“In the spring, a screenshot circulated through news feeds. Zero GPT, one of the most widely used AI detection programs, had analyzed a paragraph from Mary Shelley’s *Frankenstein* and delivered its verdict: ‘100% AI-generated.’ A novel from 1818, written by a woman who was not yet twenty years old. At that moment, it was clear to everyone that the program was not living up to its promise. It was ill-suited for identifying AI-generated texts. The consensus at the time was that no program existed capable of reliably distinguishing human-written text from machine-generated text.
Yet the AI market is evolving rapidly. And the new generation of AI detectors can do things manufacturers wouldn't have dared to dream of just a few months ago. Pangram is currently leading the pack. With version 3.0 of its detector, the company behind it—Pangram Labs—has put a figure out there: 99.98 percent accuracy. This figure is based on tests in which researchers from the University of Chicago and the University of Maryland evaluated the detectability of more than six million machine-generated texts.
Editorial offices and universities are increasingly turning to this tool to check texts, and they aren't entirely off the mark. Pangram reliably identifies the opening of *Frankenstein* as human-written. It also helped expose guest articles by Mario Voigt that were likely AI-generated. The portal ‘Frag den Staat’ stated that the program detected 100 percent AI content in his texts. Additionally, there were three verbatim quotes in his guest article for the *F.A.Z.* that could not be verified.
But what exactly does Pangram measure? Does the program actually recognize whether a text comes from a human or a machine? And what does it mean for a text to be “100% Human Written,” as Pangram indicates when it suspects a human is behind a text?
Pangram is trained to recognize stylistic regularities characteristic of language models. It looks for patterns on the surface level of the language. These include a conspicuous frequency of antitheses (famous example: "Ask not what your country can do for you—ask what you can do for your country" (John F. Kennedy)), a uniform subject-verb-object structure, and a density of stylistic devices atypical of human writing. When these patterns appear frequently, Pangram raises an alarm.
Pangram is now everywhere. On the platform X, users can simply tag it to have others' posts checked for AI content. AI hunters use the program to expose politicians or journalists. The most prominent case involved Pangram’s CEO, Max Spero—who calls himself the “slop janitor” on X; he ran 871 articles by *Guardian* sports writer Bryan Armen Graham through his program and posted the result: nine texts in two weeks were flagged as entirely AI-generated. The *Guardian* rejected the accusation, noting that Graham had been writing in that exact style for eleven years—long before language models existed. The accusation had no consequences for Graham, at least none that became public.
Journalist Matthias Meisner recently highlighted another case involving the *Ostdeutsche Allgemeine Zeitung* (OAZ), a paper founded by Holger Friedrich. A Bluesky user named “Hoywoj” ran dozens of OAZ articles through Pangram. The result: around two-thirds came back as wholly or partially AI-generated, including texts by Managing Director Dirk Jehmlich—who, just six months earlier, had warned on LinkedIn that AI was “not the solution.” However, the user “Hoywoj” has since deleted the Bluesky post containing the allegations against the OAZ. Indeed, caution is warranted with this type of “exposure.”
Pangram does not detect whether a thinking human being is behind an article; instead, it provides a style profile. If the program raises an alarm, it is a strong indication that someone received help—whether in whole or in part—in shaping the text. However, seeking assistance with phrasing does not necessarily mean an AI-generated idea was used. While that is certainly possible, the program simply cannot prove it.
It is also possible for someone to dictate their ideas and arguments to an AI assistant and have the machine flesh them out. Politicians, of course, have been doing this long before the advent of AI programs; previously, they would toss bullet points to human assistants to turn into a text or speech. Although these were written by a human, the politician was not the actual author. Furthermore, we consider it perfectly normal for journalists to act as ghostwriters for celebrity books. Or, weak articles are sometimes edited so extensively that not a single sentence remains as the author originally wrote it.
We have become accustomed to attributing the text—in all these cases—to the person named, to attribute it to a named author or speaker. It would be a mistake to deny that texts containing machine-generated phrasing constitute the author's intellectual property. Something more is required for that—fabricated quotes, for instance, or incorrect references, nonsensical contrasts, and awkward metaphors.
Here are three examples of Pangram’s output: If a journalist develops a thesis, conducts research, substantiates the argument, and then hands the draft to their language model, Pangram will likely classify the text as 100 percent AI-generated. Conversely, if someone writes their own text but adopts the idea, argumentation, and research from AI without verification, Pangram would confirm that the text is human-authored.
And anyone who manages to prompt AI models precisely enough to avoid antitheses, vary sentence structure, use strong verbs, and eschew lists can outsmart Pangram. The software sometimes classifies such AI texts as "100% Human Written."
Journalistic authorship has never been merely a matter of phrasing; it involves research, analysis, argumentation, and the formation of judgment.
Someone who uses a tool to polish their language (though it is debatable whether AI actually improves language) has done the work themselves.
Someone who—like Mathias Döpfner—tosses a prompt at an AI in "a second" to generate an opinion piece has not.
The problem is that Pangram does not distinguish between these scenarios. The program is not fit to serve as a truth machine.” [1]
1. Warum Pangram keine KI-Wahrheitsmaschine ist: Das Programm verspricht mehr, als es halten kann. Frankfurter Allgemeine Zeitung; Frankfurt. 20 June 2026: 13. KIRA KRAMER
Komentarų nėra:
Rašyti komentarą