D Magazine, in an analysis released Friday, accused Michelin of using artificial intelligence to write the Dallas entries in its latest Texas dining guide.
The French tire company updated its Texas Guide on Oct. 8, adding five new restaurants in the Dallas area.
Based on feedback from detection software and academic experts, D Magazine alleges that most or all of Michelin’s new Dallas entries were written by a large language model (LLM).
For its part, Michelin denied the allegation.
“All Inspector notes appearing on the MICHELIN Guide website and app are developed collaboratively by the Inspection team based on their firsthand dining experiences,” Michelin said in a statement to the magazine. “AI is not used in the writing of these reviews.”
The Current reached out to Michelin for comment but didn’t receive a response by press time.
“It’s worth noting that the MICHELIN Guide has its own editorial team of which the restaurant text editor is a part,” Michelin added after follow-up questions from D Magazine. The company clarified that it doesn’t use AI in the editing process either.
D Magazine began the investigation before the Oct. 8 announcement, prompted by a text from a local chef to the reporter that included an updated description of a restaurant and the question, “Do you think this is AI generated?”
The reporter, Brian Reinhart, pasted the text into an AI detector. Although AI detectors are known for producing false positives, Reinhart chose Pangram, which boasts a high accuracy rate and a low rate of false positives.
It returned results indicating the text was 100% AI-generated.
Once the updated guide came out, Reinhart ran the Dallas entries through the detector, which determined that the text was 80% AI-generated.
Reinhart compared the 2026 Dallas entries to those of San Antonio, Houston, Austin and other cities — all of which turned up 0% AI-generated. Dallas entries from prior years also registered as 0% AI generated.
Beyond the online check, D Magazine consulted with professors at Carnegie Mellon University who authored a paper on determining AI authorship. The scholars, an English professor and statistics professor, determined that the model used to produce the restaurant writeups was likely GPT-4o, an outdated model known for producing sycophantic prose.
Words used in the reviews were the tell, according to the paper’s authors. They noted that words such as ‘careful,’ ‘thoughtful’ and ‘serene’ appear more frequently in AI-generated writing than in human writing.
Nominalizations — nouns formed from verbs or adjectives, often with suffixes such as ‘-ation’ — also appear more frequently in AI writing. For comparison, Reinhart examined San Antonio’s 2026 guide and found that its nominalization rate was slightly lower than the rate typically found in human writing.
Texas politics is a mess. We make it a readable mess.
Sign up for San Antonio Current’s new “Texas Politics This Week” newsletter.
Sign Up for SA Current newsletters.
Follow us: Apple News | Google News | NewsBreak | Reddit | Instagram | Facebook | Twitter | Or sign up for our RSS Feed
