Reading Newton With AI: A Clue Is Just the Start

Key takeaways

• Accuracy and scholarly novelty require separate checks.
• Similar wording in two documents does not establish who influenced whom.
• Rediscovering an established connection can help readers, but earlier researchers deserve credit.
• Research budgets need to cover source preparation and expert verification alongside model costs.

Imagine an AI reading Isaac Newton’s alchemical notes and flagging a connection to another historical document. Before that becomes a headline about a machine uncovering a secret historians missed, someone has to establish what it actually found. As AI becomes a research tool, that distinction matters more than how impressive the answer sounds.

Correct and new are two different tests

A large language model can explain unfamiliar language or suggest texts worth comparing. Those capabilities offer useful starting points for historical research. They do not settle the interpretation.

Any promising answer faces two separate questions:

• Does the passage actually appear in the original document, and does the interpretation fit its context?
• Has an earlier researcher already made the same observation?

The first tests accuracy. The second tests novelty. A model can identify a passage correctly while repeating established scholarship. It can also produce an apparently original interpretation by misreading the source.

The discovery claim itself needs to be specific. Finding an unknown manuscript is different from proposing a new relationship between familiar texts. Each requires its own evidence.

A polished explanation cannot do that work for us.

Matching phrases do not establish influence

Take the hypothetical a step further. Suppose the model finds similar wording in Newton’s alchemical notes and material associated with Hartlib.

That gives us a textual resemblance to investigate. It does not yet establish that either writer learned something from the other.

Start with chronology: when was each document written? Then check whether both passages could draw on an older source. If the phrase was common terminology at the time, the overlap might tell us very little about a particular connection.

A claim of direct influence also needs a plausible route of access. Correspondence, records of ownership, or citations could help establish whether a writer encountered the relevant text.

The model’s suggestion is therefore a hypothesis to investigate. “Compare these passages” is a useful research prompt. “These people exchanged knowledge” is a historical claim with a much heavier burden of proof.

The distance between those sentences is where much of the research happens.

Rediscovery has value—and someone may deserve the credit

Suppose a literature review turns up the same interpretation in an existing paper. The claim to a new scholarly discovery falls away. The practical value may survive.

A newcomer might have found an important text through that exchange. A researcher might have reconnected material scattered across different collections. Making established knowledge easier to find and understand is useful work.

It should also be described accurately: AI helped recover a connection already discussed in the literature. Cite the earlier scholarship and credit its authors.

The model may have encountered that scholarship during training. But its answer alone cannot establish what its training material contained. To judge originality, examine the primary sources and prior research. Confidence of delivery is no substitute.

Efficiency needs similar care. A response delivered in seconds may still require substantial time to check citations and repair errors. Any claim of time saved has to include that work.

Fund the people who make the result checkable

For research funding, I would put more weight on an inspectable result than a dazzling answer.

That means recording the document and passage behind a claim, explaining the interpretation, and marking what has been checked. Linking manuscript images to transcriptions—the written-out text of handwritten material—makes later review easier. Keeping expert corrections alongside the model’s proposed reading helps preserve the reasoning.

All of this costs more than model access. Someone must organize the material, decipher old handwriting, interpret historical terminology, and compare the result with existing scholarship.

Leave that labor out of the budget and a project can accumulate promising leads faster than anyone can assess them.

Humanities funding should recognize verifiable research materials as an outcome alongside new discoveries. A well-documented collection can save the next researcher from repeating the same groundwork.

AI can give a historian somewhere useful to look. Calling the result a discovery requires evidence from the documents and a serious encounter with the scholarship that came before. The interesting question is what we can now establish that we could not establish before.