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

Tech magnate Elon Musk expressed dismay over the recent film adaptation of Homer's "Odyssey" by director Christopher Nolan, criticizing its diverse casting and lack of historical accuracy. To demonstrate AI's potential to produce a "historically accurate and true to the art of Homer" story at a lower cost, Musk utilized his AI Grok Imagine to create the film's opening seven minutes.

As archaeologists and art historians, our focus is on highlighting the real issue with this project: the notion of historical accuracy before the film's first frame was even generated.

While Musk's idea of combining archaeological knowledge with AI to explore and work with that information is promising, creating a historically accurate film proves challenging. Archaeology deals with incomplete and fragmented evidence, meaning any attempt to create a complete, coherent visual narrative would require AI to invent what it doesn't know. Archaeological records are limited; only a small fraction survives, and even that fraction is seldom fully excavated.

For example, the findings at the Troy site include artifacts like spindle whorls and loom weights, but no remnants of the actual textiles people wore. Thus, much of what people wore remains unknown. AI, designed to generate complete images, would fill these gaps with plausible details often influenced by other media such as films, games, and paintings.

This approach raises concerns: while the AI's output may appear plausible, it is not necessarily accurate, and the distinction between plausible and accurate is crucial in archaeology.

Homer's "Odyssey" is also a poor source for historical accuracy due to its creation centuries after the events it describes. Odysseus, Helen, and Circe are not verifiable figures, but the world they inhabit is rich in creative media. A model trained on this record would inherit both our knowledge and our ignorance, potentially producing hallucinations or imagined elements. These AI-generated visuals would appear accurate, but the honest research answer is often "fiber known, color unknown."

While AI may uncover patterns in archaeological records that a single scholar might miss, such as the spread of ceramic styles, every pattern would be based on the existing record, not a visual depiction of the past. A truly archaeological AI would clearly indicate uncertainty alongside its confidence. For instance, armor fragments found at Troy would be presented as such, with no single definitive reconstruction shown.

Instead, multiple reconstructions would be displayed, marking what is documented, inferred, and plausible. This transparency aligns with guidelines like the London Charter (2009) and Seville Principles (2011), which emphasize transparency in virtual heritage reconstructions.

The reconstructed world selected for the film is also a choice. The archaeological record supports a conflict at Troy, but it cannot establish Homer's Trojan War as one event. An AI model would choose one Troy without mentioning the choice, leading to a potentially inaccurate portrayal. Ultimately, the most accurate "Odyssey" would not be the most confident or definitive film; it would end with the artifacts and gaps found in the archaeological record.

Historical accuracy means acknowledging and displaying the gaps, rather than attempting to fill them all.