Amongst the recent headline grabbing hullabaloos about AI, there was an outpouring of academic angst. Twenty five Fields Medal winners (the Nobel prize equivalent for Mathematics) wrote in an open letter that AI could ruin the foundations of Mathematics. I think they have a point which relates to broader issues beyond maths, about learning and creativity for startup founders.
Their point was that AI firms are developing models to solve complex mathematical problems merely to benchmark the power of their own models, rather than in the spirit of genuine intellectual inquiry. They argue this is detrimental to the science of mathematics.
Their letter was in response to AI making a number of stunning breakthroughs, the latest announced by OpenAI that they had solved the Navier-Stokes problem, one of seven Millennium Problems chosen by the Clay Mathematics Institute as the hardest and most important unresolved mathematical challenges in 2000.
OpenAI seems to have gazumped Tristan Buckmaster and Levent Alpöge, two mathematicians who had been labouring on this task, and were thought to be close to a solution. OpenAI released a 166-page paper describing its solution, which employed models not yet publicly available. Despite its length however, the paper contained little of the deep explanatory detail, underlying theory or reasoning, which mathematicians traditionally provide when sharing a new discovery.
I thought this spat sounded like an intellectual version of the tutting and raised eyebrows that accompanied the introduction of electronic calculators. Mathematicians claimed this would cause the loss of mental arithmetic skills by creating a lack of sensory impact on the brain, which might have adverse consequences for numeracy. Those concerns, real enough at the time, now sound unfounded.
The concern here is that definitive answers to problems are just one aspect of maths. What is often more valuable is the process of arriving at important proofs leads to the asking of more questions and the creation of new areas of study. Iteration and retrospectives risk being lost if AI does the deep work, and doubly so if that work then isn’t released in sufficient details for others to study and learn.
As Hugo Duminil-Copin, one of the signatories, put it: airdropping someone on the summit of Mount Everest is rather different from climbing it. Whilst AI has to be part of helping humans gain new knowledge, mathematicians may start to work less openly, in order to avoid the fate of Buckmaster and Alpöge. Another, more subtle danger is that leaning on AI too much will erode habits of mind and make mathematicians less capable practitioners in their field. AI causes the mind to atrophy.
There is a historical analogy. Research by Betsy Sparrow, Professor of Psychology at Columbia University, confirmed that people turn to Google searching for answers when faced with tough questions, instead of attempting to recall what they know. They outsource their memory. Knowingly or not, the brain had calculated that if the information existed somewhere retrievable, the effort of retaining it internally wasn’t worth making. Sparrow called this the Google effects on memory.
This isn’t automatically a problem. External storage is how civilisation works. We offload memory to books and thus to other people. The issue is when the thing being offloaded is not just storage but cognition, the actual process of thinking through a problem and when that cognition is the basis of someone’s value. A solicitor, a geologist, a medical consultant with twenty years of pattern recognition: their value isn’t their access to information, it’s the quality of their judgment. The specific, hard-won, contextually nuanced thinking that comes from having done this long enough to know what good and the exceptions look like.
If AI progressively takes over the cognitive work that builds and maintains that judgment, Sparrow’s findings suggest it doesn’t sit dormant waiting to be called upon. It degrades, exactly as surgical skill degrades when a surgeon stops operating. The critical thinking, the contextual judgment, the professional instincts built over years don’t maintain themselves by default or automatically. They require deliberate, active use to stay sharp.
The Sparrow research suggests you may not notice the erosion until it has already happened. By which point the most valuable thing you brought to your startup work – your unique judgment – has quietly left the building. Another study, by Michael Gerlich of the SBS Swiss School, similarly found a negative correlation between the frequent use of AI tools and critical thinking abilities. He created the term Cognitive Offloading: When you let AI do all the thinking, your brain stops building deep understanding, critical thinking and problem-solving skills.
I’ve seen this happening with founders, many using AI way beyond the limits of reasonableness to curate their value proposition and investor pitch deck. Instead of their own original thought they use AI because they think it’s smarter and creates a more elegant prose and presentation. Instead, it’s simply navigating as Sparrow found to blandness and one-size-fits-all. Gen AI operates on statistical probability, predicting the most likely next step based on vast training datasets. For founders whose edge relies on non-linear thinking, differentiation, and contrarian insights, over-relying on AI creates distinct strategic vulnerabilities:
1. Strategic homogenization (‘regress to the mean’). AI models are trained on historical Internet data, their output naturally gravitates toward consensus. If ten founders use the same prompts to research value propositions, they will receive minor variations of the exact same output.nbsp; You end up building a statistically average company that sounds like everyone else, missing the subtle, counter-intuitive insight that creates a moat.
2. Premature anchoring and lazy thinking. When faced with a blank page or a complex problem, generating ten AI ideas in three seconds instantly creates an anchoring bias. Human brains naturally seek efficiency: once an AI presents a plausible sounding strategy, your brain stops doing the hard work of original thinking.You skip the mental struggle where breakthrough, original ideas are formed.
3. Loss of intuitive edge and tacit knowledge. Breakthrough startup strategies often come from lived human experiences, emotional resonance, and unquantifiable domain insight nuance, things AI cannot feel or experience. Outsourcing research and strategy to AI removes you from exposure to raw human signals, actual customer pain points, and edge-case insights.
4. Erosion of first-principles reasoning. Using AI to answer How should we solve Y? simply delivers aggregated pre-existing responses efficiently, rather than fundamental truths and new insights which should be the core of a founder’s new venture. Over time, founders who rely on AI risk deskilling their own ability to break down complex systems from scratch.
A Founder Operating System for AI
To capture AI’s efficiency gains without losing original creative authority, adopt a human-first, AI-second workflow.
1. Core Philosophy: AI is an Assistant, Not the Author. AI excels at synthesis, pattern matching, and stress testing. It struggles with non-linear thinking, lived experience, and contrarian insight. Founders should use AI to amplify their human creativity, not to replace it.
2. Start with a ‘human ideation sandbox’. Never open an AI tool at the start of a new strategic initiative, force yourself to spend at least thirty minutes writing, mapping, or sketching raw thoughts, hypotheses, and crazy ideas unassisted. Only introduce AI after your initial thesis is on paper.
3. Position AI as a cynical adversary, not an Oracle. Instead of asking AI to generate solutions, apply it to challenge your original thinking, to stress-test hypotheses, failure modes, and hunt for blindspots in your thinking. Using AI as a cynical adversary sharpens your original idea.
4. Develop a contrarian prompt strategy. When using AI for ideation, explicitly instruct it to break away from statistical consensus – taking a countering, contrarian perspective will add to your thinking. For example:
- The Adversarial Prompt: I am proposing the following strategy for [project/product]: [paste your raw human thesis]. Act as a skeptical VC/cynical lead engineer/conservative customer. Identify five subtle assumptions, logical flaws, or edge-case failure modes in this thesis that I could be overlooking.
- The Contrarian Prompt: We are tackling [problem/feature]. The standard approaches are [insert list]. Provide five counter-intuitive, unconventional approaches to this problem that 90% of industry incumbents would consider absurd or impractical. Explain the logic behind why each might actually work.
- The Perspective Prompt: We are designing [feature/flow] for a B2B SaaS platform. How would a luxury hotel concierge/high frequency trading broker/arcade game designer approach this same user experience problem?
5. Ground Strategy in direct customer realities. AI can process public internet data, but it has no access to private, real-time human conversations. Ensure your core thesis relies on primary qualitative data — verbatim customer interview quotes, sales call feedback, or lived operational friction and testing, rather than secondary AI research. AI summarises consensus but breakthroughs lie in edge-case customer pain points that haven’t surfaced on the internet yet.
6. Diversify your model usage. Never rely on a single model or ecosystem. Different models have distinct fine tuning, biases, and structural strengths. Running a core thesis through multiple distinct AI models exposes where outputs overlap (the generic baseline) versus where edge cases emerge.
Summary: My concern for founders is that they let AI become the default prop because everybody else is following this path, and that like the Mathematicians concern, little by little it becomes our inner authority. All this will do is synchronise our thinking and homogenise startup ideas.
We already ask it what to write, how to respond and what decision makes sense. AI can make us appear to be extraordinarily intelligent without making us more conscious, and the more convincing its answers become, the easier it may be to stop sensing for ourselves.
Does AI strengthen founder insight and discernment, or slowly replace it? Does it help us meet uncertainty and risk in startup ventures, or make dependency feel like intelligence?