Subscribe now for full access and no adverts
In the last couple of years, the rise of artificial intelligence has permeated almost every aspect of society, but (outside exhibitions and other ways of presenting the past) archaeology has remained very much grounded in the physical world. New research, though, has shown how this technological revolution can help to interpret archaeological finds, with a study (recently published in Antiquity: https://doi.org/10.15184/aqy.2025.10264) using AI to determine the possible rules of a Roman board game.
We know that board games have a long history, thanks to evidence from ancient texts and works of art, as well as excavated artefacts including dice and playing pieces. Game boards themselves are scarcer in the archaeological record, however, and when they are found their actual use is often far from certain, as they are not often found with gaming pieces or other objects to illuminate their intended purpose. Similarly, how these games may have been played remains a mystery, as written sources and images featuring these objects do not come with instructions on the rules of play. Now, a multidisciplinary team of researchers from institutions across the Netherlands, Belgium, and Australia have combined use-wear analysis with AI simulations to shed light on a stone object (04433), previously interpreted as a game board, from the collections of Het Romeins Museum in Heerlen, the Netherlands (below).
The artefact in question had been excavated from the remains of Coriovallum (Roman Heerlen), a town that was occupied from the reign of the emperor Augustus (r. 27 BC-AD 14) until the fall of the Western Roman Empire (c.AD 476) in what was then the province of Germania Inferior. Measuring 212mm by 145mm by 71mm (8in by 6in by 3in), the ‘board’ is made from white Jurassic limestone that is believed to have been quarried at Norroy in north-eastern France. As this material was typically used for large architectural elements in the Roman northern provinces, this relatively small find stood out as unusual.
It is worked on all sides, and one surface has been carefully smoothed and incised with a series of lines forming symmetrical geometric shapes. With no known parallel, the use of this object has been the source of much debate – and so, to examine whether it could feasibly have been used as a gaming board, the team created a depth map of the stone using photogrammetry in combination with a photometric stereo technique. This revealed that the topography of the smoothed surface was, on the whole, more worn around the incised lines (microscopic examination confirmed this), suggesting that some sort of hard object or objects had been dragged across them.
With such marks hinting at game play, the team turned to AI simulations to see if this use-wear pattern could also determine how it had worked. To do this they used Ludii, a general game system developed as part of the Digital Ludeme Project, which has the largest database of traditional board- game rules currently available, and allows for AI-driven playout simulations as well. By entering into the system specific criteria identified from the wear patterns (such as the board having fewer than 20 likely playing sites), they were able to narrow down the possibilities to either a blocking game (where the goal is to stop an opponent from moving) or one where the aim is to place three pieces in a row. The team next carried out a series of playouts using various criteria to see which best matched the use-wear pattern, and crucially they were also able to introduce different biases to the simulations in order to reflect more accurately the way that humans actually play board games. The results suggested that there were at least nine ways in which the game could have been played, but the one that best fits all the criteria was a blocking game in which one player starts with four pieces on a vertical side of the main rectangle and attempts to trap their opponent’s two pieces on the opposite side.
As the researchers note in their paper: ‘By combining AI simulation with use-wear analysis to identify and model traces of game play, it is possible to not only identify potential game boards, but also to rebuild playable rule-sets that may provide indications regarding the ways that people played games in the past.’
Text: Kathryn Krakowka / Image: photographs courtesy of Restaura