The AI label breeds skepticism — but it doesn’t automatically make texts untrustworthy.

Publishers often justify their reluctance to label AI-generated content by arguing that readers trust labeled text less. This sounds like a clear-cut finding. But the evidence base is narrower than the argument implies. What can be demonstrated is that many people feel uncomfortable when AI writes a text largely on its own. What cannot be demonstrated is that the same text is therefore consistently judged to be inferior, or that any label automatically destroys trust.

What surveys actually measure

The Reuters Institute’s Digital News Report 2024 does not ask whether an article is credible. It asks how comfortable respondents feel — measuring a mood, not a quality judgment. Across 28 markets, only 19 percent were comfortable consuming news produced mainly by AI with some human oversight. For content produced mainly by journalists with AI assistance, that figure rose to 36 percent. Discomfort was stronger in Europe than in the United States. In the United Kingdom, only 10 percent felt at ease with barely supervised AI use.

Amy Ross Arguedas, the author of the qualitative sub-study, describes the starting point this way: „Most people have barely thought about how AI could be used in journalism. Their default position is mostly rejection, distrust and concern.“

Discomfort grows with the visibility of AI’s role. Behind the scenes — transcription, translation, research assistance — attitudes are milder. Publicly generated texts attract more reservations. Images and video are seen as more sensitive than text; politics and crime as more sensitive than sport or entertainment. This is graduated skepticism, not a blanket verdict of „AI text equals worthless.“

The interviews also reveal that transparency is demanded but understood differently. A 40-year-old participant from the United Kingdom said he would probably not read an article labeled as an AI product. A 26-year-old from the same country considered disclosure unnecessary as long as a human continued working with these tools behind the scenes. A 28-year-old from the United States, however, wanted a label every time, so readers could decide for themselves.

What happens when readers evaluate the text themselves

A preprint study by Gilardi and colleagues, involving 599 participants in German-speaking Switzerland, picks up from there. Participants read short excerpts on Swiss politics — some journalistically written, some rewritten by ChatGPT, some fully generated by AI. Before the source was disclosed, credibility, readability and perceived expertise were virtually indistinguishable across the three groups. Equivalence tests support this: the differences were small enough to be considered equivalent. The authors write that AI-assisted or fully AI-generated news texts keep pace with traditional journalistic articles in terms of perceived quality.

After participants learned how each text had been produced, those who discovered AI involvement — either assisted or fully generated — were actually more inclined to keep reading (effect sizes of approximately d = 0.53 and 0.61 respectively). The general willingness to read AI-produced news in future, however, remained low and did not differ clearly between groups. The study itself cautions against over-interpretation, noting that the willingness to read on probably reflects curiosity-driven interest rather than approval or long-term acceptance.

Without a label, a text often registers as journalistic. With a label, it is not necessarily the quality rating that changes — it is the attitude toward the source.

Where the idea of a trust penalty comes from

Publishers speak of a trust loss that a label automatically incurs — a kind of surcharge levied on labeled text. This impression comes from experiments where the label precedes the judgment. The 2024 Digital News Report already points to early evidence that news identified as AI-generated is rated as less trustworthy than human-produced content.

More recent work sharpens the picture without unifying it. Prajod and colleagues found in a small experiment that a brief disclosure left trust scores close to those with no disclosure at all. A detailed disclosure lowered trust ratings. Both versions led readers to check sources more frequently. Questionnaire responses and subscription rates showed a trust decline only with detailed disclosure. Not every disclosure therefore creates a transparency dilemma — but a detailed one can.

A package combining a detailed production description, human oversight and an error-reporting channel did not prevent the decline. Whether human oversight alone cancels the effect cannot be cleanly derived from this. The study worked with just 40 participants and a fictional outlet. The topic and brand were also left unspecified. In the same study, the detailed disclosure had a stronger effect on lifestyle topics than on politics. Other studies cited there report credibility losses for both political and non-political headlines, but a partial penalty only for politics. An established media logo was not tested; the authors explicitly identify brand logos as an open variable.

With a few dozen participants, a debate can be sharpened. A rule for journalism as a whole cannot be derived.

What happens when there is no label and suspicion sets in later

The labeling question has a flip side that rarely appears in the studies: what happens when nothing is labeled and suspicion emerges afterward? There are no official statistics on this — the industry does not track such cases centrally. But a number of cases from the publishing world over the past 18 months are documented.

Shy Girl by Mia Ballard. The horror novel was published in November 2025 by Hachette UK and reportedly sold around 1,800 copies. Readers first grew suspicious on Reddit and Goodreads, citing repetitive phrasing and typical AI stylistic patterns. The New York Times then presented evidence suggesting the text was AI-generated. In March 2026, Hachette withdrew the novel entirely and halted a planned US edition. Ballard denies the allegation and says only a friend who edited the manuscript used AI.

Call Me, I’ll Hide the Body by Jerry Falade. The debut crime novel was sold at auction in a 14-way bidding war and reportedly fetched a very large sum, going to Minotaur/Macmillan. In July 2026, Falade’s own literary agencies withdrew the manuscript, saying they could no longer authenticate that it had been written without AI. The publisher parted ways with him. Falade acknowledges using AI for research but not on the manuscript itself.

C’était ça ou mourir by Thélyson Orélien. The debut novel by the Canadian-Haitian author was removed from the Prix Goncourt longlist in September 2026. Prior analyses, including by the detector Pangram, had suggested the text was largely the product of AI. The book remained in print; only the nomination was withdrawn. The publisher has pushed back, pointing out that AI detectors have in some cases flagged even the Bible as AI-generated.

Commonwealth Short Story Prize 2026. This is a special case. Three of five regional winners were accused after the ceremony of having used AI, triggered by Pangram analyses. The Commonwealth Foundation investigated and concluded that no AI had been used. The prizes stood. The timeline for the 2027 prize was reportedly delayed as a result of the episode.

The number of officially withdrawn novels is small. The true figure is likely higher. Publishers Lunch reports that several imprints at major publishers have quietly canceled contracts over AI suspicions, often involving self-published titles that were later acquired. At the same time, the volume of undetected or undetectable AI-generated literature is growing. One tracking study puts the share of new e-books with no detectable AI content at nearly 100 percent at the start of 2023, falling to around 60 percent by mid-2026. Kobo CEO Michael Tamblyn has described the volume of AI submissions as a „firehose.“

What is striking is what the debate now hinges on. In the case of Shy Girl, the text was reportedly criticized for flat prose and awkward metaphors — and it had already found buyers before anyone noticed anything. The question is no longer whether AI can produce convincing texts, but whether that can be proven. The primary evidence used is detectors like Pangram, which are themselves criticized for false positives, yet in the Goncourt case were treated as delivering a verdict.

The Commonwealth case illustrates what this means for authors who wrote without AI. Literary scholar Johannes Franzen, speaking on SWR’s Kultur am Morgen program, put the Orélien case in context. The starting point, he explained, was an anonymous account on X claiming the novel was AI-generated. Nothing has been proven; the author rejects the allegations; yet the Académie Goncourt withdrew the nomination regardless.

Franzen declines to render a firm judgment on whether the Goncourt Academy acted rightly, given how fiercely contested the allegations are. What is certain, he says, is that the accusation alone has already done significant damage to the author’s career. This reflects an „extremely nervous age.“ The peak of this feverish atmosphere of suspicion may have been reached — or things could still get worse.

On detectors, Franzen is cautious. They are by no means infallible; the technology is still in its early stages; the research is ongoing. That said, Pangram has recently been quite successful at exposing such texts. Nothing can be said for certain, but „the evidence does point fairly strongly in that direction.“

On the reader side, Franzen describes a reaction that runs in two directions. Some respond with revulsion, feeling deceived because they loved a text that now turns out to have been machine-made. Others reject the accusation, unwilling to appear as incompetent readers who were taken in by a machine. What is at stake, then, is not only dishonest authors but also readers‘ sense of their own judgment.

At the same time, Franzen points out that literary quality is in principle independent of how a text came about. It is a special status conferred on a text when it is perceived as particularly accomplished — the text must stand on its own. And yet something in us resists the idea that a machine could write a literary text.

Orélien had, before the allegations, been explicitly praised as high-quality and literary by critics, the publishing world and readers on platforms like Goodreads. A text of that kind, Franzen notes, makes people react all the more sensitively when something seems to be wrong. He admits to having been taken in by an AI text himself, with all the accompanying emotions. By now, almost every text provokes suspicion — one could almost speak of a general paranoia. This will normalize over time, he suggests, depending on how the relationship with AI develops.

He also recognizes the „ick“ factor: some readers believe they can detect AI while reading and feel disgust. Whether people will develop a lasting instinct for this, whether AI will eventually be able to generate any kind of text, or whether we will ultimately become indifferent — these questions he leaves open.

For the labeling question, a conclusion follows: foregoing a label does not protect against a loss of trust. It merely shifts that loss to the moment when suspicion arises — and when it does, it hits not just the source but the book and the people behind it.

Three levels that need to be kept separate

The findings address three distinct questions. The first is attitude: many people do not want news that comes mainly from AI. That is what the Digital News Report shows. The second is text quality: without a disclosure of origin, current language model outputs are often rated similarly to journalistic texts. That is what Gilardi et al. show. The third is the label effect: the phrase „AI-generated“ can reduce trust, independently of the text itself. How much depends on phrasing, context and the specific study. There is no universal figure for this discount.

Gilardi et al. describe this decoupling themselves: readers acknowledge AI articles as comparably well written but remain reserved about their origin. The argument „we don’t label because readers will trust us less“ therefore only holds up halfway. The skepticism is real. The conclusion that labeling is fundamentally harmful is empirically too thin.

Anyone who wants to create transparency while minimizing the bluntest form of label risk should not choose „written entirely by AI“ as a standard formula. More precise is specifying which task the AI performed, who reviewed the content and who bears responsibility. Whether this reliably neutralizes the trust discount is an open question. Whether silence solves the problem is equally open. The cases from the book market suggest it does not.

Silence protects short-term perceptions and erodes in the long term what publishers and journalism live on: the assumption that the stated origin of a text is accurate.

Instead of asking whether readers trust AI texts less, a more precise question is: under which label, at what level of human responsibility and on which topic does which judgment change — credibility, fact-checking behavior or reading interest? Until sufficiently large, repeated experiments are available, blanket rejection remains an understandable assumption, not an established rule.

What this means for the literary world

My thesis: the literary world damages itself not by failing to celebrate stasis and rejection as the preservation of art, but by refusing to acknowledge progress. In doing so, it awards itself a seal of authenticity that will not hold for long. AI is everywhere, and those who reject it wholesale play no part in shaping the future.

The words attributed to Mikhail Gorbachev — „Those who are late will be punished by life“ — apply here too. Those who refuse to engage with change will find themselves left behind.

Culture is subject to the same transformation as every other domain of life. It is entirely understandable not to want to throw it to the machines. But it would be equally wrong not to engage openly and freely with the advantages on offer. No one should have to feel disgust — and it is not well supported by evidence in any case. Surveys measure discomfort and comfort, and in the Swiss study, readers were actually more inclined to keep reading a text after learning it had AI involvement.

Nobody, incidentally, has ever recoiled from a Thermomix-cooked goulash because the machine stirred, weighed and regulated the temperature. You cannot see the machine in the food, and no one asks whether a human made it — because it is obvious who chose the recipe, seasoned the dish and invited the guests. Perhaps the disgust around AI-generated texts has less to do with the machine than with the unresolved question of who ultimately stands behind it.

How fragile these detection tools are was illustrated by a small but telling recent episode. According to reports on social media, several AI text-detection programs flagged passages from Mary Shelley’s Frankenstein, published in 1818, as fully or partially AI-generated. A novel written two centuries before ChatGPT fails the test. A seal of authenticity that relies on suspicion and detectors ultimately checks nothing that can be checked with any certainty.

This English version was created with the help of AI (translated from the German original) and may differ slightly from a professional human translation.