AI and archaeology: why a historically accurate fiction remains impossible

The trigger: an AI film on the Odyssey

The starting point of the analysis is a recent controversy. Entrepreneur Elon Musk criticised director Christopher Nolan's adaptation of Homer's Odyssey, faulting its diverse casting and a lack of historical accuracy. He then claimed he wanted to show that artificial intelligence could produce a work faithful to Homer at a low cost.

His tool, Grok Imagine, published the first seven minutes of such a film on 23 October. According to the article, reprinted from The Conversation, the issue is less the artistic quality of the result than the very promise of historical accuracy, which, according to its authors, fails before the first shot.

A fragmentary archaeological record

Archaeology works from a fragmentary trace of the past: only a tiny part of the remains survives in the ground, and only a fraction of that is excavated. The authors take the example of Troy: the deposits contain spindle whorls and loom weights, but no fabric. Textiles have disappeared, although all its inhabitants wore them.

Anything that cannot be recovered directly must therefore be interpreted through hypotheses, more or less solid, with uncertainties and competing interpretations. A generative AI, designed to produce a complete image, would fill these gaps with elements borrowed from other media, such as cinema, video games or painting.

Plausible is not accurate

The central risk lies in plausibility. Additions produced by an AI can seem convincing, even authentic to some viewers. Yet, as the authors stress, "convincing" is not synonymous with "accurate", and archaeology works in the gap between the two.

Homer's poem is itself a fragile witness: composed centuries after the collapse of the palaces it evokes, it blends Mycenaean memories with the worldview of the Iron Age. Odysseus, Helen or Circe are not characters that archaeology can confirm. A model trained on these traces would inherit our gaps as much as our knowledge.

AI is useful if it shows its uncertainties

The researchers do not reject AI outright. A system cross-referencing excavation reports and object databases could reveal patterns that no specialist would perceive alone, for example the spread of ceramic styles. But this would be a pattern present in the data, not an image of the past.

In their view, a truly archaeological AI should display its uncertainty as clearly as its confidence: present several reconstructions when the evidence allows, limit itself to fragments when only those remain, and flag what is documented, inferred or merely plausible. This approach echoes the principles of the London Charter (2009) and the Seville Principles (2011) on the virtual restitution of heritage.

Choosing which world to reconstruct

Reconstructing the past also means making choices. Troy covers phases separated by several centuries, and Mycenaean palaces differed from Hittite cities: there is no single "Bronze Age". The traces may support the existence of a conflict at Troy, but they cannot establish Homer's war as a single event.

Yet an AI model chooses one version of Troy without ever flagging that choice. The film will probably be forgotten within days, but the method matters. For the authors, the most faithful reconstruction is not the most assertive one, but the one that shows where the gaps lie.