Report of the workshop "IA and revision" of the Revue3.0 2025 plenary meeting

This workshop was led by Alexia Schneider, Servanne Monjour, Nicolas Sauret, and Clara Grometto.

Marcello Vitali-Rosati opens the discussion: “Artificial intelligence” is a commercial label; we can observe a correlation between the tasks taken over by AI and the fields or domains in which symbolic value is drastically lowered. The moment when a task has lost enough symbolic value, when it is delegated to subalterns, that is when a machine can do it. If a woman can do it, a machine can do it. In Jennifer S. Light’s work (When Computers Were Women), “robot” is the name of a person whose work has been devalued.

Two different angles:

Revision

Definition of revision and brief technical history

Two distinct elements:

One cannot speak about revision without speaking about writing. Writing works through trial and error. Digital writing induces changes: our relationship to the trace is modified (Derrida). We write faster, but we can also correct more quickly.

The stakes of automating textual revision

It is essential to understand that there has been a paradigm shift, following a much older historical lineage. The evolution of tools follows—or is followed by—the evolution of digital writing: confusion emerges between writing and revision.

The history of Grammar Error Correction as a secondary task parallels that of machine translation. Correcting a sentence relies on a gold standard (incorrect sentence vs. ideal corrected sentence). These strategies, based on expert systems, struggle to reflect linguistic diversity.

From the 2000s onward, semantic processing began to be incorporated.

We then shift from an ortho-typographic correction paradigm to a process of reformulation.

Today, we even attempt to mask the use of AI in text generation — creating contradictions.A matter of value systems

Word processors:

Originally, tools were created by computer scientists for computer scientists, to document code, before being adopted in business contexts. Word processing initially aimed at productivity, and the use of LLMs follows this same logic. There is a hierarchical and symbolic distribution of tasks. The goal is to delegate the secretary’s work to machines, which transforms practices, uses, and tools. Beyond questions of meaning and the loss of control over text, there is a risk of linguistic homogenization, since the richness of language does not lie in saying the most probable thing.

The academic world is particularly affected, with a push toward standardized writing and significant economic stakes tied to “publish or perish,” writing in English, and paid revision services for non-native authors. Deep semantic reformulation is becoming standard by default, as shown by the options offered in Deepl. This produces a flattening effect and grants authority to the machine. Schumacher’s 2023 dissertation illustrates this phenomenon; his study shows that cognitive delegation can be a source of error (even when working with experienced translators). Finally, these tools embody a worldview centered on productivity and speed.

Questions

M. Sinatra: Are there studies distinguishing what type of production these tools are used for? A homework assignment versus having an article corrected: the stakes are not the same.

Taboo: It is difficult to get students—like researchers—to talk about their practices.

Pressure: Produce more, produce better; a pressure found at all academic levels and ages to reach a certain linguistic standard.

Servanne Monjour: In France, students face major problems with French proficiency, notably intelligibility issues. Her department chose to completely forbid students from using AI. Observation: in-class assignments display significant orthographic and syntactic problems, while dissertations and take-home assignments are of much higher quality. She denounces an unacknowledged use of writing-assistance tools. This situation forces teachers to question their role and pedagogical practices.

Nicolas Sauret returns to the idea of déprise borrowed from Louise Merzeau. She notably spoke of a “mastery of déprise,” meaning the ability to détourne or hack the platform. He observes that today this slips out of our hands; we are in total déprise, leading to a form of proletarianization.

Is encouraging people to set up their own LLMs the solution? Considering how hard it already was to get people to adopt open formats, the outlook is pessimistic.

Bibliographic revision

Editors have always sought automation. Automation can provide some comfort, but there is always a need for a bit of manual work (“the last mile,” as described in the journal Humanités numériques).

Emphasis is placed on the tension between the visibility and invisibility of revision work.

Florence, editor of Humanités numériques: “I keep using Word because it is the proof that I have done my work.” The Track Changes feature makes the work of the “little hands” visible.

Implementation of a charter for the journal’s experiments:

Specialized LLMs — local whenever possible.

Is it actually useful?

The time saved is re-lost in checking the LLM’s output. If everyone used Zotero correctly, the problem would be solved.

Digitization has led to strong standardization of data. With LLMs, we observe a paradigm shift.

Final questions

Marcello: Does agreeing past participles still have value? What do we want to do? Do we still want to produce value? “At what point do we avoid falling into the ‘grumpy old man’ paradigm?” (regarding the idea that students write increasingly poorly, read less, etc.)

Lucia: Historically, language has been a strong political vector (“empowerment”), an emancipatory force. It is important to preserve our capacity to write and express ourselves.

Servanne: Mentions the practice of arpentage (collective reading from popular education) reintroduced at the university.

Marcello’s aside on originality (not a human prerogative, built on false premises): “I am not original. When someone says something original, it is generally stupid and wrong.”