Journals of a Time Traveller

I encountered AI (Artificial Intelligence) image generators only five months ago. Immediately, I began using them to restore and colourise scans of old black-and-white family photos. Next, I experimented with altering the ages of my ancestors, visualising them in youth or middle age. By pulling documentary evidence like height, build, hair, and eye colour from prison or military records, I could re-age them and place them in the correct uniform. I could place them in a wide variety of settings. For instance, using the military records of my mother's paternal grandfather, Alfred Henry Curtis, I could place him directly as a young man in South Africa:

I soon developed this concept further. I realised I could use AI image generators to recreate ancestors from nothing more than prison and military descriptions, combining these records with historical social conditions, local phenotypes, and a plausible likeness to their close descendants. Alternatively, I could go a step further: reconstructing them with no physical descriptions at all, but dressed in the authentic clothing of their status and time. Crucially, I could instruct the AI to prioritise raw realism over any tendency to glamourise the past.

This methodology eventually launched me into a much deeper exploration: my series on deep time, Time travel, and haplogroup ancestry. In this series, I follow an ancient story, tracing variants within mitochondrial DNA and the Y chromosome. To bring this journey to life, I use AI-generated images to illustrate plausible ancestors as they moved through different archaeological cultures.

Consequently, I had moved from simple photograph restoration and into the realm of time-travel photography. I found that I could use free, publicly accessible AI image generators to reconstruct entire landscapes:

It seems that I had stepped right into a burgeoning fad for creating AI images and videos that appear to portray modern individuals travelling into past ages. Time travel, it seems, is catching on. To illustrate, I just commissioned Google Gemini AI to prompt its image generator, for an image of myself, in 120 CE, on Hadrian's Wall in Northern Britain. I'm dressed in the segmented lorica segmentata (iron plate armour) and a heavy sagum wool cloak pinned with a fibula. I hold a pilum (javelin). Where is my army pension?

This inevitably raises the question of truth. We are told we can no longer believe what we see in images or videos, sparking a general panic about AI-generated fabrications and 'fake news'. I could point out, of course, that all images are an illusion—that nothing is quite as it appears, even to the naked eye. But philosophy aside, as a traditional film photographer, I am well aware of how easily one can manipulate even chemical silver salts to distort reality. Yet, it was never possible on such a scale, or with such casual ease.

As for how we view the past, our vision has always been coloured by prejudice. We inevitably view distant eras through the spectacles of our own culture, background, and ethnic identity. That is nothing new. The way the Victorians envisioned the 'Ancient Britons', for instance, was radically different from how the Tudors saw them, or how the twenty-first century understands the British Later Iron Age.

While I shall resist the temptation to dive deeper into the philosophy of truth, I must confront how these biases manifest today. As I continue to experiment, I keep encountering a fascinating reality: AI image generators have prejudices of their own. What follows is a breakdown of why this happens, how I spot it, and the specific idiosyncrasies I have recently noticed regarding AI visual time-travel.

The Flaws of AI Visual Reconstructions of the Past

AI image generators cannot understand the Bronze Age.

Seriously—go and ask one to generate a high-quality scene of a Bronze or Copper Age settlement. Because bronze is cast, rather than forged, and is a much softer metal than iron, its practical use translates into radically different engineering and casting shapes for weapons and tools compared to their equivalent iron counterparts.

I eventually had to give up. AI simply cannot understand bronze. Furthermore, it lazily projects an iron-forged reality even further back into the Stone Ages. Look closely at the spears it generates for pre-metal eras, or its complete inability to render early Neolithic round-bottomed pottery. I have even seen it generate Mesolithic microliths—delicate, tiny stone inserts—rendered the size of modern kitchen knives.

Cheddar Man and the Western Hunter-Gatherers (WHG).

Cheddar Man was an individual who lived in Cheddar Gorge, Britain, near the close of the Younger Dryas. While his remains were discovered over a century ago, it was the relatively recent sequencing of his ancient DNA that rewrote our visual understanding of him. Genetic analysis revealed alleles indicative of a dark, or very dark, skin tone, remarkably combined with light-coloured, blue eyes. Grafting these specific genetic markers onto a facial reconstruction based on his skull topology produced a striking, unique-looking individual.

But was this phenotype unique to him? It turns out it was not. Other individuals who lived across Europe between 14,000 and 5,000 years ago shared these identical alleles. Together, they form a distinct population that human population geneticists have termed the WHG (Western Hunter-Gatherers). They shared dark skin, light-coloured blue eyes, and were universally lactose intolerant.

The WHG looked like no 21st-century ethnicity. Yet, because AI image generators merely act as a mirror to their modern creators and users, they cannot easily conceptualise a people like Cheddar Man. If you ask an AI for dark skin, it automatically grafts on facial architecture and hair textures associated with modern-day people of African heritage—traits the WHG simply did not possess. To satisfy the prompt for blue eyes, it then inserts unnatural, alien, laser-like startling blue irises.

This happens because AI image generators inherit the 21st-century prejudices and commercial classifications of their developers. They are hardwired to create known, modern-day ethnicity, all while adhering to contemporary, hyper-polished standards of beauty and perfection.

Vikingisation and Ragnar Lothbrok.

But this 'Vikingisation' goes far beyond the eighth to twelfth centuries CE; it is routinely carried over into entirely unrelated historical periods. In fact, almost any archaeological age can fall victim to it. Ask an AI for a historical or prehistorical scene from the medieval era or earlier, and there you will find Ragnar waiting for you.

This bias isn't limited to battle scenes, either. I recently asked for a Chalcolithic (Copper Age) scene on a European river, set thousands of years before the first Viking ever sailed. The vessel it generated? A clinker-constructed longboat. Please! AI image generators will lazily default to dragon-headed longboats rigged with square sails—even when the period in question predates the very invention of the sail in that region.

Landscapes are subject to profound change across deep time, and these environmental shifts must be meticulously considered before we even begin prompting an image.