the probable past

digital public history in a global context
#ias_nuq / c²dh

frederic.clavert@uni.lu

POV: you are a woman on April 29th [1945],
and you are going to vote for the first time

This a video that the information service of the French Government published on tik tok and instagram. It is obviously AI generated. the aim was to commemorate the first time French women voted, in April 1945.

It was followed by another one, that intended to celebrate the role of French women in the Resistance during the Second World War. At the end of the video, a German soldier was celebrating Liberation with French citizens, and one person in the video was holding a japanese flag.

This seond video was removed within a fewx hours, after many historians and citizens harshly criticized it on social media. Several major and les major newspapers and newsmagazines published interviews and op-ed of historians (including me) who did not only voiced criticism against the content of the video, but also against the use of AI to generate commemorative or historical artefacts.

And it is obvious that those videos are problematic in several ways:

the aestethic is bland,
the period’s distinctive features are standardized: all women have the same clothing, the same faces, the same bikes,
there is no traces of war, whereas the war itself had not come yet to an end (with still fightings in some French areas).

why?

Why a State like France would produce such videos, whereas relationship to the past is so important in this country? Whereas archivists, librarians, historians, are often civil servants?

There are reasons that are specifically French – and among other things the concurrence between several ministries and public institutions when it comes to French memory politics, as well as the current budgetary problems and government instability.

But let’s focus on something else: the core discourse of Generative AI companies is hiding the fact that their chatbots are based on language or diffusion models – they can predict the next word or the next pixel, but are not supposed to be knowledge models.

Nevertheless, the dicourse of Gen AI companies is to make us believe that those models are good to generate knowledge. That is visible in the design of chatbots: the core of their interface are built around an incentive to ask questions or text that will act as questions, with answers that, in terms of design, are framed as answers with turstworthy knowledge, trustworthy also because they are given with some ‘belle langue’, well formulated language.

There is this idea, in this discourse, that you can access knowledge thanks to a chatbot, without using some sort of traditional mediations to knowledge and, hence, here to collective memory, so without archivists, librarians, hisorians (wether professional or not) and public historians, and even without more recent types of mediation such as Wikipedia, for instance (though models were trained on the work of all those).

This promise, within a French context in the case discussed here, is making a point: you don’t need to ask those boring gate keepers such as historians to reach out easily (or so is the intent) to younger people for instance. This is an automation of intellectual work.

how?

But when we look at the ‘how’, we can realize that chatbots are replacing traditional memory mediations (that I don’t consider perfect) to a new sort of mediation. I define here chatbots in a quite broad way: I include those which are easily generating videos, sounds, images, texts, etc, as more and more chatbots are said to be multimodal, so can handle at least text and images at the same type, as inputs and as outputs. With Sora 2, this definition may evolve, by the way. At the heart of those chatbots are models, whether language or diffusion models. Of course, chatbots have aditional features, and additional sorts of alignement, but I will focus on those models.

How do Generative AI models work? It’s basically machine learning: a training dataset is used from which the machine learning algorithm is deducing a set of probablistic relationsships between words or part of words (tokens) or set of pixels (when generating images or videos). Those probabilistic relationships (parameters) are together forming the AI model.

This mediation is hence probabilistic. When we follow thie French Female Citizen when she votes for the first time, we follow a sort of probable past, not a past that actually happened. As historians, we do not know everything about the past, there are holes in our knowledge, and there will always be some holes. The advantage of a probable past is that there’s no holes in the view on the past that is proposed by a chatbot.

cost

The price for such a probable past is very high.

Very high in terms of environment, very high in social terms. Because the training phase is often not enough, there will be a reinforcement feedback with human in the loop, a strange set of words that designates digital labor, usually underpayed high-skilled workers from the global south correcting some outputs of chatbots.

Another aspect of this price is datacolonialism: generative ai is based on the appropriation of the data we are all generating on a daily basis, when we use smartphones, computers and many other electronic devices.

A last aspect of the price of generative AI is linked to the probable past: it’s a change of regime of truth. Truth today is something probable, and this probable truth depends on the training datasets of generative AIs, datasets that have biases of very varied and different natures. Datasets that are usually closed, that we cannot analyse directly.

(public) history and the probable past

The thing is that the probable past might fit better the views on the past of a government, whether French or not, whatever its political nature. So the probable past is often an ideologically good enough past.

So, we are in a very specific context, that is a mixture of the rise of the generative ai chatbot, of rise of authoritarianism, of radical undermining of international law, of rise of ‘technofascism’, an ideology that basically exploits at its benefice this new regime of truth. In this context, what can be the role of historians and public historians?

I do not have a clear answer.

But:

it cannot be ignorance. We cannot just ignore generative AI, we cannot refuse to understand how it works, we cannot refuse to teach it when necessary. In other words, and to get back to a discussion we had 2 days ago, generative AI will not destroy us neither save us. That’s not the debate. We need to look precisely at what it does to our relationship to the past, not be doomers – doom does not help to think.
there are practices that already exists that should be emphasized. Here I am specifically thinking about public history. The best incentive to stimulate the abilities of our fellow citizens, of our students, of ourselves, whatever the country we live in, to criticise any AI slop generated about the past in our case, is to work together.
looking at our methodologies. Which one of them are underused today but could be useful? I’m a text-based historian, and that is not enough, we need to do more on images, moving or not, on sound, and on documents.
we need to look at how we teach.