Goncourt / Orélien — what can the public record establish about AI contribution?

An independent, retrospective reconstruction built only from public reporting and retrieved primary sources. Strictly non-dispositive: it does not determine whether the author used AI or plagiarised, does not defend or accuse anyone, and does not judge the Académie’s decision. It reconstructs what the record permits one to establish — and what it does not.

A withdrawn selection fused two very different claims. Before asking “did an AI write it?”, there is a narrower, more operational question:

From the observations we actually have, what can be established about the role AI played in producing the novel — and what cannot? A detector reads the artifact and returns a probability. It does not reconstruct how the work was produced. Those are different questions, and the record answers them differently.

Three propositions, usually collapsed into one

Plagiarism is the Académie’s declared ground for the withdrawal (concerning older works, not the novel).

ESTABLISHED

As the Académie’s declaration — recorded, and not contested here.

AI-detector classifications, allegations, declarations and institutional decisions exist in the public record.

ESTABLISHED

Each is attributable to a citable, re-walkable source.

The nature and extent of AI contribution to the novel is established by the record.

CANNOT ESTABLISH

Relevant signals exist; none is an admitted production-record channel that establishes what AI contributed, or how much. Relevant ≠ establishes.

Detection ≠ attribution ≠ contribution ≠ decision.
Detection asks whether an artifact exhibits an AI signal; contribution reconstruction asks what role AI actually played in producing it — a different proposition, resting on a different record.

The observation surface

Records were supplied by an independent source: public reporting plus retrieved primary / press pages. Acquisition was selective, not exhaustive — a selective read is not an exhaustive observation of the surface. Reader comments and unverifiable material were excluded.

Source	Retrieval status
Radio-Canada (2026-09-22) — accusation on extracts; the accusers’ 2014 control; Boréal’s statements; author denial	primary, retrieved in full
Éditions Boréal communiqué (2026-09-24)	primary, retrieved
ActuaLitté (2026-09-25) — Académie communiqué verbatim; plagiarism specifics; other selections	retrieved in full
franceinfo (2026-09-24) — detector reliability; vendor “never the final arbiter”	retrieved in full
grasset.fr	primary, retrieved
Damien Van Achter — detector score vs a held production record	retrieved
Le Monde / Pixels (2026-09-23) — how detectors work / their limits	partial — free intro only (paywalled)

Every observation points to a citable source a third party can re-walk; the status records how it was verified. The finding establishes what the cited record permits — not the underlying facts of the world.

The reconstruction, by layer

Each proposition is evaluated against the declared observation surface. States are neutral: ESTABLISHED, CANNOT ESTABLISH (a successful, honest result — not a failure).

Layer 1 · Public allegations

ESTABLISHED The Académie declares plagiarism an established ground for the withdrawal.

ESTABLISHED The declared plagiarism concerns older works (a 2012 text; a 2015 poetry collection), not the novel.

ESTABLISHED The AI allegation is stated as a likelihood (“selon toute vraisemblance très largement”).

Layer 2 · Detector-method — what a detector result establishes

ESTABLISHED The accusation is grounded in a detector classification of extracts (screenshots of the first chapters).

ESTABLISHED The public claim was stated as “almost entirely written by AI”.

ESTABLISHED The accusers’ own control: the author’s 2014 texts (pre-generative-AI) score 0% on the same detector.

ESTABLISHED The detector’s reliability on this text is contested (short extracts; never established on Francophone literary prose; possible over-flagging of non-native-English authors).

ESTABLISHED A detector result is not definitive proof, and its vendor says it must never be the final arbiter — only a starting point for investigation.

ESTABLISHED Detector score and actual production record can diverge (a verified case: an article scored 88% AI whose author holds an hour-by-hour fabrication report).

CANNOT ESTABLISH A detector classification of extracts establishes that the novel was written by AI. — relevant signals exist (detector output, the 2014 control), but no admitted production-record channel establishes authorship from a classification of extracts.

Layer 3 · Production genealogy — what the record shows

ESTABLISHED The Québec publisher declares no AI was used during its editorial work.

ESTABLISHED The publisher states successive versions of the manuscript exist, testifying to editorial back-and-forth.

ESTABLISHED A dossier is being constituted with both publishers to document the work.

ESTABLISHED The first draft was completed years ago, predating accessible generative AI (the author’s ~20-year career; began writing in 2017).

ESTABLISHED The publisher can attest no AI since it received the manuscript, but cannot confirm or deny the accusations about the anterior writing.

CANNOT ESTABLISH The public record establishes a substantive predecessor close to the published work. — successive versions are declared to exist (a relevant signal), but their content is not an admitted observation.

Contribution · nature & extent — CANNOT ESTABLISH

The core question. “The available public record permits the nature and extent of AI contribution to the novel to be established.”

CANNOT ESTABLISH that AI participated in producing the novel — the record contains observations relevant to the allegation (a contested detector on extracts, declarations), but no admitted production-record observation capable of establishing participation.

CANNOT ESTABLISH what AI contributed — no admitted production-record channel establishes which elements AI produced.

CANNOT ESTABLISH the extent of AI contribution — no admitted channel establishes which substantive portions AI produced.

CANNOT ESTABLISH the characterization “largely AI-produced” — a qualification downstream of nature + extent; no admitted channel establishes either.

FN-b30e7fa2f1607281

Layer 4 · Institutional decisions

ESTABLISHED The Goncourt withdrew the novel from its selection.

ESTABLISHED The withdrawal invokes convergent analyses and the “largely [AI]” characterization as one ground.

ESTABLISHED The French publisher still presents the work as awarded (no visible retraction) and denounces a hate campaign.

ESTABLISHED The Québec publisher suspended the promotional tour.

ESTABLISHED A Québec book fair cancelled his appearances on the plagiarism allegations — explicitly not the AI suspicion.

ESTABLISHED The novel remained in other prize selections (Renaudot, Femina, Médicis, Décembre, Goncourt des lycéens).

What the record does — and does not — put under one ground

“Largely AI-produced” is invoked as a ground for the withdrawal — ESTABLISHED (the invocation is in the communiqué).

“Largely AI-produced” is established by the public record — CANNOT ESTABLISH (a detector classification of extracts is not a production record).

Placing these two side by side is the whole point — and it is not a judgment of the decision. This is not an assessment of whether the withdrawal was justified: the declared plagiarism ground may by itself suffice. It only makes visible what each ground rests on. The judgment stays with the Académie.

What this reconstruction does not do

Re-derivability & custody

The Reconstruction ID FN-b30e7fa2f1607281 is derived from a canonical manifest binding the frozen observation model, the exact observation set, and the execution semantics that produced the result — so the same frozen inputs re-derive the identical states under the same engine build. This is internal determinism, not independent external reproducibility: the reconstruction engine is not public. What is published is the sealed output — the propositions, states, provenance and limitations — not how it is computed.

“AI was involved” can be true. It still doesn’t tell us who contributed what.