No, Michelin Did Not Use AI to Write Its Guide to Dallas Restaurants. D Magazine Simply Doesn’t Understand AI

by Andrew Chalk

D Magazine’s online food blog, Sidedish, contains an article entitled Michelin Used AI to Write Its Guide To Dallas Restaurants by Brian Reinhart, its restaurant critic. In the article he alleges that “most or all of  Michelin’s new Dallas entries were written by large language model (LLM) software, commonly known as AI.” He based the allegation on the results of feeding the text of Michelin reviews through software called Pangram that detects AI-written text.  

Unfortunately, Reinhart misunderstands what Pangram does and how AI can be used, so his conclusion is unfounded. To see this, consider the article Restaurant Review: Mercantile, Denver. Every paragraph of this plodding prose is mine. Cut-and-paste the text into Pangram and it instantly concludes, correctly, that this text is “Human Written". Just for grins, I cut and pasted it into chatGPT (a popular generative AI tool) and prompted it “Below is the text of my review of Denver restaurant Mercantile . Rewrite it to improve it.” chatGPT did so and I was so amazed at the results that I published them here. The text is so much easier and enjoyable to read. I wish I could write like that! 

When I ran chatGPT's output through Pangram it concluded that AI was detected. 

However, it is important to look at precisely what AI has been used for here. None of the facts are new or changed. None of the analysis of Mercantile is revised. Only the style is changed. 

AI did not write my review, it restyled it. Alternatively, one might say it polished it (there may be a term-of-art already in use among writers expressing that this is an example of using AI that does not write the article, it just improves the style). The article, specifically the substance of the article, is the facts and analysis. And they are mine, not chatGPT's.

In a similar vein, I ran Reinhart’s D Magazine article “Already Armed with a Michelin Rec, Olōyō Is Off to a Hot Start. Could It Get Even Better?” through Pangram. It correctly concluded it was human. I then cut and pasted the article into chatGPT, as I had done with mine, gave it the same prompt, and cut and pasted the results into Pangram which, as in the case of my article, detected AI. chatGPT had not changed any facts or analysis, it had only changed the style. 

AI has multiple uses, and one is certainly outright plagiarism. That is a concern, particularly in schools where faculty and parents do not want AI to replace, rather than augment, learning. On the other hand, it is often an advantage to have the better-written work of AI as the final product. Just as we prefer printed, rather than handwritten, final product. 

So Reinhart’s title is unfounded, based on the evidence. But so is mine (I deliberately juxtaposed a total contradiction to emphasize the point). The truth is, we don’t know, when we see a document that is detected as partly AI, whether the structure, rather than just the style, was changed in its creation. We have to ask if we trust the source (and Reinhart’s discovery, while astonishing to him, does not change the Michelin Guide’s stature in the culinary world).  

This means that Reinhart’s allegation of Michelin’s methods “I would have asked Michelin how it would respond to a chef who feels disrespected or insulted because their hard work was reviewed by AI. ” is a misnomer, as he has no proof that any chef’s work was reviewed by AI. He has only shown (via Pangram) that AI was used, and not that it was used to create any substance, rather than just style. 

Enjoy your future with AI, however you use it!

[The above was entirely written by a human]