Where It All Began: Brightwater Pilot

Our Pastronauts identified 2,361 archaeological features that had never been formally recorded, which was a 60% increase on the 3,925 sites already in the Historic Environment Record. But it wasn’t just the volume that stood out – the citizen-generated data was both more in-depth and accurate than the existing HER data.

The AI component was more mixed, as the model frequently confused modern roads with field boundaries, and roundabouts with round barrows. However, these early mistakes were useful, teaching us a great deal about what AI can learn from data created by citizen scientists, and shaped how we’ve approached the human-machine relationship in every mission since.

96% of participants reported improved knowledge of how archaeology informs planning decisions, and 75% developed stronger bonds with the landscape they mapped.

Data created by participants achieved an average quality rating of 94.32%, ahead of the existing Historic Environment Record’s 88.71%. Of the 3,670 features identified, 1,309 matched sites already on record, while the remaining 2,361 were entirely new discoveries.

The validated dataset was shared with Durham County Council’s Historic Environment Record, filling gaps in the official spatial database that planning and landscape management decisions across the region depend on.

The citizen-generated data also became the training dataset for an AI, as the first attempt at combining crowd and machine intelligence in this way, and the foundation for how we’ve thought about that relationship in every mission since then.

Not brilliantly – yet. The model frequently mistook modern infrastructure for archaeological features. But these pilot results were never expected to be perfect. They provided essential learning about what an AI model can learn from data created by citizen scientists, informing how we’ve designed the human-AI relationship in every mission since.