How AI retraced the age of sail

Chris Hay did not set out to rethink maritime history; he wanted to see whether an AI system, given the right tools, could run a meaningful experiment on its own.

What followed was an accidental but instructive case study in how structured AI workflows can interrogate historical data. Hay, a Distinguished Engineer at IBM, connected modular data servers—including satellite imagery, tidal models and digitized Dutch East India Company logbooks—and let the system iterate through a question he had not planned to ask: why did sailing ships appear to get faster in the late 18^th and early 19^th centuries?

“I had no clue about this,” Hay said in an interview with IBM Think, referring to what he noticed in the sailing records. “I didn’t know there was a thing. I didn’t know there was a reason. I was just messing around with data, looking for shipwrecks.”

The project began with geography, not ships.

Hay had been building MCP servers, modular endpoints that allow AI systems to call specific tools and datasets. And one of the first real-world tests he tried was using them to analyze coastal decline in the UK near where he lives. One pulled in satellite images by coordinate. Another accessed digital elevation maps. A third provided tidal information. He initially used them to examine how the coast has changed in the United Kingdom, layering satellite imagery against terrain data and asking the model to analyze shoreline retreat over time.

“It felt really good, because I could see, over a period of time, the coastal decline of the UK over periods of months,” he said.

The broader goal was compositional. Hay believes that AI becomes more powerful when it is connected to structured tools rather than left to reason abstractly from text. Satellite imagery, terrain models, tidal predictions and even lunar position data can be chained together. The question was what kinds of patterns might emerge if the system had access to all of it.

At some point in his experiments, that logic suggested a new direction. If he could model tides and access satellite images, perhaps he could estimate where historical shipwrecks occurred.

To test that idea, he located digitized Dutch East India Company logbooks. Every day at noon, captains recorded latitude and longitude during the Dutch East India Company’s global voyages from the 1600s until the company’s collapse in 1799. Similar archives exist for the British East India Company. Hay built servers to expose those records in structured form.

The initial aim was to approximate where ships might have drifted and eventually to pair that with satellite imagery near coastlines. But once the navigation data was in place, Hay noticed another pattern.

“As soon as I put all the ship data in there, then I realized, ‘Oh my God, I got 200 years of shipwrecks and positions,’” he said.

To manage the growing toolset and datasets, Hay bundled them into what he calls the Marine Archive Server—a public MCP endpoint that exposes centuries of sailing logs and related functions in structured form.

While exploring the dataset, Hay noticed that ships sailing in the late 18^th and early 19^th centuries appeared to move faster than those recorded in the 1600s and early 1700s.

“I noticed that the ships were faster in the later time versus earlier,” he said. “And I was like, ‘Huh, when did that happen?’”

He did not assume the faster speeds meant better ships or better sailors. He treated the pattern as something to interrogate.

Rather than assume better ships caused the shift, he set out to isolate competing explanations. He separated eastbound from westbound voyages, compared different national fleets and accounted for improvements in navigation tools.

The archive server eventually exposed dozens of functions that the system could be used for. Each time the model flagged missing data or uncertainty, Hay added another tool and reran the analysis.

Eventually, the system identified two apparent inflection points in sailing speeds, one around 1790 and another in the 1820s. Later, he discovered that in subsequent change had been documented. But Hay’s interest was not in claiming a discovery. It was in observing how the AI could systematically walk through the controls.

“The AI is just a calling machine,” he said. “It’s doing reasoning. It’s calling the right tools.”

Following his experiment, Hay has made the Marine Archive Server open source and publicly accessible, allowing others to connect their own AI systems to the same historical data.

“I’ve just made that available to the world,” he said. “I’ve not hidden it away. I think if you can build tools in the right way and make them available, real researchers can go and make discoveries.”

He frames the sailing project as a case study in workflow. The model did not produce insight from thin air. It queried structured data, isolated variables, then asked for additional inputs when it reached a limit.

Hay is cautious about broad claims that AI will transform science without humans at the helm. He does not present the experiment as peer-reviewed research or as a definitive answer to why ships got faster over time. Instead, he sees it as evidence that well-designed toolchains can compress exploratory analysis.

“I think if you can build tools in the right way and make them available … the more real researchers can go and make discoveries,” he said.