The Grantham schoolboy and his habits
Sometime around 1655, a twelve-year-old from Woolsthorpe began lodging on Grantham High Street with an apothecary named William Clarke. For the next five years, Newton walked to The King's School — an institution whose history stretches back to 1329 — and filled the hours outside lessons with things his classmates mostly ignored: mechanical models, a working windmill powered by a mouse on a treadmill, sundials scratched into the stonework, and notebooks recording what he observed rather than what he assumed. His teacher Henry Stokes recognised something in this behaviour that Newton's mother did not. She had wanted him back on the farm at Woolsthorpe. Stokes persuaded her otherwise, and Newton went to Cambridge instead.
These habits were not the eccentricities of a solitary genius. Arthur Storer, a Grantham acquaintance from those same years, later carried out systematic comet observations from Maryland — careful enough that Newton drew on them for the Principia. Whether or not this suggests a wider local culture of rigorous observation is debatable from a single example, but it does at least show that Newton's methods were not sealed inside one exceptional mind: they were communicable, repeatable, and shared.
TEDxGrantham's 'Rethink' theme explicitly anchors Newton as evidence that systemic ideas can come from small, non-metropolitan places. That framing is fair. But it also invites a harder question: what would someone shaped by those particular habits — record everything, test it again, show your working — make of a technology that cannot show its working at all?
What 'hypotheses non fingo' actually demands
'Hypotheses non fingo' — 'I frame no hypotheses' — appeared in the General Scholium Newton added to the second edition of the Principia in 1713. The phrase is often read as intellectual caution. It is better understood as a refusal: Newton was declining to offer any explanation whose generative path he could not trace back to observable phenomena.
His prism experiments in Opticks (1704) show what that path looked like in practice. Each step was explicit: pass light through a prism, record the angle of refraction for each colour, test whether recombination restores white light. Anyone with the same equipment could retrace every move and arrive at the same result — or contradict it. The explanation was not merely plausible; it was derivable and checkable by design, open to re-running by any curious observer.
Rule IV of Book III of the Principia extends this commitment: propositions gathered by induction from phenomena should be treated as approximately true until further phenomena contradict them. The mechanism of verification is built into the standard, not bolted on afterwards as a precaution.
What Newton was insisting upon is not the modest 'I don't know.' It is something considerably more demanding: 'I will not offer an explanation unless I can show you the steps that generate it.' That distinction — between a prediction that happens to be correct and an explanation whose mechanism anyone can independently retrace — is precisely the standard that becomes uncomfortable the moment we apply it to modern AI.
Why neural networks cannot show their working
The opacity of large neural networks is not a matter of complexity that might yield to closer inspection. A deep learning model arrives at its outputs by optimising across millions or billions of parameters simultaneously — a process that produces no readable chain of reasoning, no step that corresponds to a decision, no path an observer can retrace. Even the engineers who built and trained the system cannot reconstruct why it returned a specific result for a specific input. A model can, in this sense, be highly accurate and entirely opaque at the same time: the two properties do not constrain each other.
Researchers have developed tools — most notably LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) — that attempt to work around this. Both operate post-hoc: they build a separate, simplified model that approximates the original system's surface behaviour for a given output. The approximation can be useful. But it is not the mechanism. What LIME or SHAP describes is the logic of its own proxy model, not the internal workings of the system it was built to illuminate — an approximate explanation rather than a derivable one.
Regulatory bodies have begun to treat this gap as a practical problem rather than a philosophical one. The EU AI Act, and equivalent frameworks emerging globally, increasingly requires that AI systems making consequential decisions — in medicine, credit, or employment — provide explanations to affected parties. The demand exists in law. Whether current tools can genuinely meet it, rather than approximate it, remains an open question.
Where Newton's four-step loop breaks
Newton's loop, compressed to its essentials: observe a phenomenon, design an experiment that isolates the relevant mechanism, induce a general proposition from the results, then verify it against fresh observation. The power of the loop lies not in any single step but in its continuity — each stage feeds the next, and any observer can enter at any point to check the reasoning.
Apply it to a clinical AI that flags a patient record as high-risk for readmission. Step one holds: the output is observable. But step two — design a repeatable experiment that isolates the mechanism — fails immediately. There is no equivalent of the prism. You can vary inputs systematically and record how outputs shift, but that is probing the surface, not interrogating the interior. The generative path, spread across billions of parameters, produces no isolable step that corresponds to a decision. Induction from phenomena (step three) requires that the phenomena be the mechanism's phenomena, not merely correlated outputs; and verification (step four) can confirm a model's track record without confirming its reasoning.
Newton would note the distinction at once. A prediction that is reliably correct is not the same thing as an explanation whose steps anyone can retrace. 'Hypotheses non fingo,' read in this light, is not a methodological preference — it is a demand that the generative path itself be available for scrutiny. XAI research is candid about this gap: post-hoc proxies such as LIME and SHAP are genuinely useful tools, but they describe a separate simplified model's logic, not the original system's mechanism. The break in the loop is structural, not a problem of insufficient computing power or incomplete data.
What Newtonian scepticism looks like in practice
A Newtonian user of AI does not refuse the tool — Newton did not decline the telescope while optics remained incomplete. The practical stance is more disciplined: treat every output as provisional until its reasoning can be independently checked, and be explicit about which outputs you are willing to act on without that check.
The useful distinction is between average accuracy and specific verifiability. A spam filter catching 98 per cent of junk mail can reasonably run on track record alone — the cost of an occasional false positive is low and recoverable. But where a decision is consequential and hard to reverse — a clinical risk flag, a loan refusal, a legal summary — track record is not sufficient. Newton's fourth Rule of Reasoning holds that empirical propositions stand until contradicted by better evidence; that requires the proposition to be statable in a form evidence can actually test. An output without a traceable reasoning chain cannot be tested in that sense — it can only be repeated.
This domain distinction is increasingly codified. The EU AI Act applies its most demanding requirements to high-risk applications precisely because those are the cases where the Newtonian question — not just 'did it work?' but 'can I verify why?' — carries real weight. In practice, a Newtonian stance means recording where AI tools fail as carefully as where they succeed; resisting the temptation to generalise from one accurate result to general reliability; and asking, before acting on any consequential output, whether the explanation on offer is a mechanism or merely a proxy.
What Grantham's intellectual habit offers now
The insight that opaque systems require systematic scepticism did not emerge from a metropolitan research campus. Its Grantham thread runs through a schoolboy who built models, kept observational notebooks, and refused to record conclusions he had not derived — decades before he formalised that discipline into published method. That habit is the practical inheritance this town has in Newton, and it applies directly to the decisions arriving in local institutions now.
When AI tools enter a GP surgery's triage workflow, a secondary school's assessment platform, or a council's housing allocation system, the Newtonian question is not 'does it work on average?' but 'can the specific output in front of me be traced back to something verifiable?' Average accuracy and individual accountability are not the same thing. A resident refused a service, a student flagged by an algorithm, a patient triaged by a risk model — each is a specific case, not a statistical summary.
Newton's legacy here is method, not monument. The 'Rethink' impulse is not nostalgia for a famous alumnus; it is the recognition that first-principles scepticism — asking what derivable evidence lies behind a claim before accepting it — is the right response to any system that cannot show its working.
- [1] Early life of Isaac Newton. https://en.wikipedia.org/?curid=315685 https://en.wikipedia.org/?curid=315685
- [2] Isaac Newton. https://en.wikipedia.org/?curid=14627 https://en.wikipedia.org/?curid=14627
- [3] The King's School, Grantham. https://en.wikipedia.org/?curid=7081013 https://en.wikipedia.org/?curid=7081013
- [4] Arthur Storer. https://en.wikipedia.org/?curid=2256249 https://en.wikipedia.org/?curid=2256249
- [5] Hypotheses non fingo. https://en.wikipedia.org/?curid=4795569 https://en.wikipedia.org/?curid=4795569
- [6] Explainable artificial intelligence. https://en.wikipedia.org/?curid=54575571 https://en.wikipedia.org/?curid=54575571
- [7] Opticks. https://en.wikipedia.org/?curid=618076 https://en.wikipedia.org/?curid=618076
- [8] Machine learning. https://en.wikipedia.org/?curid=233488 https://en.wikipedia.org/?curid=233488
- [9] Regulation of artificial intelligence. https://en.wikipedia.org/?curid=63451675 https://en.wikipedia.org/?curid=63451675