Southampton Historians Used AI to Read 100,000 Pages of Handwritten Port Books

Four centuries of Britain's shipping were locked in leather-bound books that almost nobody can read quickly. The University of Southampton says it has now opened them up: historians trained AI to read about 100,000 scanned pages of handwritten port books and used the output to build a first-of-its-kind interactive map of Britain's merchant shipping from 1565 onward. The announcement is on the University of Southampton news page, and the BBC covered it on October 20.

It is a good example of AI doing something concrete and checkable in the humanities: turning an unreadable pile into structured data a historian can query. This post explains the method as described, what the researchers found, and what the release does not say.

TL;DR

What a port book is, and why it is hard

A port book is a customs record: officials at English ports noted ships arriving and leaving, who commanded them, what they carried and where from. They are a goldmine for economic historians, but they are written by many hands in secretary and mixed scripts, with abbreviations, with inconsistent spellings that drift over two centuries, and in books whose parchment is delicate. Reading them by eye is slow, and nobody could previously analyze the whole series at once.

Lead researcher Professor Craig Lambert said the project was "a massive task, completed on a scale not seen before," and that without AI extracting and organising the data, "this would have taken years. It still needed manual checking, but it sped up the process enormously."

How the AI was used

According to the release, the team scanned the pages and used specialist software to recognise the text, learn, and compile information about who was trading what, with whom, when and where. Lambert's summary of the training: "We taught AI to read the unique handwriting belonging to many different port officials and also the change in spelling of words across the decades."

Two points are worth noting for anyone building similar pipelines.

1. Training is per hand and per period. Handwriting recognition models do best when they see many examples of the specific writers and orthography in the collection. Teaching the model both handwriting variation and spelling drift is the bulk of the work.
2. Extraction is more than transcription. The goal was structured data: ship, port, cargo, date. That means recognized text must be parsed into fields and reconciled across variant spellings, then loaded into a database the GeoData Institute turned into a map.

The release and BBC piece do not name the model, the software or the character error rate. Treat statements about "AI reading" as describing a pipeline with human review, not a solved problem.

What the data shows

Dr John McAleer, the project's co-lead, said it is interesting that even centuries ago Europe was Britain's principal overseas trading region, and that although Britain had important global connections, most core trade was done closer to its own shores. That revises a popular image of imperial trade centered on far-off colonies.

Why it matters for AI practitioners

This is a case where the AI's output can be audited: the images exist, the database entries can be compared with pages, and historians can spot-check. That makes it a good template for applied document AI.

• Scale unlocks new questions. Individual scholars have studied a few books for decades. A complete series allows network and trend analysis across the whole period.
• Human-in-the-loop remains. The researchers did manual checking. If you run a similar project, budget for review and publish an error estimate so others can use the data safely.
• Public access is part of the product. The map is designed for the public, scholars and citizen historians, which broadens reuse.
• Related tools are improving. Open-weight OCR models are getting better quickly; see our coverage of LightOnOCR-3 and ParseBench, Indra OCR on a CPU and Interfaze 1 Lite. They target printed and mixed documents; early modern handwriting remains a harder, specialist case.

Questions the database can now answer

A complete, structured series changes what a historian can ask. Instead of reading a few books and generalizing, researchers can query the whole record. Examples that follow directly from the fields the release describes (ships, ports, routes, goods, shipmasters):

• Trade networks over time. Which ports traded with which, and how did those links shift between the reign of Elizabeth I and George III?
• Commodity histories. When did a cargo such as coal rise, and which ports drove it? With more than 101,000 coal entries, the series is large enough for trend analysis.
• Shipmaster careers. Following named shipmasters across years and ports, which is nearly impossible by hand across thousands of pages.
• Lost places. Cross-referencing ports that vanished, such as Ravenspurn and Dunwich, with erosion and changes in coastal trade.
• Seasonality and shocks. Whether wars, harvests or regulations show up as dips and spikes in voyages.

Each of these is an analysis a researcher would have designed for a subset, then hoped generalizes. With the full dataset, claims can be tested instead of assumed, which is the scientific value of the digitization even before any new conclusion emerges.

What this says about AI in everyday culture

Many AI stories are about chatbots. This one is about a model doing a narrow job better than any alternative on the same budget: reading handwriting at scale, with humans checking the output. It also shows where the technology should be trusted less. A transcription model can confidently misread a numeral or a place name, and a single wrong digit in a cargo quantity can propagate into a statistic. The responsible pattern, which the researchers describe, is automation for volume and humans for verification.

Caveats and open questions

• No accuracy figure. The release does not state a transcription or extraction error rate, so the database's reliability is unquantified in the material we read.
• Selection effects. Surviving port books and what officials chose to record shape the data; a map of recorded trade is not a map of all trade, and smuggling is invisible.
• Value comparisons depend on standardization. Statements such as "80% of recorded value" rely on how historians valued goods and normalized units.
• We have not used the map. This post draws on the university's release and the BBC report.

How to run a similar project

1. Digitize with consistent quality. Image quality sets a ceiling on recognition accuracy.
2. Create ground truth. Have experts transcribe a sample from each hand and period; this is your training and test set.
3. Train or fine-tune a handwriting model and measure character and word error rates on held-out pages.
4. Normalize entities. Map variant spellings of ports, ships and goods to canonical IDs, with human review of ambiguous matches.
5. Publish the uncertainty. Provide confidence flags or an error estimate with the dataset.
6. Build an interface. An interactive map or search lets non-specialists test claims against the data.

For other examples of AI applied to large archives, see our look at the full arXiv archive as a Hugging Face dataset and at DeepMind's survey of AI in science.

Related reading

• LightOnOCR-3 and ParseBench open-weight OCR
• Indra OCR: 1M-parameter CPU model
• Interfaze 1 Lite deterministic OCR and extraction
• Full arXiv archive on Hugging Face
• DeepMind Institute: AI in science survey
• Google Research: AI ultrasound for novice operators
• Sources: University of Southampton release · BBC report

Figures are from the University of Southampton release of October 19, 2026 and the BBC report; no transcription accuracy rate was published.