By Bonnie Chiu, Managing Director, The Social Investment Consultancy

We received a lot of positive feedback about our theory of knowledge in the age of AI, and a year on, we are publishing a sequel to it. As we embed more AI into our work, we realise that AI is not just changing what we know, but how we think. So – the theory of knowing – which is a process not the end product – in the age of AI.

Before the age of AI, a book that has transformed how people think is Thinking, Fast and Slow, by Daniel Kahneman. Kahneman argues that human judgement is shaped by two modes of thinking: a fast, intuitive system that generates quick impressions and decisions, and a slower, more deliberate system capable of analysis and reflection. While we like to believe our decisions are primarily rational, Kahneman shows that much of our thinking is driven by cognitive shortcuts and biases that help us navigate complexity but can also lead us astray. This simple insight has transformed how we think, and contributed to him winning the Nobel Prize for Economics, as a psychologist.

Good judgement depends not only on knowledge, but on recognising when to slow down, question our assumptions, and engage in deeper forms of thinking.

Structural shift in the economics of cognition under AI

In the age of AI, we hypothesise that fast and slow thinking are not only distinct, but increasingly diverging in their abundance, ownership, and value.

Fast thinking—rapid synthesis, recall, pattern completion, and first-order analysis—is becoming dramatically more abundant. AI systems excel in this domain, producing high-quality outputs at near-zero marginal economic cost (although at high environmental cost). As a result, the value of fast thinking is being commoditised: what once required expertise or effort is now widely accessible on demand. In this context, fast thinking shifts from being a scarce cognitive asset to an infrastructural utility—useful, but no longer differentiating.

Slow thinking, by contrast—deliberation, sensemaking, causal reasoning, and reflection under uncertainty—is becoming relatively more scarce. While humans remain essential to this domain, its production cannot be meaningfully automated because it depends on contextual judgement, values, accountability, and the integration of incomplete and ambiguous information. Unlike fast thinking, slow thinking does not scale linearly with computation.

This leads to a shift in value. The highest marginal value is no longer in generating answers, but in exercising judgement under uncertainty: deciding what matters, what trade-offs to accept, what evidence is meaningful, and what actions are justified in complex systems. This is summarised by the table below.

	Fast Thinking	Slow Thinking
Volume	Increasingly abundant	Increasingly scarce
Actor	AI excels	Humans remain essential
Value	Commoditised answers (low marginal value)	Judgement under uncertainty (high marginal value)

Two divergent futures

The structural shift in the how we think is consequential for the impact space. For those of us focused on addressing society’s biggest social and environmental issues, we actually need a lot of slow thinking. We hypothesise that there are two divergent futures. AI will either become: a substitute for thinking – crowding out slow thinking; or a scaffold for deeper thinking – enhancing slow thinking, and the answers to which future are in our hands.

Let’s delve into each of these scenarios in more details.

Substitute for thinking

There’s increasing urgency to solve all these challenges, from heightened inequality to accelerating climate catastrophe. Because AI can create answers so quickly, it compresses time, and creates overwhelm. Speed is the imperative, but it can also be seductive. AI amplifies this by flooding the system with explicit fast outputs, which creates the illusion of understanding without judgement.

A highly original article published earlier this year argues that AI’s biggest threat to humanity is not our jobs, but actually our mind. “Participants who treated AI as a starting point for their own thinking retained and even improved their cognitive performance over time. Those who accepted AI outputs passively showed a measurable decline. The difference was not in how much AI people used, but in how they used it. Skip the effortful step of making sense of what the AI gives you, and the brain’s capacity to learn weakens. Engage actively, and it can be sustained or even enhanced.” The biggest worry with AI is not just that it’s quick – but that it may reduce the amount of time people spend inhabiting uncertainty, and uncertainty is where a lot of second-order, deep thinking happens.

What we’ve learnt from Kaheman’s Thinking, Fast and Slow, is that slow thinking is activated when a problem is difficult, or we deliberately slow down. Yet humans are naturally cognitive misers; we default to conserving mental effort unless there is a strong reason not to. AI lowers the threshold for “good enough” answers, increasing the risk that we opt out of the very effortful processes through which deeper understanding is formed.

For those of us who do knowledge production for a living, the system also rewards us to use AI as the substitute for thinking. Clients increasingly expect faster turnaround times and lower costs, often assuming that AI can substitute for analytical labour that previously required sustained human engagement. At the same time, consulting and evaluation models frequently price work based on time spent, creating pressure to compress delivery cycles. The result is a structural incentive to produce more outputs with less cognitive time, inadvertently crowding out the slow thinking required for meaningful sensemaking, systems understanding, and robust judgement.

The substitution future is more likely in environments where speed, output, and efficiency are rewarded, and where access to slow thinking is concentrated among senior actors with experience and permission to reflect. This creates a potential equity dimension: junior professionals may become increasingly confined to AI-mediated fast thinking, without the experiential grounding required to develop judgement.

Scaffold for deeper thinking

The alternative future is one in which AI acts as a scaffold for deeper thinking rather than a substitute for it. In this scenario, AI absorbs much of the labour of generating first-order explanations, analysis, and synthesis, freeing humans to focus on higher-order cognition.

But this shift is not about efficiency – which is often where the conversations around AI stop. It requires a fundamental change in the nature of knowledge work: humans move from generating explanations to interrogating them—asking whether the framing is correct, what assumptions are embedded, what is missing, and how different interpretations might hold across systems, stakeholders, and time horizons.

This is qualitatively different work. It is not faster thinking, but heavier thinking. It involves sustained engagement with ambiguity, contradiction, and consequence. It is closer to judgement formation than to analysis.

And this introduces a constraint that is often underappreciated in discussions of AI augmentation: slow thinking does not scale easily across a standard working week. Unlike fast thinking, which can be fragmented and parallelised, deep cognitive work requires continuity of attention, recovery time, and periods of non-resolution. It is cognitively intensive and inherently finite. It is hard to count.

This raises an important organisational question. Even in a scaffolded future, it may not be realistic—or even desirable—for individuals to spend an entire working week engaged in this level of sustained reflective cognition. The challenge therefore is not simply to enable slow thinking, but to design systems that make it viable.

And this actually will transform our understanding of time. Time does not just get “freed up”—it gets re-valued. While AI collapses the time required to produce analysis and explanation, it does not reduce the time required for judgement. This creates a divergence between fast and slow thinking economies: one in which time becomes effectively infinite for generating outputs, and another in which time remains the binding constraint for forming understanding. In this context, time is no longer a measure of productivity, but a substrate for cognition. The quality of judgement is determined not by how quickly answers can be produced, but by whether sufficient uninterrupted time exists to interrogate, integrate, and reflect on those answers.

What next?

Find out in the second part of this blog.