Photography: cottonbro studio
When Michelin sent Fond its new Dallas description to check before publication, co-owner Brandon Moore read it and recognised the voice at once. “When they sent it to us, I thought, ‘This is clearly generated.’” A D Magazine investigation published on 16 October reached the same conclusion about six of the Guide’s Dallas texts. Michelin says no AI was used to write or edit them. Every kitchen in the Guide now has reason to ask who wrote the words beneath its name.
Highlights
- D Magazine’s dining critic, Brian Reinhart, ran six of Michelin’s Dallas descriptions through the detector Pangram. It flagged 25 of their 31 sentences as AI-generated.
- Two Carnegie Mellon researchers traced the style to OpenAI’s GPT-4o. One of them is the critic’s brother, a link D Magazine discloses.
- Michelin’s 2026 texts for Austin, Houston, New York and Edinburgh, checked the same way, showed no such signs.
- Michelin says its notes come from inspectors’ firsthand dining experiences and that AI plays no part in writing or editing them.
- Nothing suggests the inspectors skipped the meals. The dispute is over how those meals became prose.
Six texts from one city
On 8 October Michelin added five Dallas-area restaurants to its Texas selection: Fond, Jashan, Mābo, Olōyō and Resident Taqueria. It also rewrote its note on Sushi Kozy. That rewritten text was the first to catch Reinhart’s attention, and Pangram, a commercial detection tool, judged it entirely machine-written.
Run together, the six texts came to about 650 words. Pangram marked 25 of the 31 sentences as AI. The exceptions are telling. A list of tacos at Resident Taqueria and the remarks on drinks at Jashan and Mābo read as human, the sort of detail someone writes down at the table.
Reinhart then took the texts to Carnegie Mellon University. Alex Reinhart, who teaches statistics, and David Brown, who teaches English, have built a model that recognises the grammatical habits of large language models, drawing on research they published in PNAS in 2025. D Magazine states plainly that Alex Reinhart is the critic’s brother.
Their model pointed to GPT-4o. The Dallas notes use nominalisations, nouns built from verbs, at roughly twice the human rate, and lean on lists and stacked adjectives. “Thoughtful” appears three times. The researchers say no single habit proves anything, but that across these six texts the evidence is overwhelming.
The comparison carries most weight. Michelin’s Dallas notes from 2024 and 2025, and its 2026 writing on Austin, Houston, San Antonio, Santa Fe, Las Vegas, New York and Edinburgh, showed no sign of AI to Pangram or to the reporter. Whatever happened appears to have happened in one city, in one update.
© seventyfour via Canva.com
© darrenbaker via Canva.com
What Michelin says
The Guide’s statement is unequivocal: all inspector notes on its website and app “are developed collaboratively by the Inspection team based on their firsthand dining experiences. AI is not used in the writing of these reviews.” Asked again, Michelin added that AI is not used in editing either, and that its restaurant text editor belongs to its own editorial team.
It declined requests to interview the Guide’s leadership and declined to review the evidence before publication. According to Inside AI, it also would not say whether it checks texts with detection tools internally. And it made a point about vocabulary: Michelin does not usually call these texts reviews. It prefers “inspector notes” or “restaurant descriptions”.
The distinction is fair, and it changes little. Diners read the paragraph beneath a restaurant’s name as the inspector’s verdict, whatever the Guide calls it. Restaurants quote it on their websites and menus. Chefs read it as an account of their cooking by the only people sent to judge it.
The kitchen on the other side
Before publication, Michelin sends each restaurant a short sample of its text so the facts can be checked. That is how Fond’s chef-owners, Jennie Kelley and Brandon Moore, came to see theirs early. “You can tell,” Moore told D Magazine. The draft also carried a glitch, a wide gap in the middle of the word “Detroit”.
For a young restaurant, a first Michelin listing is a career moment. The text gets framed, shared and quoted in press releases. More and more, it is also read back to diners by the AI assistants they use to choose where to eat: D Magazine met a Fond guest who had found the restaurant through ChatGPT. Generic praise for careful technique does little for the cooking, whoever wrote it. When a chef suspects it came from software, it undermines what the Guide sells, which is the judgement of someone who was there.
© Filipp Romanovski via Canva.com
© MelanieMaya via Canva.com
Where the line sits
Nobody is suggesting the inspectors skipped Dallas. Reinhart notes that they need receipts to be reimbursed, so missing meals would probably have been noticed. The researchers’ own theory is gentler than the headline: an inspector may have asked an AI tool to tidy their notes, or to summarise existing reviews, and then added personal details.
None of this makes Michelin unusual. Across publishing, from news agencies to food magazines, AI now helps with research, transcription, translation and editing. The serious ones have drawn a line. When the Associated Press published its first rules in 2023, it encouraged staff to learn the tools but barred them from using generative AI to produce publishable content, and told journalists to treat anything it produced as “unvetted source material”. The tool can help with the words. The judgement, and the responsibility for it, stays with the person who did the work. For a guide that sells the experience of a meal, that line matters more than for anyone.
That is where the real argument lies. Software that organises an inspector’s observations leaves the judgement intact. Software that supplies the adjectives, the sense of a sauce’s balance or the verdict on a dish has written about food the inspector tasted and the machine did not. Detection tools cannot say which of those happened. They can also be wrong, which is why the researchers rest their case on the pattern rather than on a single score.
Michelin’s answer is that neither happened. It has not described how its texts are produced in enough detail for readers to judge that for themselves.
What a note should carry
A good inspector’s note does a few things well. It names what is distinctive about a kitchen, records something only a person at the table could notice, and gives the reader a reason to book that no other restaurant could supply. A line such as “the beurre blanc is held back so the langoustine leads” could only come from someone who ate the dish. “Thoughtful combinations and bold flavours” could describe a thousand restaurants, and that is the trouble.
Generic language is not proof of software; tired human writers produce it too. But it is the clearest sign that observation has gone missing, and observation is the only thing a guide has to sell. For chefs, the test is this: if the paragraph could move to another restaurant without changing a word, it was not written about yours.
If your description arrives
Michelin sends restaurants a short version of their text before publication so the facts can be checked. It is the one moment a chef has any say, and it is worth using well.
- Read it against the menu you served this year. The dishes, ingredients and techniques should be ones your kitchen recognises.
- Correct every factual error. Names, places and spelling matter. Fond’s draft arrived with a gap in the middle of “Detroit”.
- Flag anything generic. Michelin describes the check as factual, but pointing out that a phrase could apply to any restaurant is a fair request.
- Keep the version you were sent. D Magazine saved copies of the Dallas texts in case they changed. A dated copy protects you too.
- If it reads as though no one ate there, ask. A guide that stands by its inspectors should be able to say who wrote the words.
A question of trust
Michelin’s authority rests on an arrangement chefs accept and diners take on trust: anonymous inspectors, private deliberation, little published method. Its international director, Gwendal Poullennec, has argued that AI strengthens the case for that model. “With AI, it comes together with a lot of fake information as well,” he told The Staff Canteen in 2025, adding that Michelin could “safely continue to work and carry on with the human-based approach.”
“But a secretive process only works if it is founded on trust.”
Brian Reinhart, dining critic, D Magazine
Chefs MAG’s view is clear. An inspector tidying their notes with software is not the scandal. Publishing words that read as judgement when the judgement is missing is. A guide can keep its inspectors anonymous and still be open about how a meal becomes a paragraph. Michelin has said what it does not do. It should now tell chefs what it does.
That matters most on the other side of the pass. A listing can change a restaurant’s year. The chefs who cook for the Guide’s inspectors are owed a description written by someone who sat at their table, and the confidence of knowing it was.
© seventyfour via Canva.com
Sources & Further Reading
- Brian Reinhart, “Michelin Used AI to Write Its Guide to Dallas Restaurants”, D Magazine, 16 October 2026.
- Alex Reinhart, David West Brown et al., “Do LLMs write like humans? Variation in grammatical and rhetorical styles”, Proceedings of the National Academy of Sciences, 2025. Preprint · Carnegie Mellon summary
- Sophia Andreou, “Michelin Denies AI-Written Dallas Restaurant Reviews After Investigation Flags 80% of Descriptions”, Inside AI, 16 October 2026.
- “How Michelin inspectors are navigating AI to remain anonymous”, The Staff Canteen, 22 March 2025.
- Associated Press, standards on the use of generative AI, 16 August 2023.
