The Knowledge: what’s still worth learning when the machine already knows the way

Since 1865, no one has driven a licensed black cab in London without passing a test called, simply, the Knowledge. To earn the badge, a candidate must hold the whole tangle of central London in their head — every street and landmark within a six-mile radius of Charing Cross, roughly 113 square miles, organized around 320 set routes in a book drivers call the Blue Book. Transport for London describes the result plainly: by the end you will have learnt “thousands of streets and points of interest.” One review of the research puts it at more than 26,000 streets. It takes, on average, three to four years. People fail, and start again.

You can be forgiven for asking what all that is for. A passenger wants to get somewhere. GPS has gotten people somewhere, reliably, for two decades. Strip away the pride and the craft — and they are real — and the encyclopedic requirement starts to look like a monument to a problem we already solved. That is the uncomfortable question the Knowledge puts to any curriculum: how much of what we still make people memorize is a service to the learner, and how much is nostalgia?

But before we answer too quickly, the Knowledge has a twist that should slow us down.

The effort was building something the map can’t

When neuroscientists put London taxi drivers in a scanner, they found something remarkable: the drivers’ brains were physically different. The posterior hippocampus — a region tied to spatial memory — was enlarged, and it grew with years on the job (Maguire et al., 2000). The skeptic’s obvious reply is that people with bigger hippocampi simply become taxi drivers. So Maguire’s team followed trainees over four years. At the start, there was no difference. Four years later, grey matter had grown in the posterior hippocampus of only those who passed the exam — not in those who trained and failed, and not in the controls (Woollett & Maguire, 2011). The learning itself, the effortful acquisition of it, had reshaped the brain.

The effort, in other words, was building something the map on the dashboard does not: a mind. And there is a suggestive coda. In a 2024 analysis of nearly nine million death records across 443 occupations, taxi and ambulance drivers had the lowest share of deaths attributed to Alzheimer’s disease (Patel et al., 2024). The authors are careful, and so are we: this is an association in death-certificate data, not proof that navigating wards off dementia. But it sits beside the brain-imaging work and makes you wonder what we spend when we stop doing hard things in our heads.

Because there is a bill. Those same expert drivers were measurably worse at forming certain new spatial memories than bus drivers were — expertise came with a trade-off (Maguire et al., 2006). And the tool that replaced the Knowledge appears to charge its own quiet fee: people who lean more heavily on GPS tend to have poorer, and declining, spatial memory when they navigate without it (Dahmani & Bohbot, 2020). Offloading is not free. (It is also not a catastrophe — the “digital dementia” headlines run well past what the evidence supports.)

So the Knowledge leaves us holding two true things at once. GPS delivers the output — it gets you there. The years of study built a capacity — a durable, expensive thing that the output never required. The interesting question for education is not which of these to worship. It is how to tell them apart.

The same question, pointed at the curriculum

Every curriculum has its Knowledge: content we ask learners to hold in their heads because we always have, some of which a machine can now summon in a sentence. When a tool arrives that answers the question we built the exercise around, the honest move is not to pretend the tool isn’t there. It is to ask, of each thing we teach: is this effort building a mind, or is it just producing an output a machine now produces better?

Here the reflex is to say “so teach skills, not facts — creativity, critical thinking, the ‘21st-century’ skills.” I want to be careful, because the evidence does not actually say that, and saying it would be its own kind of nostalgia.

Thinking runs on knowledge. The cognitive scientist Daniel Willingham puts it bluntly: critical thinking “is not a skill” you can learn once and deploy anywhere; “the processes of thinking are intertwined with the content of thought — that is, domain knowledge” (Willingham, 2007). You cannot think critically about a subject you know nothing about, and — this is the part that matters most in the AI era — you cannot evaluate what a machine tells you about a subject you know nothing about. The ability to catch the confident, plausible, wrong answer is not a generic skill. It is knowledge, wearing work clothes.

And the “transferable skills” we hoped would let us skip the knowledge mostly don’t transfer. The most sobering finding in this whole literature is that far transfer — training a general mental skill in one domain and having it carry to another — is, across chess, music, and brain-training studies, close to nonexistent once you control the experiments properly. Sala and Gobet, reviewing it, call far transfer “a chimera,” and a meta-analysis of meta-analyses concludes the effect and its variance “equaled zero” after correcting for placebo and publication bias (Sala & Gobet, 2017; Sala et al., 2019). Critical thinking can be taught — a large meta-analysis finds a real, if modest, effect — but the approaches that work embed it in subject content, not in a stand-alone skills class (Abrami et al., 2015). Competence, it turns out, is not general. It is grown in a field, from the soil of what you actually know.

Notice where this lands. It does not license us to keep every dusty fact because “knowledge matters.” It says the opposite of a shortcut: the goal is depth in things worth going deep on — the knowledge that becomes judgment. The OECD’s own framework for the future of schooling says as much. A competency, it insists, is “knowledge, skills, attitudes and values” together — knowledge and competency are “neither competing nor mutually exclusive,” and “students need to learn core knowledge as a fundamental building block” (OECD Learning Compass 2030). And its report on crowded curricula names our exact problem: we keep adding content for every new demand “without proper consideration of what needs to be removed,” until learning goes “mile-wide, inch-deep.” Its prescription is four words worth pinning above every syllabus: “learn deeper, not more” (OECD, Curriculum Overload, 2020).

That is the rebalance. Not knowledge or skills. Not shed everything the machine can do. Keep the effortful learning that builds durable, judgment-bearing minds; hand to the machine the labor that was only ever producing an output; and spend the recovered time going deeper. The taxi driver’s modern equivalent is not “memorize the map or trust the satnav.” It is knowing the city well enough to know when the satnav is wrong.

The part we keep getting wrong: banning the tool

Which brings us to AI in the classroom, where the debate has collapsed into two bad options: pretend students aren’t using it, or let them use it with no guardrails at all. The best evidence we have says both are mistakes — and it points to a third way.

In a controlled trial with about a thousand high-school students, researchers gave some students unguarded access to GPT-4 while they practiced math. Their practice scores jumped 48%. Then the researchers took the AI away for the exam. Those same students scored 17% worse than classmates who’d never had it — the AI had been a crutch, and the learning hadn’t happened underneath it. But a second group used a guardrailed tutor, designed to give hints rather than answers. That group showed no such harm (Bastani et al., 2025). Read that result carefully, because it is the whole argument in miniature: unguided AI let students produce the output while skipping the effort that builds the mind — exactly the Knowledge’s warning — and the fix was not to ban the tool but to design how it’s used so the effort stays with the learner.

This is where I’ll stop reporting other people’s findings and tell you what I believe.

What I believe

A personal conviction, stated as mine — not as a research finding.

The more we resist these changes, the more damage we do — to students, to the system, and to ourselves. The damage isn’t only disengagement and lost interest. It’s in the very things we say we care about: critical thinking, problem-solving, creativity. Students already use AI. They will use it tomorrow, as professionals, in a world that expects them to. If we insist on resisting, we don’t stop them — we push them toward either cheating or irrelevance, and toward a justified contempt for a system that pretended the future wasn’t happening.

I keep coming back to something simple: you cannot expect a process to produce the results you want if you are not part of the process. If we are not part of how young people learn to use technology, social media, and AI — if we stand outside it, arms crossed — we have no standing to be surprised when they use it badly. AI does not make us irrelevant. It makes us more necessary, because someone has to do the thing a model cannot: guide, show judgment, model the ethics, explore alongside them, and sometimes learn with them. We — the adults, the educators, the institutions — have the vision, the experience, and above all the responsibility to be in the process. When we abdicate that, we don’t protect learning. We hand it to whoever is willing to show up.

None of this is new, and that should comfort us. Every genuinely useful tool — writing, the calculator, the internet — arrived to the same funeral for learning, and learning went on, changed and usually better, once we stopped guarding the door and started teaching people how to walk through it. (We’ve written elsewhere about that recurring panic.) The Knowledge earned London’s drivers a magnificent map of their city. We don’t need to grieve that GPS exists. We need to decide, deliberately and subject by subject, which maps are still worth building in a human head — and then protect the effort of building them, precisely because a machine will offer to skip it.

Keep the essential. Transform the accidental. Change it — but change it well.

The factual claims in this essay are backed by the verified sources collected below; the closing section is my own stated conviction, not a research finding. — Carlos Miranda Levy

Four perspectives

Sources & further reading

This essay is a personal argument, but its factual claims are not. Every study cited below was gathered through deep research and checked against the primary source. Where the evidence is genuinely mixed — the Alzheimer’s link is an association and not causation, the far-transfer null comes from cognitive-training studies rather than school curricula, and the AI-and-learning trials are still early — the notes say so. The closing conviction is Carlos’s own, labeled as such in the text.

The Knowledge & the taxi-driver brain

• Maguire, E. A., et al. “Navigation-related structural change in the hippocampi of taxi drivers,” PNAS 97(8), 2000. doi:10.1073/pnas.070039597
• Woollett, K., & Maguire, E. A. “Acquiring ‘the Knowledge’ of London’s layout drives structural brain changes,” Current Biology 21(24), 2011. doi:10.1016/j.cub.2011.11.018 (Grey matter grew only in the trainees who passed — not in those who failed, and not in controls.)
• Maguire, E. A., Woollett, K., & Spiers, H. J. “London taxi drivers and bus drivers: a structural MRI and neuropsychological analysis,” Hippocampus 16(12), 2006. doi:10.1002/hipo.20233 (The expertise came with a measurable trade-off: worse at forming certain new spatial memories.)
• Patel, V., et al. “Alzheimer’s disease mortality among taxi and ambulance drivers,” BMJ 387:e082194, 2024. doi:10.1136/bmj-2024-082194 (An association in death-certificate data across 443 occupations — not proof that navigating prevents dementia.)
• Transport for London, Learn the Knowledge of London (official guidance) and the Introduction to the Knowledge booklet, March 2022.

Cognitive offloading & its cost

• Dahmani, L., & Bohbot, V. D. “Habitual use of GPS negatively impacts spatial memory during self-guided navigation,” Scientific Reports 10:6310, 2020. doi:10.1038/s41598-020-62877-0

Knowledge, thinking & the limits of “transferable skills”

• Willingham, D. T. “Critical thinking: why is it so hard to teach?” American Educator 31, 2007 (ERIC EJ794281).
• Sala, G., & Gobet, F. “Does far transfer exist? Negative evidence from chess, music, and working memory training,” Current Directions in Psychological Science 26(6), 2017. doi:10.1177/0963721417712760
• Sala, G., et al. “Near and far transfer in cognitive training: a second-order meta-analysis,” Collabra: Psychology 5(1):18, 2019. doi:10.1525/collabra.203 (Far transfer called “a chimera”; the effect “equaled zero” after correcting for placebo and publication bias.)
• Abrami, P. C., et al. “Strategies for teaching students to think critically: a meta-analysis,” Review of Educational Research 85(2), 2015. doi:10.3102/0034654314551063 (Critical thinking is teachable — but the approaches that work embed it in subject content, not a stand-alone skills class.)

Rebalancing the curriculum

• OECD, OECD Learning Compass 2030 (concept notes), OECD Publishing, 2019. A competency is “knowledge, skills, attitudes and values” together; knowledge and competency are “neither competing nor mutually exclusive.”
• OECD, Curriculum Overload: A Way Forward, OECD Publishing, 2020. doi:10.1787/3081ceca-en (Names our exact problem: curricula grow “mile-wide, inch-deep.” Its prescription: “learn deeper, not more.”)

Guide, don’t ban: AI in the classroom

• Bastani, H., et al. “Generative AI without guardrails can harm learning: evidence from high-school mathematics,” PNAS, 2025. doi:10.1073/pnas.2422633122 (Unguarded GPT-4: +48% on practice, −17% on exams once removed. A guardrailed, hint-giving tutor eliminated the harm.)