negotiating the past
towards a new framework for digital memory studies?

frédéric clavert

5.11.2025 – ai and digital humanities – siena

historical representations

generative artificial intelligence

collective memory

Good morning.

First, I would like to thank you for inviting me today to talk. Thanks for your organisation. A special thank you to Michele who cannot be there. I am going to read a bit too much what I want to say, basically because I wannt to be precise – and some discussions we had yesterday show that we need to be precise, I think.

This talk will explore the intersection of historical representations, artificial intelligence, and collective memory.

Negotiating the past? Negotiating the past is not something new. There are many spaces where, as social groups, as a society, we negotiate the past: committees that define which historical knowledge students should learn, for instance, are spaces of negociation about the past.

Today, my argument will be that, though spaces of negotiations about the past are already existing, generative AI platforms open new ones, that have their own specificities.

I’ll try to examine how users “negotiate” with AI systems to express their conceptions of the past, how these interactions can reveal tensions between user expectations and AI-embedded historical patterns, and how to insert this into memory studies.

I’ll try first to show how LLMs encode historical perspectives, trying to set up a theoretical framework.
I’ll then show how I am using myself AI and LLMs to set up a corpus of prompts to analyse, and how I do analyse them.
I’ll then comment the analyses before concluding on chatbots as new spaces / frameworks for negotiation on the diverse ways we collectively see the past. I’ll gop back to this very draftuy theoretical framework.

(draft) theoretical framework

chatbots as medium of memory

I define here ‘chatbots’ as generative AI platforms designed for wide audiences, made of an interface that encourages users to enter a prompt, in order to generate texts, images, videos. Those platforms are more and more multimodal – inputs can be texts or images, as well as outputs. So we consider here chatbots that are based on diffusion systems and/or large language models.

I also consider specifically chatbots, and not only their underlying engines (large language or diffusion models), because chatbots have additional layers of filtering, feedbacks, alignments and fine tuning.

For instance, DeepSeek – the online chatbot – is well known to refuse to answer a question on the Tiananmen massacre. But if you used one of the DeepSeek R1 models, when they were released, directly on your computer (I’ve tested the 7b parameters one), it did not refuse to answer questions on the Tiananmen massacre, though it clearly indicated in its reasoning that it should be respecting the chinese law and sensibility on the subject. Even if, since then, they have been updated those models in order either to refuse to answer or to deny it even existed, the fact that the initial models runable on your computer would answer a question that the chatbot refused to answer show that there can be very strong differences between a model and a chatbot using this model.

The Tiananmen example brings us back to memory: in those additional layers of alignment and fine tuning that are made to allow wide audiences to use easily chatbots, there are additional views on the past that are modified or embedded.

As a consequence, those chatbots can be considered as medium of memory. I use here Astrid Erll’s work (Memory in Culture, 2011) where she writes that medium of memory are « “constructs versions of a past reality” and plays a role “in the encoding and decoding [Stuart Hall, 1980.] of that which is (to be) remembered.” » (p. 120ff). For Erll, medium of memory performs several functions:

they store information
they allow a form of ciruclation of this information
they can also be collective memory triggers

For instance, let’s take the example of the French numerous Monuments aux morts, that can be considered as medium of memory:

they in some ways allow for the storage of information,
though it is not their main function, they allow this information to circulate – basically they are a message to people passing by,
they are trigger of collective memory with the many commemorations that are organised around them.

Let’s go back to chatbots and let’s try to define them as medium of memory:

chatbots are based on models that are storing information. Models are sets of parameters deduced from a training phase, parameters that contain patterns based on training datasets. In this sense they can be seen as storage of information, including information on the historical past,
chatbots allow a form of circulation of information, when their users query them, even if their stochastic ways to restitute information, that does not include any sense of truth that can lead to hallucinations, should be carefully considered.
chatbots are also triggers of collective memory. That’s their interface, based on, often, a single box where users can type questions. Studying the past, in a professional way or a more amateur one, is all about asking questions and trying to find ways to answer them. Prompts are hence a huge incentive / trigger to query the past, even if it is not the main use of generative AI platforms, of course.

So, if we consider chatbots as medium of memory, we should also try understanding how they embed views on the past.

historical patterns

We’ll take the example of text generation through LLMs.

LLMs are made to “generate text”. They do that by using patterns, deduced from a training dataset, that is usually not very well known.

We should consider training datasets as primary sources, as an object of the historian’s inquiry and critique: a training dataset is an archive, it is made of historical records. But as all historical records, they contain specific views of the past – I could speak of ‘biases’, but I think this word is really not pertinent here.

And, the way the training phase is embedding those views of the past into the model is not at all the way we are dealing with historical records as historians. There’s no critical appraisal of those historical records, nor by the model itself or the engineers who conceived it. What matters are patterns, probabilities, hence the famous article of Bender et al. on stochastic parrots.

It says something about the fact that chatbots are changing our regime of truth – truth becomes a matter of probabilities – and this change of regime of truth have consequences on how we consider our past: through chatbots and our prompts, we inquire for a probable past. This probable past is not only the result of large language or diffusion models, but also the results on how chatbots are fine tuned and aligned (and of course of the prompt, so the user agency).

aligning the past

We should also look at the alignment, feedback (with or without humans in the loop), fine-tuning and any other operations that can be performed to create the chatbot itself. Alignment, for instance, is supposed to ensure that the answers of a chatbot, whatever their modalities, fit with the value of a society, whatever that means.

But this “alignment” also ensures that the firm publishing the chatbot gets money from it, which probably means that the answers must please a large number of users (and also the firm’s shareholders). That poses the question of minority voices and makes the choice of the training dataset a real decision of memory politics (see Smits et al and Walden-Richardson et al., 2024).

Of course, this alignement is obvious for chatbots that are published in an authoritarian state. We have seen the DeepSeek Tienanmen example before. I could also give the example of ruDALLE, a russian image generation platform, that will send back an image of flowers to all prompts mentioning something related to Ukraine.

Here is a different example: midjourney has filters that encourage the representation of minorities on the images they produce, which is obviously of good intent. But this can lead to abberations. In this image, the aberration is to represent Leopold II as a black king colonizing Congo.

the latent space of the past

Let’s go back to large language models and how they are created. One of the step of the creation of models is the latent space. Latent spaces are a way to compress a dimensional representation of data. It allows the model to generate new data based on the original data, while being more efficient because compressed.

Let’s take the notion of latent space as a metaphor to understand how LLMs, in this case, represent and render historical concepts, events, etc.

In this metaphor, historical concepts are encoded as points in a multidimensional space, made of word embedings that capture semantic relationships. Then the relationships between historical events, figures, and concepts are captured in the distances and directions between these vectors. Temporal relationships are encoded – or we can suppose so – in semantic proximity and cultural associations, including related to memory. This multidimensional space is ‘compressed’ in a latent space that captures all those relationships and, when activated by a prompt containing a reference to the past, will send back an answer. This latent space reflect collective memory patterns from the training corpus.

see Alban Leveau-Vallier’s talk and book

chatbots as frameworks

So, chatbots (and LLMs / Diffusion models behind them) are playing their role of encoding / decoding what is to be remembered through this “latent space of the past” – an expression, I prefer to state it again, that should be considered as a metaphor – this latent space is a specificity of chatbots considered as medium of memory.

Let’s go back to Astrid Erll and media of memory. One of their functions is also to play their role as media frameworks. Their affordances, including anthropomorphism, the ways they are structured, are framing what users can expect from those systems.

What are memory frameworks? This notion leads us back to Halbwachs and his seminal work on collectiev memory in the 1920s, les cadres sociaux de la mémoire. What does Halbwachs mean by “memory frameworks”: basically that you can not understand how an individual remembers if you do not consider the social group in which they are living, including how this social group is structured. It is this relationship to the group that frames the individual memory, as the group will have its own view on the past, view that is a consequence of the social relationships between individuals of the group, and of the insertion of this group into a wider society.

empirical approach

So my first part was an attempt – probably still too drafty – of theoricizing what are chatbots in terms of collective or cultural memory. I’ll get back to this in the conclusion, and will try to compare it to other theoricizing attempts.

But, let’s try to link this to an empirical approach. I am a historian, I don’t work without primary sources.

What I am trying to do here is to look at prompts and how users are negotiating with the chatbot to get an artefact with a view of the past that fits their wishes. That implies that users may have different prompting strategies, that their prompts evolve while discussing with the chatbot (through image or text) – I call this discussion part ‘stochastic maieutics’.

I will focus on prompts made of text, though images can be prompts too. I have not yet tried to insert into my research sora 2 – though, as you might have heard, the use of historical figures is quite popular, with very different ideological backgrounds. But that’s the next territory I will try to investigate, though collecting data will be hard, if not impossible.

In the next few slides, I’ll explain how I have built two corpora of prompts. Note, that I am using prompts that are published open access. They usually comes from open source communities, or are available online on website that offer an API to get prompts. The result is that those prompts are mostly coming from Stable Diffusion-based platforms, as Stable Diffusion is open source: those prompts are designed to generate images, not text, as stable diffusion is one of those text-to-image systems. That might change in the future, that might have already changed – maybe it will be possible to work on other corpora released by larger platforms. But to my knowledge, it’s not possible yet.

what is a reference to the past?

army of the european union invades budapest 2 0 2 2, highly detailed painting, digital painting, artstation, concept art

army of the european union fighting on the streets of budapest 2 0 2 2, highly detailed illustration for time magazine cover art

army of the european union with tanks fighting on the streets of budapest 2 0 2 2, highly detailed oil painting

Let’s take an example: the three prompts here are not refering explicitly to the past. They were issued in 2022 about a 2022 event, basically negociations, hard negociations, between Hungary and the European Union’s Commission, around cuts to European fundings to Hungary, because of Hungary’s tendency not to respect basic democratic rights.

Copy-paste the last of those prompts in an image search engine: all the results are historical. We have here an implicit reference, probably to the hungarian revolution of 1956, that was put to an end by a soviet intervention. Well, the results are now a bit biased, let’s say, as it’s not the first time that I am using this example, so the search engine also finds some previous prez…

So, the problem here is to build a robust identification strategy for implicit and explicit references to the past.

corpus 1

krea corpus (sample) → claude api → 5000 prompts

corpus 2

keyword (‘european union’) → api lexica.art → 2000 prompts

prompting to find references to the past

Let’s focus a bit on Claude: this prompt was co-written with Claude to help Claude reasoning for each item of my corpus. It’s a mixed approach: reasoning and personae, the prompt being enginered through a negociation between me and Claude (the chatbot).

It’s the use of Claude that worked the best (my very human evaluation). It’s far from perfect: the implicit part is not fully taken into account and some prompts are refering to the past, but not the historical past.

I must precise that for now I did not do any benchmarking, which is an obvious weakness, that I plan to adress. Furthermore, I used only a sample of the krea corpus for cost reasons. The plan is at one point to find prompts with references to the past in the full 10 mmillion lines corpus, which should allow me to get around 900 000 prompts with references to the past.

As I said, Claude is just sending a ‘yes’ or a ‘no’. I could also ask for a short ‘why’. Here, for now, I have limited to a simple binary answer for costs reasons. But, asking a ‘why’ would be interesting to analyse, including with distant reading tools.

In the next few slides, I’ll analyse them through scalable reading, where distant reading is at the same time used to start the interpretation and as a search tool for more refined analyses, including analyses of precise prompts or series of prompts.

negotiating the past

corpus 1

corpus 2

This is an anlysis of the European Union corpus, based on prompts published on lexica.art.

What you see here is interesting, because it’s a lot linking the European Union and Europe themes to all sorts of empire themes and to some sort of medievalism. We also find back the ‘propaganda’ cluster – probably because those prompts were produced in the wake of the russian agression against Ukraine.

Those distant reading analyses are interesting, but confirm more than discover: in a way, we expect Europe to be linked to the concept of empire, we expect that lots of references to the past are linked to the ‘style’ part of a prompt.

But those distant readings allow us also to go back to specific prompts. By looking at prompts that are very similar, we can trace the evolution of a prompt written / re-written several times by a user. And it’s here, that we can see that gen AI platforms, seen as frameworks, encourage users to negotiate with the machine the past they want to see or read.

This negotiation can be seen as a confrontation between several kinds of collective memories: the one that are embedded in the genAI platform and that comes from the way the LLM or diffusion system was trained and from the corpus it was trained on; the collective memory of the group the individual belongs too; the individuals own vision of the past.

I’ll give two examples.

Ursula von Der Leyne [sic] and Emmanuel Macron, Peter [sic] Pavel in the image of knights of the round table

I have showed before prompts that relate to the hungarian revolution – unfortunately I do not have the images.

Here is another example. Several prompts of this kind were written, with some differences, but results that are very similar.

As you can see, we can consider there are several references to the past here.

the reference to a myth, the knight of the round table, that is supposed to be medieval (even if it’s been heavily re-used since the XIXth century), but that is illustrated by an image that looks more Renaissance than medieval - which gives a hint on how collectively we see the Middle Ages (and it’s coherent with the 19th century revival of the round table myth).
as this image is from 2023, it is also a way to see how the memory of some politicians is being built – and obviously, Petr Pavel’s memory outside of the czesh republic is not very well built…
we could say also that Macron is slightly more recognizable than von der Leyen (who looks like an average 50s-60s-ish european woman). Though the fault in the prompt might induce that, Julien Schuh (2024) showed that there are differences between representations of famous men and women, with strong gender biases.

The user never managed to have Petr Pavel on their images.

joe biden doing a nazi salute, in front of brandenburger tor. huge nazi crowd in front of him. face of joe biden is clearly visible. canon eos r 3, f / 1. 4, iso 1 6 0 0, 1 / 8 0 s, 8 k, raw, grainy

This is a striking example of negotiation with a machine to get something in the present with references to the past that serves as an ideological reading of the present.

The user never got what they wanted. Biden doing a nazi salute in front of nazis – that just does not exist. Nevertheless, the prompt activated patterns from the second world war and maybe from the cold war.

I may be overinterpreting, but I can see influence, of course, from nazi footages – the way the masses are represented could be seen as similar to the ways masses are represented in Leni Reifenstahl’s Triumph des Willens, but also from De Gaulle in August 1944 (26) in the Champs Elysées, Kennedy in front of the Brandenburger Tor, etc.

The political goal of the prompt writer here is partly in failure: this political goal is confronted with the collective memory of the second world war embedded in the model, that are preeminent over their views.

we need more empirical research.

Chatbots function as media of memory that store, circulate, and trigger collective memories. They are medium of memory because AI models encode historical perspectives. Metaphorically, we could call that the collective memory latent space, that reflects collective memory patterns from training corpora. On top of the training process, alignment and fine-tuning embed specific views of the past (e.g., DeepSeek/Tiananmen, minority representations). In this sense, we can also consider chatbots as memory frameworks.

When we look at users’ prompt that have references to the past, we see how users are writing those references to the past, but also how they negotiate with gen AI systems to obtain their desired vision of the past, creating a confrontation between different collective memories, but also, often, their vision of the present.

This negotiation reveals tensions between user expectations and historical representations (frameworks) embedded in AI.

Beyond this, I think we should remind something quite important about LLMs and Diffusion System: they are the products of artefacts from the past (the training dataset), they are producing primary sources (prompts, images, texts), and they are triggers of collective memory.

In this sense, GenAI systems are fundmentaly products from history and memory.

I’d like to go back a bit on my promise to speak about a new framework for digital memory studies.

The reference is Andrew Hoskins Digital Memory Studies book, in 2017. Hoskins tries to explain how the web, social media, big platforms have been transforming our relation to the past into a restless past that generates a memory of the multittude rather than a collective memory through a media, the web, that is ‘bigger than us’, that human can not encompass fully.

Since then, some more critical articles and ChatGPT. Two special issues have tried to speak about collective memory and generative AI: one in Memory Studies Review that I co-coordinated with Sarah Gensburger (Sc. Po.), and another one (well, rather ‘a collection’) led by Andrew Hoskins for Memory, Mind and Media. Some authors are common, and we are working on an article for the Memory, Mind and Media collection.

In his introduction to the collection, Hoskins evolved from the restless past and the memory of the multitude to the ‘third way of memory’, neither human nor machine, but hybrid, that he articulates with what he calls the untethered past and conversational memory.

I prefer speaking of probable past and negotiation around the past. Negotiations because there is, from what I could see in the interactions between users and chatbots, often a confrontation between the users’ view of the past and the embedded views of the past into the model and the different layers of alignment and fine tuning that makes the chatbot. I think the user agency is far higher than Hoskins think it is, all the more that models remain the result of human agency, and of human data – for now. But even synthesised data has something human.

Furthermore, there is the question of social relationships, that is not adressed enough by this “third way of memory”, or that are evacuated. But without social interactions, there would be no training dataset. We are in a paradox, that is described by Pierre Depaz in an article on GitHub and StackOverflow considered as a collective memory of developers: the training dataset is based on social interactions between developers, that created a training dataset for chatbot to be good developers, but destroying the social ethics of developers that allowed the creation of this collective memory. In Carl Öhman’s view, it’s a balance of power between our present us and our past us, which is embodied by the models.

What is certain is that we need to engage more with chatbots and their effects on memory, on our disciplines, on social life in general. We need empirical research about generative AI, and its encoutners with history and memory studies.

thank you

Gensburger, S., Clavert, F., « Is AI the future of collective memory ? », Memory studies review, 2024
Hoskins, A., « AI and Memory », Memory, Mind and Media, 2024-…

negotiating the pasttowards a new framework for digital memory studies?

frédéric clavert

5.11.2025 – ai and digital humanities – siena

(draft) theoretical framework

chatbots as medium of memory

historical patterns

aligning the past

the latent space of the past

chatbots as frameworks

empirical approach

what is a reference to the past?

corpus 1

corpus 2

prompting to find references to the past

negotiating the past

corpus 1

corpus 2

thank you

negotiating the past
towards a new framework for digital memory studies?