A suspicion can now begin with a simple feeling: a text sounds too polished, too predictable, too much like something generated by a machine. Increasingly, that suspicion is being tested through Pangram, an artificial intelligence detection system that has become one of the most discussed tools in the growing debate over AI-generated writing.
The software, developed by New York-based start-up Pangram Labs, is designed to determine whether a text was created by artificial intelligence. Its results have been used to support accusations involving authors, journalists, researchers, and public figures. But as the technology gains influence, so does the controversy surrounding its reliability and the consequences of its judgments.
In an interview with Die Zeit, Pangram co-founder Max Spero defended the company’s technology and argued that the software is not creating suspicion around AI-generated content but responding to concerns that already exist.
“People read something and think: this sounds like AI,” Spero explained. Pangram, he said, simply provides a technical assessment that reinforces or challenges that initial impression. Without such software, he believes public skepticism toward machine-generated writing would have emerged anyway.
Spero says Pangram was created in response to a rapidly changing information environment. Artificial intelligence tools can now produce large volumes of text within seconds, creating opportunities for productivity but also raising concerns when people present machine-generated work as their own.
The issue has become particularly significant in fields where originality and individual effort are central. Spero points to academic research, where AI-generated papers could place additional pressure on peer-review systems. He also refers to controversy surrounding the Commonwealth Short Story Prize, where one winning entry came under scrutiny after claims that artificial intelligence had been used in its creation.
Pangram analyzed the winning text and concluded that it was fully AI-generated. However, the Commonwealth Foundation later stated that after reviewing drafts and speaking with the author, it found no reason to believe the accusations were accurate. The case highlighted the uncertainty surrounding AI detection tools when applied to real-world situations rather than controlled tests.
Spero maintains that Pangram’s purpose is not to replace human judgment but to provide additional evidence. The company’s system, he explains, is based on a concept called “synthetic mirrors.” Researchers begin with human-written texts and ask AI models such as ChatGPT to create similar pieces without access to the originals. By repeating this process thousands of times, Pangram builds a dataset that allows its own model to identify differences between human and machine-generated writing.
The system does not rely on a single suspicious word or phrase. According to Spero, there is no individual linguistic feature that proves a machine wrote a text. Instead, Pangram evaluates the overall structure and patterns of choices made throughout the writing process.
Human writers generally show greater variation in their language decisions, while AI systems tend to produce patterns that are more consistent. Pangram attempts to identify these broader statistical signals rather than searching for specific expressions commonly associated with chatbots.
The growing influence of AI detectors has also exposed the risks of public accusations. On platforms such as X, users have shared Pangram results as evidence against writers, sometimes before the circumstances behind a text are fully understood. Critics argue that such results can damage reputations because authors often have limited ways to prove how their work was created.
Spero acknowledges that false accusations can occur. He points to a case involving a US journalist whose article was accused of being entirely AI-written after a short excerpt was analyzed. Pangram later determined that the complete article was human-written.
The incident, he said, was caused partly by the limited amount of text submitted for analysis. The original sample contained only about 51 words, while longer texts provide the system with stronger statistical information. Following the controversy, Pangram adjusted its interface to display the length of submitted texts alongside results, making it more difficult to circulate misleading screenshots.
The question of accuracy becomes even more complicated as AI-generated writing evolves. Many people do not ask chatbots to create entire articles or essays but instead use them for rewriting, editing, or improving tone. Pangram attempts to distinguish between these uses.
If artificial intelligence is used only for minor corrections, such as fixing spelling mistakes, the system should identify the work as human-created, Spero says. But if a chatbot substantially rewrites a person’s ideas or generates a complete article from a few notes, Pangram would classify it as AI-supported or AI-generated.
The broader challenge, according to Spero, is that society has not yet agreed on acceptable boundaries for AI use. Different institutions are creating their own rules, but questions remain unresolved in areas ranging from political speeches to academic writing and opinion articles.
Pangram, he says, should not become the final authority in these debates. Instead, communities must decide how artificial intelligence should be used and what level of disclosure is required.
For Spero, the issue is not simply about writing quality. AI systems may already produce prose that rivals or surpasses the average human writer, but readers still care about whether the words they encounter came from another person or from a machine.
Pangram continues to update its model roughly once a month as AI technology develops. Spero argues that newer AI models are not always harder to detect because companies also train them with recognizable styles and behaviors that can create new patterns for detection systems to identify.
The company’s biggest limitation, he says, remains that it only analyzes text. As AI-generated images and videos become increasingly sophisticated, distinguishing real from artificial content is becoming a broader challenge.
The debate surrounding Pangram reflects a larger transformation in how society understands authorship, trust, and creativity. As artificial intelligence becomes a routine part of communication, the question is no longer only whether machines can write — but how people will decide when those words still represent a human voice.
