain parameter thresholds, entirely new capabilities appeared. Not improvements on existing capabilities. New ones. Capabilities that did not exist in smaller versions of the same architecture, and that had not been trained for explicitly.
Chain-of-thought reasoning. Multi-step arithmetic. Analogical inference. Code generation from natural language specifications. The ability to identify logical fallacies. To revise its own reasoning. To debate. These capabilities were emergent — a word that carries precise meaning in complexity science: a property of a system that cannot be predicted from the properties of its components. You cannot derive consciousness from neurons by examining any individual neuron. You cannot derive reasoning from parameters by examining any single weight. Yet at scale, something qualitatively new appeared in both cases.
No one taught the system to think. It crossed a threshold, and thinking began.
This is not anthropomorphism. ‘Thinking’ here is operationally defined: the capacity to take known information and derive conclusions, solutions, or frameworks not explicitly present in the training distribution. When a language model solves a novel legal hypothetical it has never encountered, applies a physics framework to a biological problem, or identifies a logical contradiction in a CEO’s strategic memo, it is not pattern-matching. It is reasoning. The functional difference is real, testable, and strategically significant.
Daniel Kahneman’s framework is useful here. System 1 thinking is fast, automatic, pattern-driven. System 2 thinking is deliberate, sequential, effortful, the mind reasoning through unfamiliar problems step by step. For decades, AI operated entirely in System 1 territory. The emergence of chain-of-thought reasoning, self-consistency checking, and multi-step inference marks AI’s entry into System 2 territory. Not because anyone designed it that way. Because scale changed the rules.
The Cognition Asymmetry
Here is the strategic problem. Most organizations are still deploying AI as if it operates in System 1. Their governance models, use-case roadmaps, ROI calculations, and risk frameworks are calibrated to a learning system. The actual system they are operating is increasingly a thinking system. This gap — between the AI executives believe they have and the AI they actually have — is what I call the Cognition Asymmetry.
The Cognition Asymmetry has three dimensions. The first is capability underestimation. Organizations systematically underestimate what their AI systems can do because they are measuring against learning-AI benchmarks: speed, accuracy, recall. When a thinking AI drafts a strategic scenario the management team hadn’t considered, or identifies a logical inconsistency in a board presentation, it is operating outside the measurement framework. The value, and the risk, goes unrecorded.
The second dimension is governance mismatch. Frameworks for learning AI focus on data quality, model bias, output accuracy, and explainability. These are the right questions for pattern-matching systems. Thinking AI raises different questions: What reasoning is it applying? On what values is it arbitrating between competing conclusions? When it exercises judgment, whose judgment is it approximating? These are not software questions. They are leadership questions. Most governance frameworks do not ask them.
The third dimension is authority displacement. This is the one executives are least willing to discuss, and most urgently need to. The defining capability of senior leadership has historically been judgment: the synthesis of information, experience, and values into a decision under uncertainty. Learning AI doesn’t contest that. It supports it. Thinking AI does something different. It generates judgment. Not always better judgment than an experienced executive. But often faster, broader, and scalable in a way no human team can match. The authority of the senior leader is not eliminated. But it is, for the first time, contestable.
What the Boardroom Missed
This was not, in retrospect, unforeseeable. The economics of language model training followed a consistent pattern: more compute, more data, more parameters, better performance. But ‘better performance’ masked a category shift. The performance curves weren’t linear. They were punctuated, long periods of incremental improvement, then sharp discontinuities where qualitatively new behaviors appeared. The AI research community called this ’emergence.’ Business strategy treated it as ‘progress.’ The two words carry very different implications. Progress is manageable. Emergence requires a different response entirely.
The data tells the story. In 2020, GPT-3 demonstrated that large models could generate coherent prose and basic code, impressive, but understandable as sophisticated pattern completion. By 2022, models of similar architecture but greater scale were demonstrating multi-step reasoning, causal inference, and the ability to catch their own errors mid-generation. By 2024, frontier models were solving problems at PhD level across mathematics, science, and law, not by retrieving answers, but by reasoning toward them.
The three-year gap between ‘impressive autocomplete’ and ‘contestable expertise’ is the gap in which most enterprise AI strategy was written. Many boards approved AI roadmaps calibrated to the former. They are now implementing them in the era of the latter.
The Psychological Challenge for Leadership
The cognitive dissonance this creates for executives is real, and worth naming directly. Human identity at the senior level is deeply intertwined with cognitive authority. We are paid for our judgment. Our careers were built on it. Our credibility rests on it. The proposition that a machine can reason — that it can think, in any operationally meaningful sense, is not merely a technology claim. It is an identity challenge.
Rene Descartes gave us ‘Cogito, ergo sum’ — I think, therefore I am. For three centuries, the capacity to reason was the defining distinction between human and machine. Learning AI left that distinction intact. A machine that matches patterns isn’t thinking; it’s calculating. But a machine that generates novel inferences, identifies logical flaws, and revises its own reasoning under new information, that machine is doing something that looks, from the outside, like thinking.
The executives who manage this well are not those who deny it. They are those who restructure their cognitive authority around what thinking AI cannot yet do: hold values under ambiguity, navigate political and social context, make decisions with full ethical accountability, and build the human trust that executes strategy at scale. The role of the senior leader does not disappear when AI can think. It changes. And leaders who don’t make that change consciously will have it made for them.
What Boards Need to Understand
The board’s role in this environment is not to become AI technologists. It is to ask three questions that most governance charters currently do not contain.
First: Where in our organization are thinking AI systems operating, not just learning ones? The answer in most large enterprises is: everywhere. The language models embedded in strategy tools, knowledge management platforms, and executive decision-support systems are thinking systems. Treating them as search tools or summarizers is a category error with material risk implications.
Second: Who is accountable for the judgment the AI is generating? In a learning AI world, accountability is relatively clear: the data scientist is accountable for model accuracy, the business owner for use-case appropriateness, the CRO for risk. In a thinking AI world, the system generates judgment autonomously. Accountability for that judgment is, in most organizations, undefined. That is not a technology gap. It is a governance gap.
Third: How are we calibrating the authority boundary between human and AI reasoning? This is the hardest question, and the most important. Organizations that get this right will not be those that minimize AI reasoning or maximize it, but those that define, with clarity and intentionality, where human judgment is non-negotiable and where AI reasoning can be trusted to operate autonomously. That boundary needs to be set by leaders, not defaulted to vendors.
The Position
We are approximately three years into an era that most business institutions have not yet named correctly. The machines we built to learn have become machines that think. That sentence should land with the weight it carries. Not as science fiction. Not as techno-optimism. As a material fact with direct implications for governance, strategy, leadership, and organizational design.
The executives and boards that understand this, who can distinguish between a learning system and a thinking system, who have redefined accountability accordingly, who have chosen their authority boundaries with intention, are operating in a different strategic reality than those still calibrating to 2020’s AI roadmaps.
Catching up is not primarily a technology challenge. It is a cognitive one. And in the end, that is the deepest irony of this moment: the gap between human and AI thinking is most consequential not because AI is catching up to humans, but because humans have been slow to think clearly about AI.
© 2026 Norm Murray. All Rights Reserved. No part of this publication may be reproduced, distributed, or transmitted in any form without the prior written permission of the author.
