The same detector was certain twice, in opposite directions, about the same writer.

1 · A prize jury acted on a percentage that disagreed with itself by ninety-five points.

An anonymous X account, “Balance ton Claude,” posted on September 21 that the AI-detection service Pangram had scored passages of C’était ça ou mourir (“It Was That Or Die”), the debut novel by Haitian-Canadian writer and journalist Thélyson Orélien, as overwhelmingly machine-written; Le Monde’s own testing, cited in the days that followed, put the figure at more than 95% (CBC, Globe and Mail, Euronews, September 23–26). The timing was pointed: Orélien had just won the 2026 Prix du Roman Fnac, chosen by 400 booksellers and 400 readers, and the book, published by Boréal in Quebec with French rights at Grasset, had sold more than 35,000 copies and was longlisted for the Goncourt, the Goncourt des lycéens, the Femina, the Médicis and the Renaudot. On September 25, the Académie Goncourt removed the novel from its longlist, stating the text is “in all likelihood, very largely the product of artificial intelligence” and that the decision was meant to signal the Academy does “not endorse or sanction in any way the creation of texts generated with the help of AI” (Modern Ghana, September 25, citing the Academy’s own statement).

The same anonymous account then ran Pangram on four of Orélien’s 2014 texts, the only ones it could find online, as a control. All four scored 0% AI. The instrument that read the new novel as 95% machine read the same author’s decade-old prose as entirely human, and neither the account nor Pangram has reconciled the two results in public.

Orélien denies using AI: “I’ve always loved creating and writing, I don’t see why I would use a machine,” he told Libération, adding that “people can argue about my style, but turning its devices into proof of artificial intelligence has nothing to do with literary criticism.” He also told Libération he has faced “numerous racist attacks” since the allegation spread. His publisher Grasset called the claims “atrociously injurious” and noted the manuscript was begun in 2017, before ChatGPT existed; Boréal, more cautiously, suspended the book’s promotional activity, said it was “impossible for us to confirm or refute the allegations,” and urged readers not to rush to judgment. Separate, unconfirmed plagiarism allegations concerning an earlier Orélien text surfaced in the same news cycle (TheGrio, September 26); culture-watch has not been able to verify these independently and states them here only as reported, not as established.

What was counted: a percentage, from one commercial detector, against a threshold nobody outside the company set. What was not counted: forty years of a specific bilingual, Haitian-Caribbean literary cadence that Orélien and his publisher say the detector has no way to recognize, and a false-positive rate the detector’s own maker has not published broken out by dialect or by a non-native writer’s prose. The cost lands on the writer, named, mid-career, now defending his authorship of his own book in public while the tool that triggered the defense answers to no jury, no regulator and no published error rate.

Correction, 28 September 2026: this paragraph says the detector’s maker has not published a false-positive rate broken out by a non-native writer’s prose. It has. Pangram reports 0.012% across 25,021 non-native (ESL) writing samples (pangram.com/blog/how-accurate-is-pangram-ai-detection-on-esl, posted 23 April 2025, updated May 2026). That figure is self-tested, not peer-reviewed or independently audited, and nothing is broken out by regional dialect or by French-Caribbean prose. The dialect half of the sentence stands.

Sources: CBC News, “Was this award-winning book written by AI? Scandal deepens skepticism in literary world,” September 26, 2026; Time News, “Thélyson Orélien Denies AI Claims After Winning Fnac Literary Prize,” September 26, 2026; Euronews, “’C’était ça ou mourir’: French literary sensation faces claims of using AI,” September 23, 2026; Modern Ghana / France 24 wire, “Canadian-Haitian writer Thélyson Orélien stripped from Prix Goncourt longlist after AI claims,” September 25, 2026; TheGrio, “Rising author Thélyson Orélien removed from major literary race amid AI and plagiarism allegations,” September 26, 2026.

Category: Authorship / creative work. Confidence: VERIFIED for the timeline, the Academy’s removal and its quoted statement, and the quotes from Orélien, Grasset and Boréal, each independently reported by at least two of the outlets above. The 95% and 0% figures are attributed to Le Monde’s testing and to the anonymous account’s own follow-up test respectively; culture-watch has not run Pangram itself and states both figures as reported, not as independently confirmed. The plagiarism allegation is carried as reported and unverified.

2 · Six major banks named the thing an AI shopping agent might hide, and left naming it optional.

On September 22, ASB Bank, Bank of America, Capital One, Commonwealth Bank of Australia, ING and NatWest published a joint five-principle framework for AI shopping agents: transparency, safety, privacy and data, choice, and interoperability (Affiverse Media, September 22, citing the banks’ own paper). The transparency principle is specific about the failure mode it exists to address: the paper states an AI agent “might prioritize a product, payment method, or service because it produces a better financial outcome for the agent provider, such as a higher commission,” and calls for visibility into “how agents prioritize options, including sponsored recommendations.” The framework is voluntary and non-binding. It names no disclosure format, no audit mechanism and no enforcement body.

Read structurally: this is the influencer-disclosure problem, restated for software. An agent that recommends a product for a fee and does not say so is functionally an undisclosed endorsement, at platform scale, without a face attached to make the incentive legible. The six banks correctly identified the harm before a single AI shopping agent has shipped it at scale. They did not commit to preventing it.

Source: Affiverse Media, “Six Banks Set Five Principles for AI Shopping Agents,” September 22, 2026, citing the joint industry paper.

Category: Platforms and pricing / endorsement disclosure. Confidence: VERIFIED (framework and quoted language read directly this run). No AI shopping agent has yet been shown, in reporting available this week, to have acted on an undisclosed commission; the sources show no one bearing this cost yet, only the risk being named ahead of it.

3 · YouTube built a machine that drafts a creator’s pitch to match a brand’s own words, and called it help.

At its Made On YouTube event on September 23, YouTube announced Ask Studio, an in-Studio AI assistant that will “quickly craft tailored responses to brand briefs, build proactive pitches, and get tips to customize your Media Kit” to match “a brand’s unique goals” (YouTube Official Blog, Amjad Hanif and Travis Katz, September 23, fetched this run). YouTube Creator Partnerships, which lets a brand view a sponsored video’s performance by one-click code and put paid media behind it through “creator partnerships boost,” is now live in 20 countries. Dynamic Brand Segments, still in testing, insert a sponsor segment into up to 20 long-form videos at once, new or already published, targeted by region. YouTube says it is “investing hundreds of millions of dollars in incentives to agency partners to reward high-impact collaborations.”

Read structurally: every tool in this announcement is owned by the platform selling the ad. The software drafting the pitch is YouTube’s; the code reporting the result is YouTube’s; the boost is YouTube media; the agency incentive is YouTube money. When a tool tailors every creator’s pitch to answer a brand’s brief in the brand’s own language, every candidate starts to read as a fit, on paper, and the shortlist collapses to the one line a template cannot flatter: the number. What the tools cannot manufacture is the record: what a creator recommended before the brief existed, and whether the audience believed them last time. That evidence sits outside the loop YouTube just built, and it is the one thing a brand should still be reading for itself.

Source: YouTube Official Blog, “Bringing brands and creators together to help you earn more,” Amjad Hanif (VP, Creator Products) and Travis Katz (GM & VP, Shopping), September 23, 2026.

Category: Platforms and pricing / creator economy infrastructure. Confidence: VERIFIED (primary source, fetched this run).

This week carries no item from the creator market. The one candidate repeated ground already covered, so it was cut.

F2. AI-detection scores are becoming institutional verdicts with no accountability attached to the instrument.

[INFERRED] Pangram’s 95%-versus-0% result on the same author this week is not, on its own, proof the tool is broken; detectors are probabilistic and no vendor claims certainty on a single sample. What is new is an institution, the Académie Goncourt, treating a contested score as sufficient grounds for a public, named sanction, ahead of any independent audit, published false-positive rate by dialect or writing tradition, or right of technical reply for the author. If that pattern holds, the detector’s percentage becomes load-bearing in exactly the way a follower count once was: cheap to produce, easy to cite, and only loosely tied to the thing anyone actually wants to know. Would confirm: another public institution, in publishing, hiring, education or elsewhere, cites a named AI-detector’s score as its stated reasoning for a punitive decision within the next quarter. Would kill: institutions begin requiring a published, dialect-broken-out false-positive rate or multi-detector corroboration before acting on a single score.

A ninety-five-point swing on the same author’s prose, depending only on which pages you feed the machine, is this week’s cleanest instance of a pattern running under three unrelated stories. In each one, an institution had a cheap number in hand and a harder question it did not have to answer to act.

The Académie Goncourt did not establish that Thélyson Orélien used AI to write C’était ça ou mourir. It established that a commercial detector scored the manuscript high enough, and that the score was enough. The same detector, pointed at the same writer’s older, undisputed prose, returned zero. Nobody has explained the gap, and the Academy did not wait for an explanation before acting. What got counted was a percentage. What did not get counted was whether a detector trained mostly on the mechanics of contemporary English and French prose has any real purchase on a bilingual, Haitian-Caribbean cadence Orélien and his publisher say is simply his own. The author says the accusation has traveled with racist commentary. That is his account of what the past week has cost him, not culture-watch’s inference, and it is the sharpest version of the general risk: a countable score, deployed against a writer whose actual voice the count was never built to recognize.

The other two stories are the same failure in earlier stages. Six banks correctly named the mechanism by which an AI shopping agent could quietly favor whichever merchant pays it more, then left disclosing that mechanism optional. Reach is what brands pay for. Fit is what pays back. YouTube’s new tools are extremely good at manufacturing the appearance of fit at scale, tailoring every pitch to a brand’s own words, while leaving untouched the one thing that would actually distinguish one creator’s judgment from a template: a record of calls made before anyone was watching.

A number that changes by ninety-five points depending on the sample is not a measurement. It is a coin that landed twice. This week, one institution still spent it as if it were money.

Eighteen calls stand on the record. Zero graded. The first grade lands October 8, when Call 11 matures.

New calls this issue, FINAL:

Call 17. No AI-detection vendor, Pangram or otherwise, will publish a peer-reviewed or independently audited false-positive rate broken out by non-native or regional-dialect prose by 2027-03-31. The gap this week’s case exposed, between a 95% score on a Haitian-Canadian author’s new novel and a 0% score on his own older prose, is exactly the kind of finding a vendor under real pressure to prove its instrument would publish. Checkable against Pangram’s published methodology and independent academic literature on AI-text detection. Due 2027-03-31.

Correction, 28 September 2026: the rationale above inherits the same error. Pangram has published a self-tested false-positive rate for non-native prose (0.012% across 25,021 samples). Call 17 stands as worded: its bar is a peer-reviewed or independently audited rate, and none has been published.

Call 18. None of the six banks’ own AI shopping agents, or a comparable major bank-backed agent, will ship a live, user-facing disclosure of sponsored or commission-based prioritization by 2027-06-30. The five-principle paper names the exact harm and commits to no format, no audit and no deadline for fixing it. Checkable against each named bank’s published agent documentation and product release notes. Due 2027-06-30.

Eighteen calls now open. First grade lands October 8 with Call 11, then Call 13 on October 14, then Call 15 on October 20, then Call 8 on October 31.

#  n | due (ISO)  | status | call
 1 | 2027-03-31 | OPEN | Meta's Partnership Ads mandate will not shift material brand budget toward fit-vetted creators in 2026
 2 | 2027-03-31 | OPEN | No major FTC enforcement action against a creator/brand for undisclosed AI-generated endorsement before Q1 2027
 3 | 2026-12-31 | OPEN | A second major platform ships a user-facing AI-content feed control
 4 | 2027-06-30 | OPEN | A mainstream platform launches creator-set per-use pricing for licensed creator likeness in AI features
 5 | 2027-03-31 | OPEN | A second major studio/IP holder signs a creator-remix licensing deal with a social platform
 6 | 2026-12-31 | OPEN | At least two more national consumer brands face copycat influencer-disclosure class actions
 7 | 2027-08-31 | OPEN | Instagram or TikTok moves its public view metric to a first-frame/served standard
 8 | 2026-10-31 | OPEN | Inaugural CreatorFronts yields at least one announced Fortune 500 upfront-style creator commitment
 9 | 2027-03-31 | OPEN | X's Original Content Rewards does not move brand budget: no top-10 holdco or major agency network publicly announces a dedicated X creator-spend commitment
10 | 2027-06-30 | OPEN | No industry body (IAB or comparable) publishes creator pricing guidance
11 | 2026-10-08 | OPEN | X will not publish a per-creator payout formula for Original Content Rewards by October 8, 2026.
12 | 2027-03-31 | OPEN | A second major retailer joins a platform's in-video shopping-affiliate program by March 31, 2027.
13 | 2026-10-14 | OPEN | TikTok's September 24 Community Guidelines, as published in effect, will contain no policy category for human "ghost creator" persona accounts.
14 | 2027-03-31 | OPEN | No major platform (TikTok, Meta, YouTube, or X) ships a policy category, label, or enforcement rule naming human-fronted persona accounts as distinct from AI-generated content by 2027-03-31.
15 | 2026-10-20 | OPEN | The USTA will publish no creator conduct, etiquette, or credentialing standard tied to its creator cohort by 2026-10-20.
16 | 2027-03-31 | OPEN | The IAB will publish no standard definition of a creator "view" by 2027-03-31.
17 | 2027-03-31 | OPEN | No AI-detection vendor will publish a peer-reviewed or independently audited false-positive rate broken out by non-native or regional-dialect prose by 2027-03-31.
18 | 2027-06-30 | OPEN | None of the six named banks' AI shopping agents will ship a live, user-facing disclosure of sponsored or commission-based prioritization by 2027-06-30.

Culture-Watch Log is culture-watch’s public record: dated calls, graded where everyone can see. culture-watch.com/intelligence