Our Blog

Most of us are using AI every day. We appear to be  producing more, moving faster, and feeling increasingly capable. And yet something nags.

Are we actually thinking better? Or are we just thinking less – with better-looking results?

The research now has an answer. And it points in a direction most of us would rather not look.

Fast, slow – and artificial

In 2011, Daniel Kahneman gave us a map of the human mind. System 1: fast, intuitive, emotional – the part of the brain that makes snap judgments, reads a room in seconds, and finishes familiar sentences before they are complete. System 2: slow, deliberate, analytical – the effortful mind that calculates, compares and reasons.

For fifteen years that map has shaped how we understand decision-making, leadership and human error. It is still right. It is just no longer complete.

AI has added a third system. A paper published in early 2026 from the Wharton School of the University of Pennsylvania makes a compelling argument that the way we think about thinking is now out of date. Their argument: the dual-process model now has a fundamental gap. AI constitutes a third cognitive system – external, algorithmic, embedded in how we work and decide. They call it System 3.

System 3 is not intuition and it is not deliberation. It is something qualitatively new: vast in processing capacity, free of emotional colouring, immune to fatigue, capable of sounding authoritative regardless of whether it is right.

Used well, it extends our capability beyond what was previously possible. Used poorly, it quietly replaces it. And the uncomfortable finding from the Wharton experiments is that when people stopped applying their own judgment and simply followed AI – even when AI was wrong – their confidence went up anyway. They were wrong. And certain.

The system that looks like thinking

There is a subtlety to AI that most of us miss, and it matters enormously.

Large language models are not, architecturally, careful deliberate thinkers. They are closer to what researchers have called a ‘giant pseudo System 1’: extraordinarily fast pattern-matching engines that have absorbed vast amounts of human thought and learned to reproduce its surface characteristics – fluency, structure, apparent logic – without necessarily the underlying substance.

This creates a specific trap. When AI produces a confident, well-structured response, our own System 1 tends to accept it. The output looks like careful thinking. The brain registers it as careful thinking. And System 2, which might otherwise scrutinise it, disengages.

Andy Clark, the philosopher who first argued that the mind already extends beyond the skull, addressed this directly in a 2025 paper in Nature Communications. We have always offloaded cognitive work to tools – writing, calculators, search engines. AI is the latest in that arc. But Clark argues this extension demands what he calls ‘extended cognitive hygiene’: applying to AI outputs the same critical standards we would apply to our own thinking. We cannot outsource judgment without first developing the judgment to know when outsourcing is appropriate.

From offloading to surrender

Not all AI use is the same. There is a spectrum, and where we sit on it determines everything.

Cognitive offloading is deliberate and purposeful. If you are writing a book, for example, asking AI to verify the academic research behind a chapter while you focus on the argument is a good trade. It saves hours and you stay in control of the thinking.

Cognitive debt is what accumulates when offloading quietly becomes dependency. Six months later, you realise you can no longer start a chapter without first asking AI to structure it for you. You are still writing. But the thinking that shapes the argument has slipped away without you noticing.

Cognitive surrender is the endpoint. Submitting a chapter where AI wrote the argument, structured the ideas and chose the examples – and you changed a few words and called it yours.

The most unsettling finding is this: cognitive debt builds silently. The person experiencing it almost never notices. Which is precisely what makes it dangerous at organisational scale.

Why the results aren’t showing up

The individual gains from AI are real. Studies consistently show productivity improvements of 14 to 55% on specific tasks. So why aren’t those gains showing up where it matters?

The macro data is striking. A survey of nearly 6,000 executives across the US, UK, Germany and Australia found that over 80% of firms report no measurable productivity gains from AI despite rising adoption. McKinsey found that 94% of companies with deployed AI are not seeing significant value from their investment. The UK government’s controlled trial of Microsoft Copilot found that users completed individual tasks faster – but there was no evidence those time savings had improved overall productivity.

The time saved on tasks is being consumed elsewhere. In verification overhead. In iterative refinement that goes well past the point of good enough. In the elaboration loop that AI sycophancy creates.

Individual efficiency is rising. Organisational value is not following. And the gap between the two is, in large part, a thinking problem.

The discipline that makes the difference

The organisations navigating this well are not using AI less. They are using it differently. The most effective personal discipline, backed by research published by Harvard researchers, is straightforward: form your own view first. Write it down. Identify what would surprise you. Only then consult AI – and use it to challenge your hypothesis, not generate it.

Hypothesis first. AI second. That single practice, consistently applied, is the most powerful guard against cognitive surrender.

After using AI, it is worth building the habit of asking three questions:

• Did I think before I prompted?
• Did I accept this because it was right – or because it was confident?
• What would I have concluded without it?

The research is clear: most of us are largely unaware of our own cognitive surrender as it happens. Awareness has to be deliberately cultivated. That is what metacognition means in practice.

Mindset matters more than most people realise

There is a deeper variable underneath all of this, which we will explore in a future article: mindset.

Carol Dweck’s growth mindset research turns out to intersect powerfully with how people use AI. Those with a fixed mindset tend to use AI as a performance tool – to produce better-looking outputs more quickly, without investing in understanding why the output is good or developing the judgment to know when it is wrong. Those with a growth mindset use AI as a development tool – a thinking partner that challenges their reasoning, exposes their blind spots and accelerates their learning, while they retain ownership of the judgment.

The first approach produces polished output. The second produces a better thinker. They are not the same thing, and over time the gap between them compounds.

What this means if you lead people

Leaders face a specific version of this challenge. High status, high time pressure, significant investment in appearing to have the answers: these are precisely the conditions that make cognitive surrender most tempting – and most costly.

Our emotional and cognitive states are genuinely contagious. Research consistently shows that a leader’s internal state shapes the psychological climate of everyone around them. A leader who privately surrenders judgment to AI while publicly presenting AI-generated conclusions as strategic thinking does not keep that habit to themselves. It radiates. It becomes the culture.

The most capable leaders of the AI era will not be those who use AI most. They will be those who use all three systems – intuition, analysis and AI – in the right sequence, for the right tasks, with the awareness to know which is which.

The question I ask myself – and the one I’d encourage you to sit with – is not ‘am I using AI?’ It is: when I use it, am I still thinking?

Continue the conversation

Join the live webinar

The AI Productivity Paradox: Why AI Adoption Isn’t Translating into Organisational Impact

Join Tom Flatau on 7 July for an interactive discussion exploring the research behind AI’s productivity paradox, the emergence of “System 3” thinking, and what leaders need to do differently.

Register Now

Book a Conversation with Tom

Tom works with organisations worldwide on AI adoption, leadership, mindset and organisational culture, helping leaders translate neuroscience into lasting behavioural and business performance change.

Schedule Now

Tom Flatau

CEO, Teamworking International  |  International Keynote Speaker

Tom has spent 25 years applying neuroscience to leadership and business. He works with organisations navigating AI adoption, leadership development and organisational change. Clients include BBC, HSBC, Unilever, Siemens, Emirates and Kier Group.