2026-04-29_CtrlAltHistory

## From *Ctrl-Alt-Del*...

<a href="https://commons.wikimedia.org/wiki/File:Three-finger_salute.svg#/media/File:Three-finger_salute.svg">A Qwerty keyboard (scheme), STRG-ALT-ENTF highlighted. Sven. CC BY-SA 3.0

Note:
There is a theoretical framework on the temporality of updates, crashes and reboots as a cultural regime, in a quite important work by Wendy Hui Kyong Chun [@Chun_2016].

She uncovers a paradox: we update and restart precisely in order to stay inside the same habitual media environment. In this framework, rebooting is not seen as an isolated gesture but as a structure of habitual digital life: we restart constantly in order to continue.

The reboot is what we reach for when we no longer understand what is happening inside the machine. It is, in a way, an act of epistemic surrender: we give up trying to debug, and we start over.

Since 2022 and the rise of chatbots that high-lighted the current trend of generative-AI-based innovations, have we, as historians, lost our understanding of what is happening around us and inside our main tool, the computer, so that we need to reboot?

## ...to *Ctrl + Alt + **History***

**History as what is threatened**

**History as what allows us to reboot differently**

<a href="https://commons.wikimedia.org/wiki/File:Nintendo-Super-Famicom-Console-FL_(Reset_Button).jpg">Evan-Amos, Public domain, via Wikimedia Commons

Note:
This conference replaced *Del* with *History*. I read this substitution in two incompatible ways, and I am going to keep them both alive across the whole keynote.

- First reading: **history as the threatened term.**

The *Del* key, in this reading, is generative AI. History is what stands to be deleted: automated archival work, interpretation replaced by statistical prediction, hermeneutic traditions flattened into tokens. In this reading, Ctrl-Alt-History is a cry of alarm.

- Second reading: **history as the response.**

History, here, is what we reach for when our system of knowledge freezes. When the epistemology of the chatbot leaves us speechless, when we do not know how to evaluate what a machine "tells" us about the past.

THen we reach for the historian's craft.

In this reading, Ctrl-Alt-History is a programme: use history to reboot the debate.

I am going to refuse to choose between these readings today.

I like the tension between them, a tension I see as a strong motivation. I will come back to this tension in the conclusion.

## Look, they have six fingers

Author with DALL.E (2023)

Note:
I will add one more comment to this introduction. We need to move away from chatbots, or at least from the way chatbots are putting together words to form some sort of narration of the historical past. Those narrations (if there's no RAG, if there's no correct prompt) look like the images that were generated three years ago: metaphorically they have six fingers.

This is not nothing. Factual errors in historical discourse matter, and the scale at which generative systems are now producing such discourse is genuinely new -- the scale is new, not misinformation, of course. I am not here to dismiss these concerns.

But *this is the wrong question* around which to organise our debate. If the only question we know how to ask is "does AI get the facts right?", we remain stuck inside an image of AI that is obsolete — the AI-as-oracle. And the real transformation of our practice is happening elsewhere, in the places where this question does not reach.

So something is blocking when we speak about AI -- here generative AI -- to many historians, students, etc. I want to argue that what is blocking is our image of what AI is. By 'our', I mean historians collectively.

## A diversity of epistemic signatures

- Language models (LLMs)
- Agents and automated workflows
- Research infrastructures
- Corpus interrogation tools
- Systems connecting heterogeneous sources

Note:
- AI ≠ chatbot. The landscape:
  - LLMs (often silent inside pipelines — classify, extract, embed)
  - Agents/workflows (not answering — *operating*, with tools)
  - Research infrastructures (→ P2)
  - Corpus tools (→ P3 Lester)
  - Systems connecting heterogeneous sources (the decisive one)
- Each has different failure modes and epistemic commitments
- Collapsing them all = "microscope = all scientific instruments"

**A language model** is a statistical system trained to predict the next token in a sequence. It can be deployed as a chatbot, usually with many other layers of software, but it can also be deployed silently inside a pipeline — to classify documents, to extract entities, to align translations, to generate embeddings for semantic search. Most of these uses have nothing to do with dialogue.

**Agents and workflows** are systems in which a language model is given tools — a search function, a database query, a file reader — and invoked iteratively to accomplish a task. The model is no longer "answering"; it is operating. This is a very different epistemic object.

**Research infrastructures** are the layer where AI meets the cyberinfrastructure tradition — I will return to this at length in Part 2.

**Corpus interrogation tools** are what let us ask semantic questions of messy archival material — more on this in Part 3.

**Systems connecting heterogeneous sources** are for me a decisive point. The most interesting thing generative AI does, for historical research, is not produce text. It is connect things that did not previously connect.

Each of these has different failure modes, different epistemic commitments, different relationships to the historian's craft.

## Connecting

<a href="https://commons.wikimedia.org/wiki/File:Pont_du_Gard_v2_082005.jpg">Pont du Gard, août 2005</a>, Patrick Clenet. CC BY-SA 3.0

Note:
The main epistemic shift introduced by generative AI in our field is not about what these systems *say* or *answers*. It is about what they allow us to *ask* — and in a way, this question has been present since the advent of digital humanities and digital history as we know them, for the past twenty years.

For instance, semantic search across a noisy corpus means we can find conceptual proximities where keyword search would have failed. There were techniques that existed before machine learning, but machine learning and language models today enable something much more efficient.

Dynamic query reformulation means the research question can evolve *during* the search, not only before and after it. Bridging structured and unstructured sources means we no longer have to pre-model our archive in order to interrogate it.

And — this matters — these tools *negotiate* rather than *impose*. An agent asks itself whether it has enough information. It calls one more tool. It revises its approach. This is categorically different from a database query, which either returns results or does not.

I will demonstrate this concretely in Part 3 with the Sean Lester dataset. For now, I ask you to hold this substitution in mind: not AI as *speaker*, but AI as *connector* and hence as infrastructure.

## The displacement we need

| From... | To... |
|---|---|
| AI as *oracle* | AI as *infrastructure* |
| "Does AI get things wrong?" | "What does AI change about what we can ask?" |
| Content reliability | Transformation of practice |

Note:
- **Oracle → Infrastructure**

An oracle is consulted; an infrastructure is inhabited. We do not evaluate an infrastructure by asking whether it tells the truth. We evaluate it by asking what it enables, what it makes impossible, whom it serves, and whom it excludes. Moving from *oracle* to *infrastructure* is moving from a Delphi frame to a Latourian frame — from a question about pronouncements to a question about the sociotechnical assemblage.

- **"Is AI wrong?" → "What does AI change?"**

The first question is evaluative and terminal: we give the system a grade. The second is genealogical and ongoing: we ask how our practice is being reconfigured. The first question is answered once and for all; the second has to be asked again every time the practice shifts.

- **Content reliability → Transformation of practice**:

This is the most important shift for us as historians. The epistemology of our discipline is not primarily about the reliability of individual statements; it is about the traceability of how we came to those statements. The transformation of practice — *how* we search, *how* we connect, *how* we revise — is where the epistemological action is.

## Uncertainty as a core feature

<a href="https://commons.wikimedia.org/wiki/File:Set_of_dice-Saint_Raymond-IMG_0107-white.jpg">Rama</a>, <a href="https://creativecommons.org/licenses/by-sa/3.0/fr/deed.en">CC BY-SA 3.0 FR</a>, via Wikimedia Commons

Note:
We read fragmentary evidence. We formulate hypotheses. We revise them in the light of new sources. We make arguments that could, in principle, be overturned by the next find.

We have not, as a discipline, developed a shared vocabulary to describe this probabilistic navigation. We have left it implicit, as a kind of tacit craft. One of the things generative AI is doing — and this is what I want us to notice — is making this probabilistic navigation *explicit*. Not because the machines reason the way we do, but because working with them forces us to put into words what we do.

The AI system that revises its own queries is a mirror for the historian who revises her own hypotheses. This mirror is, perhaps, the most interesting thing AI gives us.

## A long established diagnosis

The **archipelago syndrome**

> Dacos, Marin. <a href='https://shs.hal.science/hal-00871765'>‘Cyberclio. Vers Une Cyberinfrastructure Au Cœur de La Discipline Historique’</a>. *L’histoire Contemporaine à l’ère Numérique / Contemporary History in the Digital Age*, edited by Frédéric Clavert and Serge Noiret, P.I.E.-Peter Lang, 2013, pp. 29–41.

Note:
In 2009 (published 2013), during a conference I organised the first time I worked in Luxembourg, Marin Dacos, at the time head of what would later become OpenEdition, talked about the Humanities as a digital archipelago (in terms of data, methods, platforms, etc.).

The archipelago syndrome is about dispersed corpora, inert documents, an impossible cartography. His answer was -- remember we're in 2009 -- a **cyberinfrastructure** (borrowed from NSF Atkins Report 2003).

In a way -- and I don't want to underestimate what many projects on linked open data for instance did since then -- we can argue that the current trend of innovation in AI could be the most serious attempt, more than 15 years later, to deliver this cyberinfrastructure that would connect the different islands of the historians' digital archipelago to one another.

## Did the archipelago persist?

> Historical data is not a kitten, it’s a sabre-toothed tiger (Lermercier / Zalc)

Note:
Between 2010 and 2025, massive investments were made — at national level (Huma-Num in France, CLARIAH in the Netherlands), at European level (DARIAH and CLARIN as ERICs, European Research Infrastructure Consortia), at project level (countless Horizon 2020 projects).

And yet: any historian in this room, sitting down to a new research project today, still experiences the archipelago. The project-specific database. The local metadata schema. The tool that works for one corpus but not the next. The shared standards that required so much upstream ontological work that most scholars simply did not use them.

Here, the different epistemic instances of AI remove the requirement of prior standardisation. AI agents, for instance, negotiate across heterogeneity rather than imposing homogeneity, of course with limitations.

During a meeting at the C²DH, my colleague Caroline Muller (from Rennes 2) and I set out a series of provocations. One of them was: history deals with messy data. Claire Lemercier and Claire Zalc even argue that it's what is good about historical data: "Historical data is not a kitten, it's a sabre-toothed tiger".

| Moment | Answer | Limit |
|---|---|---|
| 1990s | Databases, GIS | Disciplinary silos |
| 2000s | Linked Open Data, cyberinfrastructures | Rigid upstream standardisation |
| 2010s | DARIAH, CLARIN | Institutional scale, low agility |
| Today | AI agents | To be evaluated — the subject of this keynote |

- Walk through the four rows on the vertical sub-slides (↓)
- Each row: a different *kind* of answer to the same archipelago problem
- Today's row = what we are evaluating — *structurally different* answer

## From linking data to linking practices

<a href="https://www.instagram.com/nadning/">Nadia Nadesan</a> & <a href="https://digital-dialogues.co.uk/">Digit</a> / <a href="https://betterimagesofai.org/images?artist=NadiaNadesan&title=DeceptiveDialogues">Deceptive Dialogues</a> / <a href="https://creativecommons.org/licenses/by/4.0/">Licenced by CC-BY 4.0</a>

Note:
Previous generations of connective technology operated at the level of data: they required the sources to be pre-modelled in a shared representation. Databases needed schemas. Linked data needed ontologies. The intellectual labour was all upstream.

AI agents operate at the level of *practices*. An agent can read a PDF of poor scan quality, a structured database, a blog post, an archival description, and a tweet, and reason across them without requiring that they be pre-modelled in a common schema. This is not magic; it is the combination of two things: (a) language models can produce robust semantic representations of unstructured text on the fly, and (b) agents can decide, mid-task, to call additional tools or sources.

This is a genuinely different epistemic object. Whether it delivers on its promise is an empirical question.

## Stochastic maieutics

> AI as the interlocutor who helps thought to be born.

Note:
Go back to the Sean Lester experiment.

- part of this keynote was born in a conversation with Claude about MCP servers → drifted to cyberinfra → Dacos 2009 → structure emerged
- Not a confession — **the demonstration in act** of what I'm arguing
- The keynote is itself a cyborg artefact; genesis traceable in a log
- So is much of the intellectual work I produce today — and much of yours, named or not
- Turn now: from demonstration to consequence

- *Maieutics* = Socratic, **Plato, *Theaetetus* 148e-151d** — philosopher as midwife, adds no content, facilitates delivery
- *Stochastic* = interlocutor is probabilistic (LLM generates statistically plausible)
- Empirical finding: those plausible responses are precisely the ones that let me formulate my next thought
- Labour mediated by stochastic process → thinking-*with*, not outsourced thinking
- ≠ oracle (consulted for answers); = maieutic interlocutor (engaged for the labour of thinking-with)

## Assuming a probabilistic epistemology

Note:
Historical work has always been probabilistic navigation. We read a fragment. We form a hypothesis about what it suggests. We read another fragment. We revise. We identify patterns that are never certain. We make arguments that are always defeasible. This is not a weakness of our craft; it is its specific mode of rigour.

The AI agent — with its visible sequence of query revisions, its explicit uncertainty, its iterative reformulation — holds up a mirror. In the mirror we see, externalised, something we have always done internally.

- Question: is the AI moment the one where the discipline finally **owns** the epistemology it's always practised?
- If yes: new ways to think *error*, *revision*, *disagreement*; new vocabularies for dialogue with public, other disciplines, machines now in the room

I want to leave you with a question, not a claim.

Our discipline has resisted the explicit thematisation of its probabilistic character. We tell ourselves we do empirical research. We produce "findings". We "establish" facts. The vocabulary is indicative, not modal. We do not say "this claim holds with probability *p*" — we say "this happened" or "we do not know".

But our practice, at its best, is modal all the way down. The tentative interpretation, the competing hypotheses, the arguments from silence, the weighing of sources of different reliability — all of this is probabilistic reasoning in a vocabulary that refuses to say so.

The question I leave you with: is the AI moment — this moment in which we find ourselves working daily with systems that are openly probabilistic — the moment in which our discipline finally *owns* the epistemology it has always practised?

If so, this is no small thing. A discipline that owns its probabilistic character can think about error differently. Can think about revision differently. Can think about disagreement differently. And can — perhaps — find new vocabularies for dialogue with the public, with other disciplines, and with the machines now in the room.

-->