In Defense of Pangram

November 2026

Recently there have been a number of controversies involving the Ai-writing detector Pangram and prize-winning stories.

  • In May 2026, Pangram reported that 100% of Jamir Nazir’s story, “The Serpent in the Grove”, which won the Commonwealth Prize, was AI-written.
  • In November 2026, Vauhini Vara, writing for the Atlantic, used Pangram to point out that winner of the Caine Prize - “the Blue Fighter” by Simbarashe Steyn Kundizeza - seems to be an AI-written story.

In both cases, the outcry has been bizarre. Despite Jamir Nazir’s story having obvious AI tells such as a women who “had the kind of walking that made benches become men” or “smiled like sunrise over a sink” - both prizes doubled down.

Internet commentators cried that this statistical detection of AI writing is somehow racist. There have been claims that AI writes in these stilted patterns precisely because it ingested the writing of black and brown authors from the global south, not that authors have used AI.

Tools like Pangram, rather than being a defense of human writing seemed to have instead been hand-waved away, leaving the field wide open for exploitation.

In the case of the Caine Prize even their statement on Instagram seemed to be AI generated:

Image

Human centred, evidence led, thorough. Multiple lines of evidence, not a single score. Every account aligned.


There are two problems here.

The first is that the people who are rightfully opposed to AI in writing seem to have absolutely no critical engagement with the technologies involved, and therefore no real understanding of current generation capabilities, writing styles or patterns.

We already know that untrained humans are notoriously bad at recognizing AI-generated text. To quote Wikipedia:

Do not rely too much on your own judgment. Humans are notoriously bad at distinguishing human and LLM-generated text. While research on humans’ abilities to detect AI-generated text is still limited, a 2025 study has shown that human ability to distinguish LLM text from human is no better than random chance. Another 2025 study on German theses has shown that humans managed a “recognition rate of 57 % for AI texts and 64 % for human-generated texts”.

In a 2025 preprint, a subset of ‘expert’ annotators correctly identified 90% of AI-generated text (and misidentified 3.3% of human-written text) as AI-generated. Study participants who didn’t use LLMs much did only slightly better than random chance (in both directions).

Note, also, that human speech and writing is being influenced by LLMs, and thus they are becoming more similar. This was already evident in 2024, as shown by a study that detected a significant LLM influence in spoken content (e.g. conversational podcasts). Further studies seem to confirm this influence on language, including semantics and word choices. However, this is a lagging and patchwork process, and the reverse is not true: LLM output is not becoming more similar to that of human writing.

Unless you know AI tells, you tend to be pretty bad at recognizing it.

One would assume that judges of literary awards and literary merit - fundamentally human aspect of writing and the community - would at least be technically competent enough to distinguish machine writing from human, or at least have trained themselves in that regard, but well, one assumes wrong. In fact, the literary world is sleepwalking into a crisis, as Kelsey Piper so forcefully points out.

The second, which complicates the matter, is the fact that writing from the global south seems to be expected to be this tortured, heavy, and quite frankly, bizarre strain of navel gazing. Actual readable writing seems to be left to the Anglosphere.

From the rest of us, it seems the world expects endless meditations using analogies and metaphors that only make sense if you put on some heavy Orientalist glasses and say, oh, but they’re so brown and exotic, their worldviews are so alien. As if somehow we were a different species of humans altogether, as if we did not use the internet or everyday applications, as if we did not understand Microsoft Word, or have access to thesauruses and dictionaries; as if it is simply not possible to convey our social media without doing the literary equivalent of beating up a thesaurus in an empty parking lot.


Let’s revisit the case of Jamir Nazir.

Image

There is a very fine article by Africa as a Country on how to read post-colonial writing, which points out the nature of this otherization and this alien gaze within the literary sphere.

What exactly were the judges thinking when they selected this story? Surely, there were other, more worthy winners. According to Sharma Taylor, the judge for the Caribbean region, the story was selected for the following reasons:

Jamir Nazir’s language is sublime—precise yet richly evocative—conjuring vivid, lush imagery with remarkable economy. Through sharp sensory detail, he renders the Grove as a living presence, where labour, landscape, and memory are intimately entwined. Polished and confident, this is a story with a melodic voice that lingers long after the final line. Jamir Nazir’s prose pulses with a voice of restraint and quiet authority—a beautifully told and assured piece of storytelling.

If these feel like vague and uninspiring platitudes, that’s because they are. It is almost as if someone put postcolonial keywords in a blender and served up the slop. What they reveal is a critical vocabulary that has become entirely decorative—terms like “richly evocative,” “sensory detail,” and “melodic voice” floating free of any engagement with the actual sentences on the page. Nazir’s selection, therefore, is less about him using AI and more about the irony of Euro-American conceptions about the unintelligibility of postcolonial writing. Despite years of literary criticism seeking to undo the idea of the inscrutable Other, here we were. . . .


Let’s revisit the case of Jameer Nazir.

Like, look, this shit is bad writing. It is objectively terrible. And it is AI. You don’t need a pangram to understand how bizarre some of these metaphors and sentence structures are.

Image

In fact, if you do but a cursory search for the man himself, here’s what you come across:

  • A LinkedIn profile with obvious AI written, engagement-farming slop:

    Image

  • Two books on Amazon.in self-published that have a completely different writing style from his award-winning story:

    Image


And the Caine Prize winner?

Image

Much of the anger directed at the Atlantic seems to assume that Vauhini Vara, the author of that article, relied solely on Pangram, but she didn’t. That article actually pointed out the AI tells that were obvious and common in AI-produced writing: the repeated spam of not X but Y sentence constructions; the glut of metaphors; the aphorisms that didn’t make sense, cliched stereotypes and cultural tropes.

Brunella Urbano writing for the Point Mag (translated from Spanish), pointed out the exact same patterns in the award-winning novel Donde el agua trabaja, by Peruvian writer David Silva Rojas. And again, the reaction to that analysis was identical: allegations of racism, mistrust of statistical methods, “a display of ignorance and obscurantism on the part of many intellectuals from our continent.” And Brunella Urbano notes the exact same thing that Vauhini Vara did: this “It’s not X. It’s Y” pattern.

It would be fabulous indeed to believe that all these different writers from different parts of the Global South (remember we are not a monolith), seem to coincidentally write exactly like the state of the art AI models available today. What a fantastic feat of magical realism.

However, one is an interesting anecdote. Two is the start of a suspicious pattern and multiple is a problem.

Using cries of racism to cover up crap like this is in fact racist in itself.

It is the assumption that writers for whom English is a second language cannot be competent enough to write a sentence that makes sense to another human being. It also infantilizes writers, as if we are some sort of strange beings who are not capable of being liars, cheats, thieves or assholes (when in fact a cursory history of interesting writers will show you the very opposite).

Now, no model is 100% accurate. I have a statistics and machine learning background: it’s a very common aphorism in statistical modeling that if your model is very accurate, then all you’ve got is an overfit.

But in my opinion, any fool should be able to see the value in using these tools in an age where increasing volumes of text are generated by large language models with - if you know how they work - have ridiculously common tells. So much so that Wikipedia has an entire section on how to identify signs of AI writing in Wikipedia articles.

If it’s not clear, I’m on the side of Pangram here. People have tested Pangram (medium post), on Twitter. And not just on Twitter or medium post by random people, there have been actual studies on the subject.

In fact let me name a few people who are actually studying this. Tuhin Chakrabaty, an assistant professor at Stony Brook University, has studied books that get flagged by Pangram (that is to say 14,419 self published fiction books sold on Amazon from 2023 to 2026 March: see their paper at https://arxiv.org/pdf/2607.20349) with Paramveer Dhillon of the University of Michigan, Xinyue Liu of Stony Brooke, and Jane Ginsberg at Columbia Law.

He also ran [(]one hundred and eighty six African novels through Pangram]([)https://tuhinchakrabarty.substack.com/p/are-ai-detectors-really-racist-or). You won’t know what he found? Pangram isn’t racist towards African writing. There’s zero indicator that Pancram sees African writing and starts screaming.

Pengram is in fact very good at detecting AI slop. If something shows up as 2% AI on Pangram, that can be dismissed as statistical noise if something registers as a hundred percent - well that’s time to sit up and pay attention.

Critically, despite Pangram having a open bounty for pre-LLM writing that detects as AI, nobody has so far managed to find such a thing. The only thing so far that the Internet has dredged up is a joke poem where the same 6 words are repeated four thousand times.

Meanwhile, on the other side - aka “pangram is evil” - the best we seem to have were Timnit Gebru’s “trust me, bro,” claims that she has spoken to students who’ve said they’ve been unfairly flagged for AI writing.

The slightly more coherent counterclaim is a strawman argument: saying statistical tools like this have been pretty bad, and some have even flagged the US Constitution as AI written. Indeed. But are they the same statistical tool? Is a hammer like a spade?

Look, let me do a Gebru. I know my shit on this. In fact, since 2019, I’ve been training machine learning models, I’ve been building my own large language models, and I run AI locally for programming, data extraction and policy research. Hell, I think OpenAI and Anthropic are unambiguously evil, but I’m a massive advocate of open source AI that everyone has access to.

Hell, I’ve done this before. I have a post on this site (Can You Vibe-Code A Novel) in which I specifically had Claude generate a 50,000+ word “novel” just to evaluate how far the public interfaces for these tools have come.

I even have a series of novels, The Salvage Crew series, where I deliberately set myself the challenge of using local AI in interestingly oddball ways that I explain in needless detail in the forewords and afterwards. I don’t pretend it’s high literature, but critically, I’m not trying to fool people. If I can figure out how to use a tool in an unconventional way, I’m going to talk about it.

TL;DR I cannot tell you how many hundreds of thousands of chats I have seen, the exact same type of language that these so-called literary award-winning writers demonstrate.

So trust me, bro.

If this is what wins awards, then we should perhaps shut the awards down and go home - because Claude and ChatGPT and Quen 3.8 flash next and GLM 5.3 and DeepSeek 4.1 and whatever else I’ve forgotten to mention here -

they’ve already won.


But let’s explore this racism allegation a bit further. A critical part of the racism allegation is that it is anti-black and or anti-African.

James Yu, who makes money by selling AI software that specifically writes stories for authors, has been keeping track of competitions where AI use was suspected: https://jamesyu.org/award-winning-ai-literature/.

Now, I assume you does this because it is relevant to his business somehow. I find that particular business banal, but I’m going to appropriate his table to point out something. Look at the list of countries:

CountryWorkCreatorPrizeStatusNotes
🇰🇷 South KoreaFiction: Unnamed winning entryUnnamed11th Aviation Literature Award, Adult Fiction Excellence Award, 2025Organizers or publishers said AI was usedA prompt and an apparent AI reply were still in the story when it appeared in the winners’ anthology, according to ministry records released by a South Korean lawmaker. A reader filed a complaint. The transport ministry revoked the award in August 2026 and moved to recover the prizes.
🇫🇷 FranceNovel: C’était ça ou mourirThélyson OrélienPrix du Roman Fnac, 2026Organizers or publishers said AI was usedThe novel won Fnac’s prize, then was removed from the Goncourt longlist after AI allegations. On October 21, Boréal dropped Orélien after an external review reportedly found AI markers in working-document metadata. The publisher said it still did not know how much AI had been used. Orélien had denied that any sentence was machine-generated.
🇸🇬 SingaporePoems: Migrant Writers of Singapore poetry competitionEight winners (not named)Migrant Literature Festival, 2024Organizers or publishers said AI was usedEight winners were announced, then the entire list was recalled after organizers said three winning entries were AI-generated. Their review relied partly on Scribbr and QuillBot.
🇺🇸 United StatesPoems: Poetry.com November and December contestsNot reliably identifiable (removed entrants)Two Monthly First Prizes, 2024Organizers or publishers said AI was usedPoetry.com said both original winners were created “with assistance” from AI, changed the results, and redistributed the prizes. The removed poems and writers are no longer reliably identifiable. The names on the current result pages are the replacements, not the accused entrants.
🇫🇷 FranceNovel: L’Inconnue du quai de JavelPhilippe JaenadaPrix Goncourt, 2026Confirmed AI useAfter winning, Jaenada said he had turned to AI three times while working on the novel. On RTL, he revealed the half sentence from ChatGPT he had deliberately hidden in the book. ChatGPT had offered to invent a cliché. RTL interview
🇯🇵 JapanNovel: Sympathy Tower TokyoRie Kudan170th Akutagawa Prize, 2024Confirmed AI useKudan first guessed that about 5% of the novel involved ChatGPT, then gave a much smaller figure: about one page out of 143. She consulted its output for the replies written by the AI inside the novel. The prize stood.
🇨🇳 ChinaShort story: The Land of Machine MemoriesShen Yang, writing as “QuartZen”5th Youth Popular Science and Science Fiction Competition, Second Prize, 2023Confirmed AI useAI generated the title, prose, illustrations, and pen name over 66 rounds of prompting. The organizer invited Shen and hid the experiment from the judges. This is often called a national literary prize, but it was a broad Jiangsu association contest: 18 stories shared second prize and 90 of roughly 200 entries received awards.
🇯🇵 JapanWeb novel: Modest Skill “Tidying Up” Is the Strongest!Not named18th AlphaPolis Fantasy Novel Awards, Grand Prize and Reader’s Choice, 2025Confirmed AI useThe author said a large majority of the web novel was made with generative AI. AlphaPolis later canceled the print edition and manga adaptation and changed its rules. Reporting confirms those cancellations, but not that the two award labels were formally revoked.
🇺🇸 United StatesMemoir: How to Win One Million Dollars and $#!T Glitter!Luke StoffelIBPA Book Award, Benjamin Franklin Silver Medal for Neurodivergent Communities, 2026Confirmed AI useStoffel says he used ChatGPT for grammar and structural scaffolding and disclosed it in the book and award submission. IBPA’s guidelines did not prohibit AI-assisted work, and the book received Silver. IBPA has not publicly commented on his AI use.
🇵🇪 PeruNovella: Donde el agua trabajaDavid Silva RojasAlmadía–Ventosa Arrufat International Novella Prize, 2026Suspected or disputedEl País reported that Pangram classified 99% of the text as AI-generated. Silva Rojas denied using AI. Almadía reviewed materials from his writing process and consulted specialists. On November 6, the publisher upheld the €20,000 prize and kept its review documents private. Author’s statement
🇿🇼 ZimbabweShort story: The Blue FighterSimbarashe Steyn KundizezaCaine Prize for African Writing, Winner, 2026Suspected or disputedPangram flagged almost the entire story as AI-generated, according to The Atlantic. Kundizeza denied using AI. On November 11, the prize upheld his award after interviews and independent forensic analysis of draft metadata, saying it found no evidence of generative AI use. Pangram cofounder Max Spero then alleged that the prize’s Instagram statement clearing him also appeared AI-generated. That claim concerns the statement, not new evidence about the story.
🇬🇧 UKShort story: The Serpent in the GroveJamir NazirCommonwealth Short Story Prize, Overall Winner, 2026Suspected or disputedPangram called it 100% AI-generated. Nazir denied it and described six or seven drafts made with Android speech-to-text because of a chronic condition. The Commonwealth Foundation reviewed drafts, timestamps, notes, and testimony, accepted his account, and kept the award. The allegation is disputed, not proven.
🇬🇧 UKShort story: The Bastion’s ShadowJohn Edward DeMicoliCommonwealth Short Story Prize, Canada and Europe Winner, 2026Suspected or disputedPangram also returned 100% here. DeMicoli produced handwritten drafts, notes, and years of research and denied using AI. The Foundation accepted the evidence and the prize stood.
🇬🇧 UKShort story: Mehendi NightsSharon AruparayilCommonwealth Short Story Prize, Asia Winner, 2026Suspected or disputedAruparayil denied using AI at any stage and kept a paper trail. She also found that adding three Marathi words and removing one word made a flagged section register as 100% human. The Foundation kept the award.
🇬🇧 UKShort story: DescendChanel SutherlandCommonwealth Short Story Prize, Overall Winner, 2025Suspected or disputedPangram scanned the story a year later and labeled it 88% AI-generated. Sutherland denied using AI and said its repetition and rhythm come from Vincentian oral storytelling. Her publisher and agent backed her. No public provenance review of the 2025 story has been released.
🇬🇧 UKShort story: Back and ForthKavyta KayHarper’s Bazaar UK Short Story Competition, Winner, 2026Suspected or disputedA screenshot of a 100% Pangram result spread online. No author response, organizer investigation, admission, or revocation has surfaced.
🇬🇧 UKPoem: Peter Doolan and an unnamed winnerPeter Doolan and an unnamed winnerPlaza Prizes, 2025Suspected or disputedThe Plaza disqualified both after detector flags while admitting it could not know how much AI had been used, “if any.” Both denied it. Doolan said his poem had already been published in 2018. This looks less like an AI scandal than a detector-driven procedural failure.
🇺🇸 United StatesComics anthology: Stardust the Super Wizard AnthologyEditor and contributors (not named)Eisner Award for Best Anthology, Nominee, 2026Nominations and shortlists withdrawnOne page was an AI-generated comic credited to the invented “Alex Irons.” The editor and other contributors reportedly did not know. After Comic-Con said the judges would not have nominated it had they known, the editor withdrew the anthology from the ballot.
🇧🇷 BrazilIllustrated edition: Frankenstein, Clube de Literatura Clássica editionVicente Pessôa (designer) and MidjourneyPrêmio Jabuti Illustration Category, Finalist, 2023Nominations and shortlists withdrawnThe book openly credited designer Vicente Pessôa and Midjourney and was marketed as fully AI-illustrated. Two judges said they did not realize Midjourney was AI. The edition was disqualified under rules that had never anticipated generative art.
🇬🇧 UKIllustrated book: Unnamed illustrated titleNot namedInk Book Prize, Shortlisted, 2026Nominations and shortlists withdrawnThe prize removed one title after complaints about AI in its illustrations. A different shortlistee withdrew in protest. Comparing the old and new lists suggests possible titles, but neither the organizers nor the reporting named the accused book, so I won’t either.
🇳🇿 New ZealandNovellas + stories: Angel Train and Obligate CarnivoreElizabeth Smither and Stephanie JohnsonOckham New Zealand Book Awards, 2026Nominations and shortlists withdrawnBoth books were initially barred because their publisher used AI-generated covers, then reinstated after the awards trust decided the rule had arrived too late. Neither reached the longlist.

South Korea, UK, Brazil, New Zealand, Japan. You have to really, really ignore the evidence and try hard to claim that this is somehow a phenomenon unique to black co-African authors.


Look, I find this whole thing ironic.

In 2018, I wrote a story for Brian Walton called The Writing Contest. In it, a poor writer enters a writing contest. The writer desperately needs money for his mother’s care. In a spate of uncertainty, he spins up an AI model and enters like a dozen more entries. And he is utterly crushed when it is the AI-written story that wins.

This is the future that we are in; this is the future that we have been heading into for some time. What I find extremely unpalatable is that the people who desperately want to be guardians of certified organic human written literature are the people least competent at detecting it.

In 2019 I co-founded a fact checker and grew that into a data journalism lab. One of the things we learned the very hard way was Brandolini’s law: it takes an order of magnitude more effort to refute bullshit than to produce it.

Whether you like it or not, the problem is at a scale and at a volume that simply cannot be dealt with by people who aren’t experts at both human and machine constructions.

I honestly think it’s about high time that we have more Pangrams. Perhaps each literary prize should in fact concoct, if not a statistical tool, at least a tree of decisions that they can follow to understand whether something is human or not.

Perhaps they should conduct their own independent audits on tools. Perhaps they should have not one, but many, similar to how universities use automated tools to flag plagiarism. Perhaps these tools should be private, so as to prevent future LLMs from learning to game them.

It won’t last long. As I always say, if you think you can detect AI writing, that’s because you’re looking at unsophisticated users. But we do collectively need to start holding ourselves to standards if we are to be serious about this whole business of prizes and literary merit.

Without it, all of this is tosh. Correction: at best it’s tosh. At worst it’s just Orientalism with extra steps.


Back to top

To get in touch, reach out to hello [at] yudhanjaya [dot] com, or contact finegan [at] zenoagency [dot] com to talk to my agent (Stevie Finegan). Responses from my end may take time.

/ search
rss · substack