Fabricating the past: can AI Really help reveal hidden histories in the classroom?

"Ladybird - Adventures from History" by ulle.b is licensed under CC BY-NC-SA 2.0 (via Flickr).

So recently I was introduced to the Any Human Ever web project. The project begins from the premise that the stories we tell in history often do a poor job of representing the fullness of the human the past. As the authors state:

"We know far more about the sequence of Frankish kings than we do about a typical peasant's routine in 5th century Avignon. Think of your notion of history, and contrast it to the fact that most lives ever lived were Asian, poor, and brief."

I agree. We know that much of what we encounter through history is just a tiny piece of the bigger tapestry of human experience. Any Human Ever therefore proposes a radical alternative - to create human stories based on data. A statistical way to get a better sense of what the human experience was like. Again, they explain their approach: 

"If you could draw a life at random from all of human history, what would it look like? Roll the dice, and see a genuinely random life, depicted as well as the data allow. Do it again and again, and what kind of pointillist painting emerges from all those colorful drops of humanity?... That's the idea of anyhumanever. Over 100 billion humans, each with a story."

But despite the inclusive goals of the project, what I ended up with was a feeling of deep concern about the ways in which we now marry big data and AI to fabricate the past. Far from creating a better picture of human experience, I am more an more convinced that projects like this actually obscure the past and contribute to the undermining of what it means to think historically about our world.

I feel this is particularly pertinent for history teaching at the moment, as the movement to place historical storytelling front and centre continues to gather pace. I have already seen it suggested that teachers might turn to AI to help them build interesting 'character cards' or create stories to introduce pupils to historical periods. Stories, and especially human scale ones, do indeed draw us into history. But just as we might criticise a 'Ladybird' rendering of Richard I or Nelson, so too we should critique stories which are pulled from vague period knowledge, or AI hallucinations, and which are designed for memorability, or emotional connection rather than respecting the limits of what we can know. The key question for any history teacher is not so much 'what is the story?' but 'how did we get this story and is it the only one?' A key question therefore is what we gain and what we risk in the process of bringing AI into classrooms to try to fill in the missing people of the past through human-scale narratives.

First Impressions
I will begin by saying that I think that using data to try to reveal things about hidden lives is an incredibly valuable historical approach. AHE lets you specify a year and a place and will use its sources to show you the statistical picture for humans at that point in history. For instance, I could choose the year 1050 and quickly get a statistical picture of life expectancy, key causes of death, common work patterns etc. for a whole range of places. This has some interesting potential. For instance, teachers of Medicine Through Time courses might find it helpful to 'roll the statistical dice' and see and compare the varying life chances of people in England from the middle ages to the present.

​The sources themselves for this data are cited (if limited), which is a positive. But even here, I found I had concerns. This may feel like a minor point, but it is never really made clear how the very specific percentages have been derived. The linked sources don't illuminate too much either, not least because some are behind paywalls which no school teacher would be able to afford to remove.

From Data to Fabrication
​Beyond the initial statistical overviews, users are invited to generate individual lives based on statistical probabilities. ​First users can generate a timeline their invented person's life. This is the first major step away from the source material and into fabrication. A skeleton life story emerges which draws on the statistics to work out potential life events. ​I assume this is AI, but it could also be a simple statistical model or probability tree. 

Once the timeline is created, users are further invited to ask an AI model to fabricate a life story based on the timeline, relevant statistics and a handful of source materials. What is produced is a 'person' with a name and short life story.  Through this fabrication process we are encouraged to connect with them and to see their lives as plausible historical realities. The user is being invited to confuse a statistical past with the lived past.  

The point of moving away from the general statistics and into the realms of supposition is exactly when a good historian would signal the need for extreme caution. We don't get that here. In fact, the AI model becomes less cautious at each stage. Where the initial statistics noted possible outcomes, the AI story paints a picture in bold colours, complete with an AI generated portrait. 

Of course, historians do tell stories and they do rely on data and statistics to explore hidden histories. As Trouillot notes, the past is full of silences. Historians spend whole careers trying to reconstruct missing lives, especially those who have been actively forgotten. I am reminded here of the amazing work of Edward Baptist who used many statistical records and other tangential sources to try to gain glimpses into the lives of enslaved Americans in the 19th century. Lives which would otherwise be largely  unknown. But that is painstaking work which demands a deep respect for people, their dignity as human beings and defensible historical methods. Crucially, Baptist's work has been critiqued through disciplinary debate - there are ongoing discussions about how this work is done. But if this work is done well, and carefully, then the stories of real people like Charles Ball, a central figure in Baptist's book, may well begin to emerge, if only faintly, and provisionally, from the records. 

AI by contrast, despite the disclaimer at the top of the page, shows none of this caution. It uses its data set to breathe false life into imagined people. Roger never saw Halley's Comet. He didn't breakfast on oat bread or chat with the traders in the town. He didn't leave any descendants, not because he died of dysentery, but because he never lived.

Deafening Silences
And here is the rub. The Any Human Ever project sets out with a laudible aim - to restore the presence of "the unrecorded 99% of past humanity". But like every project where the hubris outstrips the reality, for every silence the project challenges by making big data available, it layers on still more silences in what gets missed out of these fabricated stories.

Let's take Meena, 'born', according to AI, near Amritsar in the 1830s. 'Meena's story, tells us that she sees Halley's Comet streak across the heavens, as that she witnesses the skies turned red by the eruption of Cosiguina. Yet strangely she does not witness the violence of the first Indian War of Independence, and the  summary execution of those who took part. Her India is not fighting against repression and for independence because, for whatever reason, the AI model chooses to ignore these things. 

The AI generator boldly claims that its generated life stories are 'created from the statistical reality' of the chosen place and time. Yet not a single AI person from the twenty I generated on the Guinea coast in the 1750s had so much as heard of the impacts of the Transatlantic Slave Trade, let alone felt them themselves. According to the sources used to generate the stories, Atlantic slavery was apparently not a statistical reality for people in West Africa in the 1750s!

By contrast, enslaved people do appear in the American South. But their life stories don't tell of the impact of enslavement. People are not beaten, punished or abused. Nor is enslavement linked to the premature death of so many enslaved Americans. What historian would look at Isaac's story (below), if he were real, and not ask why Isaac's father died so young, what happened to his wife, or who ensured his travels were so limited? Enslaved AI African Americans in the 1830s did not, according to the story generator, have to cling to their history and humanity in the face of social death - enslavement was worn so lightly by these fictional people as to be almost non-existent. Their experiences are devoid of reality. ​I do not think this is the intention of the project, but that view is not to dissimilar to the re-writing of the history of slavery by the Far Right in the USA now.

And the more we look at these fake lives, the more the questions start to emerge. Isaac is noted as being literate but I wonder if the statistical model considers the statistically likelihood based on Isaac's enslaved status. And if, like Susie King Taylor, Isaac did manage to learn to read despite the prevailing conditions, the question of at what cost is neither pursued nor even hinted at. The true experiences of millions of enslaved peoples are summarily extinguished beneath a glib AI fantasy. I wonder too what the AI story-weaver would say of white plantation owners, or Indigenous Americans being forced from their homes to make way for cotton, or being pulled into enslavement themselves. I suspect very little. 

These are not isolated examples either. Ask the project to generate a life on the  California coast in the mid 1800s, a period during which the California government recently acknowledged a genocide was perpetrated against Indigenous peoples, and what do we find? Again, no sign of violence, forced labour, military invasion, or resistance to colonisation. Certainly no genocide. What we get instead is a series of generic Indigenous people, isolated from the realities of the period and of history.

And if we choose to go back further still, we find that Land Bridge theory erases the presence of Indigenous peoples in North America altogether. Go back 20,000 years and, despite solid evidence showing Indigenous peoples in North America at this time,  you don't get the option to generate a person at all. Instead, North America is shown as an empty land. 

An absence of sources
Despite everything AHE purports to do, the weakest link is actually the evidence base. When we dig into the nature and utilization of the source materials, we see some very interesting things. The only plus point here is that the model is relatively transparent about this, even if the links are hard to verify. As shown below, only some of the cited sources are directly use in producing the 'life' stories. Those which are used tend to be  statistical. We see far fewer studies which explore  the deep structures of life in the past: gender norms, kinship, education, emotions etc. There are gaping holes where foundational works of scholarship should be. And when they do appear, they are paraphrased by AI in unexplained ways. 

Does it really matter?
The big question I suppose is whether any of this really matters. The project sets out with a positive aim and is attempting to fill gaps in history and provide some entertainment at the same time. ​Am I taking this too seriously?

Maybe. But we are increasingly seeing AI used to make claims about all sorts of things. And the tendency is for people to use AI to fill gaps where our knowledge is weak. This is precisely when it is at its most dangerous, because we lack the knowledge base to challenge its claims. Fabrication therefore is quickly able to become part of the fabric of history. This is not a new phenomenon. This kind of fabrication has always happened with national heroes and stories. The difference is that the access to it is so easy and the nature of the authorship lends a false sense of rigour and neutrality. 

Constructing the past is fundamental to history. But if we are going to start to weave fabricated lives into our classrooms, whether created by AI, or as stories written by teachers, we need to be absolutely clear, both with ourselves and our pupils, about how, and crucially why, we have decided to engage in this process. 

Most importantly, whatever stories we use in a classroom, and at whatever scale, from the personal to the local, regional, national, or even international, we need to open them up to critique. We need to be reminding students always that this is 'a' story not 'the' story and encouraging them to constantly ask 'how do we know?', 'who is telling us this?', 'why are they telling us this?' and so on. In fact, the main use I can see for any of the stories created here is to help pupils to grapple with what they fail to do and how they fundamentally undermine human experiences of the past. A few lessons correcting and improving some of these caricatures  would be time worth spending I think.

Good historical construction relies on deep respect for people and the traces they leave. If we cannot, in good conscience, say that the fabrications we use pass this test, then we should be thinking hard about whether we use them at all. If we fail to grapple with the growing presence of AI (and other) fabrication in the classroom we compound the issues of erasing the people already forgotten by history, and risk fundamentally undermining the humanising process of studying and attempting to reconstruct the past.