AI won’t ever make a historically accurate film – but it could help archaeology create better stories

Tech billionaire Elon Musk was up in arms about the recent adaptation of Homer’s Odyssey by director Christopher Nolan, which he heavily criticised for its diverse casting and accused of lacking historical accuracy. Claiming he wanted to show that artificial intelligence could bypass traditional Hollywood studios and deliver a story “historically accurate and true to the art of Homer” at a fraction of the cost, he set his AI Grok Imagine to the task.

On October 23, the platform posted the first seven minutes.

As archaeologists and art historians, we’re not here to review the film for its artistic quality or even its historic accuracy. Instead, we’re here to point out the real problem with the whole endeavour: the promise of historical accuracy failed before the first frame was even rendered.

Contrary to what many of you might assume, as historians we think Musk’s underlying idea – bringing together everything archaeology and history have to offer and using AI as a tool to explore and work with that knowledge – is a genuinely exciting one. We see considerable potential in that approach: a system spanning excavation reports and object databases might reveal patterns no single scholar can hold in mind, such as how ceramic styles spread. The problem is turning that into a film that claims to be historically accurate.

Any attempt to produce a complete, coherent visual narrative would inevitably require AI to go beyond the evidence and invent what we do not know.

Incomplete images

Archaeology is inevitably working with a fragmentary record: only a tiny proportion of the past survives in the ground, and only a fraction of that is ever excavated. For instance, open the boxes in the Troy depot and you’ll find spindle whorls and loom weights; however, you won’t find the cloth they made. Everyone at Troy wore textiles but the textiles are gone.

So, like the textiles, everything that cannot be recovered directly from the archaeological record has to be interpreted and explained through hypotheses, however well-founded they may be scientifically. There will always be gaps, uncertainties and competing interpretations.

AI, by design, would inevitably fill these gaps.

Suppose Grok Imagine was given everything the Troy Project holds: a century and a half of excavation, thousands of plans, pottery drawings and photographs, plus the small-finds catalogue. It would generate connections, details and entire scenarios where the historical and archaeological evidence offers no firm answers.

Generative AI is built to produce complete images, even when there are visible gaps. These gaps are often filled with influence from other existing media like film sets, games, paintings.

This is precisely where the problem lies: additions by AI may be plausible, even convincing. Some sceptical reviewers concede that some of those seven minutes pass for real. But convincing is not accurate, and the space between the two is where archaeology lives.

The Odyssey is also an interesting target for accuracy. It’s a poor witness to the past: shaped centuries after the palaces its about fell, it is a palimpsest of Mycenaean memory and iron age worldview – not a snapshot. Odysseus, Helen and Circe are not people archaeology could confirm. However, there is a lot of creative media about the world these characters are believed to have inhabited.

Back to those textiles. A model trained on that record would inherit our ignorance along with our knowledge. It will hallucinate or imagine because it must paint every pixel. It will draw on all the evidence, media and more connected to the themes and history to render a finished cloak, dyed and draped. This rendering will look like a convincingly accurate garment where the honest research answer is “fibre known, colour unknown”.

What survives is not what happened

None of this makes AI useless. A system spanning excavation reports and object databases might reveal patterns no single scholar could see – such as how ceramic styles spread. But every such pattern would first be a pattern in the record, not a picture of the past.

A genuinely archaeological AI would do what generative AI rarely does: show its uncertainty as clearly as its confidence. Take armour. Only fragments survive at Troy. Where evidence permits several reconstructions, it would show several rather than present one confident, definitive option. Where only fragments survive, it would show only the fragments. And where the evidence runs out, it would say so, clearly marking what is documented, what is inferred and what is merely plausible.

This approach to rendering possibilities is close to what the London Charter (2009) and the Seville Principles (2011), two international guidelines for virtual reconstructions of heritage, have long demanded: transparency rather than smoothing.

Which world gets reconstructed is also a choice. Troy spans phases centuries apart, and Mycenaean palaces differed from Hittite cities: there is no single “bronze age”. The record can support a conflict at Troy; it cannot establish Homer’s Trojan war as one event. However, an AI model picks one Troy and never mentions the choice.

The film will likely be forgotten within a news cycle or two; what remains is the method. The most accurate Odyssey would not be the most definitive or confident one; it would end where we began: on the boxes of whorls and weights, and the empty space where the cloth should be.

Historical accuracy does not mean filling every gap. It means showing where the gaps are.

The authors do not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and have disclosed no relevant affiliations beyond their academic appointment.