Timnit Gebru urged people on November 16 not to use Pangram to determine whether writing is AI-generated, sharing an earlier critique by Emily M. Bender. “Please don’t use pangram to determine what is ‘AI generated’ vs not,” Gebru wrote in a quote-post. Source
The essay she linked, “Pangram is not an oracle,” was published October 19 in Mystery AI Hype Theater 3000: The Newsletter. It argues against treating a detector’s output as conclusive evidence about who—or what—wrote a particular text. A percentage sign does not, in Bender’s telling, turn a prediction into a verdict. Source
What the accuracy argument leaves out
Bender distinguishes two errors: a false positive labels human-written text as AI-generated; a false negative lets generated text pass as human-written. She argues that readers need both rates to understand a detector’s performance, rather than relying on an overall accuracy figure or a case study covering only one kind of error. Source
She also questions how evaluation results translate to the writing someone actually wants to check. Literary fiction, high school essays and sports reporting are different applications, she notes, with further variation across language communities and time periods. Users need evidence from “sufficiently similar language use” before applying a benchmark result to their own situation, she argues. Source
Bender acknowledges that Pangram 4’s technical report addresses both error types. An excerpt reproduced in her newsletter reports a company-measured false-positive rate of 0.0041% on more than a million human-written English texts. But she argues that the report gives too little detail about the evaluation data to judge how that result applies to other writing. Source
The pushback—and the stakes
In a quote-post of Gebru, @Afinetheorem disputed her characterization of the newsletter as technical and asserted that papers address whether detectors can empirically or theoretically achieve approximately zero false positives. “They can & do,” the account wrote. The post’s text does not name those papers, and its broad claim about detectors does not by itself establish Pangram’s performance. Source
Another reply, from @tranche_cheddar, argued that “a whole body of evidence should be considered,” rather than a single detector score. Source
Bender argues that the stakes extend beyond technical error rates, particularly for Black authors facing demands to establish their credibility. She warns that false positives could lead to student punishment or damage an author’s reputation. Source
“A system like Pangram, if used at all, could be a starting point of the conversation, not the end of it,” she writes. Source
Comments