Helen Edwards from the Artificiality Institute tells the story of the researcher who, before LLMs came along, read every transcript of every interview he had ever done. He handed that job to an AI and then mid-presentation he had to stop because he realised that he couldn’t speak to any of it. The analysis might have been perfectly plausible, but he just didn’t own it.
Research from Wharton, published a few months back, called this phenomenon ‘cognitive surrender’. The researchers, Steven Shaw and Gideon Nave, argued that Kahneman’s two systems, fast intuitive judgement and slower, more deliberate reasoning, needed to be expanded to incorporate a third system. System 3 thinking happens outside of the brain when humans are outsourcing decisions or reasoning to an AI. Across three experiments with more than 1,300 participants they found that when AI was right, accuracy actually improved by 25 percentage points but when it was wrong, it fell 15 percentage points below the level of people working with no AI at all. Deference to AI was highlighted as the key problem, and the study found that even when the AI was wrong about half the time, the participant’s level of confidence went up.
For strategists cognitive surrender is almost like an existential threat. Ask AI a strategic question and it will give you a coherent, plausible-sounding, confident answer. But in doing so it has already set the framing for the conversation. Accept that, and everything clusters around one reasonable sounding answer. AI models are built to resolve tension but good strategic work requires the strategist to hold several possibilities open long enough to explore them and to hold different approaches in tension with each other. Premature convergence kills good strategy.
If your institutional answer to cognitive offloading is having a ‘human-in-the-loop’, you are likely not solving the fundamental problem. Having someone review outputs or sign something off might feel like deliberate thinking but it’s mostly just pattern matching to someone else’s checklist. Kahneman spent a career showing us that we’re often very bad at telling from the inside when we have stopped deliberating and started recognising. Fluent, plausible and confident AI outputs makes this even harder. The reviewer feels like they are thinking when they are actually ratifying.
Dr Philippa Hardman recently did an excellent job of assimilating all the recent research into the phenomenon of cognitive surrender (in a learning/education context but applicable much more broadly). She noted how studies show that cognitive offloading is typically a default rather than an exception, and the kind of AI/human collaboration that effectively avoids it only emerges organically in a minority of conversations (7.5% according to one study). Studies also showed that cognitive surrender happens when AI gives feedback before a learner reflects (‘if AI judges first, the learner never learns to judge’), and also arises when the task rewards outputs over method (deadlines and grades rather than exploration). Offloading can look like success if you only measure performance, but this hides the real, longer-term impacts.
Hardman’s conclusion is that rather than AI degrading thinking, it actually holds the potential to extend it (and her six principles are a useful checklist). I’ve written before about specific techniques for maintaining cognitive sovereignty in AI-augmented strategy, but how do you systematise this? Helen Edwards puts it this way. Summary and synthesis are two completely different things. AI is brilliant at summarising but is far less good at synthesis because whilst a summary is faithful to its sources, a synthesis is answerable to something specific like a purpose, a decision or a consequence: ‘A summary says here’s what everything said and a synthesis says here’s where I’ve figured I land and I can defend it.’
Psychologist and writer Philip Tetlock talked about the idea of the ‘Dragonfly eye’ – lots of lenses fusing into one image. The best forecasters that Tetlock studied were able to synthesise many diverse perspectives into a single, cohesive view. Synthesis carries with it a judgement about what matters, what good looks like and what you’re willing to be wrong about. It involves applying weight to different elements and those weights can only come from a person who has something at stake. Summary does none of this.
Asked to synthesise, AI will look across millions of documents and produce an averaged out summary of what synthesis should look like. The brain is a prediction machine, constantly looking ahead to guess what’s coming, and it’s the little prediction errors, the small surprises, that helps the brain to learn. Reading the raw material, says Helen, is full of those misses – the unexpected contradictions, asides, comments, disclosures. That’s what synthesis is.
So given this, here are my (seven) principles for improving cognition when using AI in strategy:
1. Match the approach to the context
The intuition when trying to avoid cognitive offloading is to use AI less, and a bit more carefully. But the mistake is often in applying the same casual approach to different kinds of decisions or contexts. Two key questions: how reversible is this decision, and what other things will be built on top of it? A one-way door decision needs much greater depth of thinking and human involvement. A two-way door decision is reversible so you can move quicker, bring AI deeper into the process, and course correct if necessary.
2. Push back on convergence
AI can reliably expand the list of options in front of you but it cannot do the weighting. So use it for what it’s good at – giving you conflicting interpretations of the same situation, opening new avenues to explore, looking at the same thing through different framings. But push back on the model’s urge to arrive prematurely at a plausible solution or answer. Keep optionality there long enough for you to decide what the framing should be, and what will be important in reaching a resolution.
3. Measure friction, not output
Decide what evidence would be enough to act on and then go and see whether it exists. Committing to a standard, or a hypothesis, before you read the output helps to avoid the model setting the framing for the conversation and you just accepting it. If you commit to a position have the AI argue against it. And don’t settle for generic push back, ask it to challenge you from a named stance. If you judge the session based on how much the AI changed your mind or generated unexpected angles that’s a signal that you’re using it to extend or reshape your own thinking.
4. In strategy, the labour is the thinking
Delegating the labour and keeping the thinking sounds good but in strategy the labour is often where the insight is. Reading twenty interview transcripts rather than a summary of them is often where the genuine and most useful insight reveals itself. In many ways the labour is the thinking. Being involved enough in the labour of the process enables better synthesis because you can judge the weighting of the elements that requires. Ask an LLM to synthesise and what it produces is a plausible average of how conclusions tend to go. But do delegate retrieval, structuring, generation (not selection), production. Avoid the state where you’ve given away enough to stop owning it but have kept enough to still be doing all the work.
5. Focus on process, not the artefact
Strategy works to deadlines and under pressure it makes it a lot more likely that we collapse into offloading. So also consider how the output, and therefore the incentive, is being framed, not just the conversation. If everyone is focused on the artefact (the deck, a report, or a board paper) offloading will likely increase. If it is discovery and synthesis, that rewards thinking. Shaw and Nave also found that when participants were given accuracy incentives and immediate feedback, they became far more likely to question and override wrong AI advice.
6. Protect the variance
AI is very good at producing smoothed outputs which can be problematic at a team level. Challenging a rough draft feels like a helpful contribution but challenging a polished one can feel like you’re obstructing. This can mean that variance collapses in the preparation, because six people are using overlapping materials and models, but also in the discussion afterwards because a polished output has increased the social cost of disagreeing. Form a view before going to the AI, and consider assigning different starting framings so healthy disagreement is designed in.
7. Judge the reasoning
Helen Edwards notes how traders develop an intuition for the markets through years of interoception – feeling and interpreting the signals from the body from rapid feedback. It’s literally gut feel. Strategists can’t benefit from such rapid feedback so you have to manufacture it. Write down what you decided, what assumptions you had to take, what uncertainties were involved, and what would tell you the decision is wrong. It’s a mechanism by which judgement can measurably improve rather than degrade. A decision record like this can also prevent what is, I suspect, an increasingly common pattern inside organisations – an AI finding or output gets included in a deck labelled as directional but then gets repeated in other documents until it reads as what customers actually think.
Eisenhower once famously said that ‘plans are worthless but planning is everything’. What he meant was that the value sits in what the process leaves behind in the people who have to act, not in the document itself. AI is very good at producing persuasive looking plans, but it can’t do the planning for you. Outsource that and you’ll get something plausible but mediocre, with no shared understanding beneath it.
A version of this post appeared on my weekly Substack of AI and digital trends, and transformation insights. To join our community of over thirteen thousand subscribers you can sign up to that here.
To get posts like this delivered straight to your inbox, drop your email into the box below.
Leave a Reply