Max Spero, a former Google software engineer, relies on intuition to identify AI-generated text, focusing less on common indicators like em dashes or polished tone and more on what he describes as a lack of "information density." Spero is the co-creator of Pangram, an AI detection tool launched in 2023 designed to differentiate human-written content from machine-generated text. Pangram and similar tools have been used in high-profile controversies involving accusations of AI use in literary and journalistic work, including the removal of a novel by author Mia Ballard amid allegations of AI involvement, which she denied. This summer, Pangram flagged a short story by Trinidadian writer Jamir Nazir as AI-generated, leading to significant debate and dispute.
The growing prevalence of AI in producing written content has complicated the landscape of authorship verification. Some estimates suggest that up to a third of new webpages and a sizable fraction of eBooks may be AI-generated, spurring efforts to develop reliable methods to detect synthetic texts. Platforms like Substack have begun scanning newsletters with AI detectors, citing concerns over the integrity of human communication and the potential for AI-generated content to "pollute the commons," according to Substack CEO Chris Best.
Despite the advances in detection technology, skepticism persists over the reliability and fairness of these tools. Universities including Oxford, Cambridge, and Harvard endorse the use of generative AI by students but reject commercial AI detectors for academic assessments, pointing to the risk of false positives. Instances where canonical human texts, such as the U.S. Declaration of Independence, have been incorrectly flagged as AI illustrate the challenges in drawing a clear boundary between human and artificial writing.
Pangram’s detection method involves training on millions of human-written texts predating AI models and comparing these with AI-generated “synthetic mirror” texts to identify statistical patterns. These processes analyze token sequences, word placement, frequency, sentence structure, and paragraph organization to estimate the likelihood of AI authorship. However, the proprietary nature of detection algorithms and training datasets leaves the precise workings of such tools opaque.
The case of Jamir Nazir further highlights the controversy around AI detection. His short story, "The Serpent in the Grove," which won the Caribbean regional Commonwealth Prize in May, was flagged by Pangram as 100 percent AI-generated. Nazir has consistently denied using AI, describing the stylistic elements pointed to as “AI giveaways” as intentional literary techniques rooted in mysticism. He expresses frustration over what he views as paranoia and unfair witch-hunts targeting writers. Nonetheless, some academics and critics maintain skepticism, noting repeated AI-associated language patterns within his work.
The Commonwealth Foundation publicly supported Nazir, but Granta magazine decided to stop publishing the prize-winning stories, citing the difficulty of verifying textual authenticity amid AI concerns. Granta’s publisher, Sigrid Rausing, expressed openness to transparent, experimental AI use in literature, provided authors disclose their methods.
Looking ahead, regulatory frameworks are anticipated to emerge, with the European Union planning new measures aimed at deterring undisclosed AI-generated text. Some AI developers, like Anthropic with its Claude models, intend to embed imperceptible watermarks to identify synthetic content. However, these markers might be vulnerable to removal through editing, paraphrasing, or copying, and their absence does not definitively signal human authorship.
Experts acknowledge the challenges inherent in this evolving domain. Ethan Mollick, an AI researcher at the University of Pennsylvania, likened the widespread integration of AI in text creation to a geological extinction boundary, suggesting that society is still negotiating the balance between human and machine contributions in writing. Edward Tian, creator of GPTZero, anticipates a future equilibrium where human and AI-generated writing coexist, emphasizing the importance of preserving the critical thinking and authenticity that underpin human communication.
As AI's role in writing expands, unresolved questions remain about appropriate disclosure, ethical use across various fields, and how detection tools will influence trust and verification. The debate continues over how best to protect the integrity of human expression while harnessing the benefits offered by artificial intelligence.