Why Detecting AI Use at Work Is Easier Said Than Done
Companies see value in AI detectors, but deploying them is far from straightforward.
Anyone who scrolls through LinkedIn has likely noticed a seemingly endless stream of posts that feel distinctly machine-generated. Telltale signs often include formulaic constructions like “It’s not X, it’s Y,” gratuitous em dashes, and overused buzzwords like “delve” or “underscore.”
While these stylistic quirks aren’t definitive proof of AI use, their repeated appearance naturally raises suspicions. This creates a modern workplace dilemma: many employers actively encourage AI to boost productivity, but obvious reliance on it prompts uncomfortable questions. Is the employee using AI as a helpful assistant, or is the tool doing so much of the heavy lifting that it calls into question who truly deserves the credit?
The High Stakes of AI Detection
The consequences of blurred authorship are already rippling through creative and professional industries. In March, Hachette Book Group canceled the U.S. publication of Mia Ballard’s novel *Shy Girl* and discontinued its U.K. edition after allegations of AI-generated content. An AI text-detection tool, Pangram, had flagged 78% of the text as machine-generated. Shortly after, a Commonwealth Short Story Prize winner faced similar suspicions, prompting *Granta*—which had published the prize’s winning stories for over a decade—to step back from external publishing partnerships. In both cases, the authors denied personally using AI.
Even seasoned editors are grappling with false alarms. James Taranto, op-ed editor at *The Wall Street Journal*, recently noted that Pangram flagged three freelance articles he had accepted for publication as AI-generated—a conclusion he disputed after personally investigating the authors’ workflows.
Where Detection Tools Fit In
Amid growing AI adoption and occasional scandals, employers and platforms are trying to figure out where detectors belong in their workflows. Are they a source of friction, or a necessary risk-management tool? Experts suggest their results should be treated as signals, not definitive judgments.
Social media and content platforms are already taking action. LinkedIn recently introduced a “seems like AI slop” button to help users flag suspicious content, and on July 21, Substack rolled out a Pangram-powered feature allowing users to scan published text for likely AI use. “We’re not against people using AI to assist their work,” Substack co-founder and CEO Chris Best noted. “But people should know what they’re getting.”
Similarly, Qwoted, a two-sided expert network connecting journalists with sources, has used detectors to foster transparency. After starting with GPTZero in December 2024, the company switched to Pangram last August. Both journalists and sources can use a “Check for AI” button. If a journalist reports a suspicious submission, Qwoted reviews it and may issue a warning, temporarily disable the account, or, for repeated violations, remove the user entirely, according to Shelby Bridges, Qwoted’s head of user success.
Beyond text, detectors can also help verify the authenticity of audiovisual content. Ilana Golbin Blumenfeld, a partner at PwC US leading the firm’s responsible AI work, notes that when image or video manipulation is suspected, detection tools can help answer foundational questions: *Is this doctored? Is it AI-generated? Did this event actually happen?*
Process Check or Bottleneck?
Despite these use cases, reliability remains a major hurdle. Blumenfeld cautions that because detectors are not infallible, they should only be one input in a broader verification process.
Some organizations have abandoned them altogether. Danya Henninger, editorial director at Philadelphia-based media company Technical.ly, says her publication does not use AI detectors because the results are too unreliable to inform editorial decisions. “The detectors do not work. They are completely fallible,” Henninger says, pointing out that simply changing a sentence or a few words can alter a detector’s result, making the tools easy to game. Instead, Technical.ly relies on internal discussion and human review. Henninger also noted the value of proactive tracking, referencing her past work at OKhuman, a startup that records writing activity to document human involvement in real time, rather than analyzing finished text after the fact.
Henninger also highlighted a 2025 incident where a Deloitte report commissioned by the Australian government included a fabricated quote from a federal court judgment and references to nonexistent academic papers—a stark reminder that in high-stakes situations, human fact-checking is irreplaceable.
Furthermore, Blumenfeld warns that embedding detectors into daily workflows can send mixed messages, especially when organizations are simultaneously encouraging employees to adopt AI. Ultimately, she notes, employees must remain responsible for the work they produce.
Signals, Not Judges
Experts advise that organizations should first define what they want detection tools to accomplish. Tuhin Chakrabarty, an assistant professor of computer science at Stony Brook University, explains that detectors can offer valuable information about a text’s provenance, but it is up to the organization to decide how to act on that data. “Whether you use it for moral policing is on you,” he says.
While AI-generated boilerplate text may be harmless, its use in creative or analytical work raises valid questions about authorship. However, the value of a detector’s signal depends entirely on its accuracy. Chakrabarty notes that low-quality detectors, which have historically flagged classic human-written works as AI-generated, have undermined confidence in the entire category.
Pangram, for its part, claims a false positive rate of about one in 10,000, based on evaluations of hundreds of thousands of human-written documents. Co-founder and CEO Max Spero adds that the company retrains its model every few weeks to keep pace with newly released large language models. While Chakrabarty considers Pangram highly accurate, he acknowledges that any machine learning system will occasionally produce false positives.
The Policy Imperative
Ultimately, an AI detector is only as useful as the policy governing it. Spero emphasizes that organizations must clearly define what kinds of AI use are acceptable. “If you don’t have an AI policy today,” he says, “then it’s basically just a policy that all AI use is fair game.”
As AI-generated text becomes ubiquitous, the question of disclosure will only grow more pressing. Companies must weigh the risks of nondisclosure, which can lead to a massive loss of trust if AI use is discovered, against the risks of transparency. Chakrabarty points out a difficult psychological reality: including a disclaimer that content was AI-produced or assisted might cause audiences to reject it outright.
“I’m very careful about calling disclosure the future,” Chakrabarty says, “because from what I have seen, there is a huge rejection of AI-generated text on the internet. There is so much psychological bias against the use of AI in writing that people might not even engage with your content. They’ll just say, ‘Oh, this is AI slop.’”
Navigating this new reality requires more than just software; it requires nuanced policies, human oversight, and a clear understanding of the tools' limitations.