Anyone who scrolls LinkedIn will likely encounter a seemingly endless slew of posts that appear to be AI-generated. Constructions like “It’s not X, it’s Y,” gratuitous use of em dashes and words like “delve” and “underscore” are among the purported AI tells.
Of course, the mere inclusion of these words and phrases isn’t proof that AI was used. But repeated use of the same words and structures may raise suspicions.
While many employers encourage AI use to help employees become more productive, obvious signs of it can raise uncomfortable questions: Are people just using AI as an assistant, or is it doing enough of the work to call into question who deserves the credit?
In March, Hachette Book Group canceled the planned U.S. publication of Mia Ballard’s novel Shy Girl and discontinued its U.K. edition after allegations that it contained AI-generated material. Pangram, an AI text-detection tool, had flagged 78% of the text as AI generated. In May, a Commonwealth Short Story Prize winner was suspected of using AI. The literary publication Granta, which had published the prize’s winning stories for more than a decade, decided to pull back from external publishing partnerships after the controversy. Both authors denied personally using AI.
In both cases, detection tools suggested that at least some of the work was AI generated, though experts and practitioners disagree about how reliable they are. James Taranto, op-ed editor at the Wall Street Journal, recently reported that Pangram flagged three freelance articles he accepted for publication as containing AI-generated writing, a conclusion he disputed after investigating the authors’ use of AI.
Amid growing AI adoption—and scandal—employers are figuring out where detectors fit in the workplace.
Are they a source of friction, or a risk management tool to determine the extent to which outputs are AI generated? Companies say detectors have value depending on how they are used. Their results, however, should be treated as signals rather than judgments about whether AI use is appropriate.
Where AI detectors fit
Social media networks and content management systems have begun introducing detection capabilities. This week, LinkedIn added a “seems like AI slop” button to help the company more easily flag AI-generated content. And on July 21, Substack rolled out a Pangram-powered feature that lets users scan text published on the platform for likely AI use.
“We’re not against people using AI to assist their work, and we think people should be free to choose which tools they use to express themselves,” Chris Best, Substack’s cofounder and CEO, wrote. “But people should know what they’re getting.”
Other platforms have also begun using detectors in their workflows. Qwoted is a two-sided expert network where journalists seek sources; experts and their public relations representatives can respond to inquiries and send story pitches. The company suggests detection brings transparency to communications. It’s used detectors since December 2024, first with GPTZero (another detection tool) before switching to Pangram last August.
Both sides can opt to use a “Check for AI” button. Reporters can check responses for possible AI use, while sources can see if their text could be flagged, says Shelby Bridges, head of user success at Qwoted.
When a journalist reports a submission, the company reviews it. Depending on what it finds, the platform may give the source a warning or temporarily disable the account. Repeated violations may result in removal, Bridges says.
Detectors can be used to verify the authenticity of audiovisual content, says Ilana Golbin Blumenfeld, a partner at PwC US who leads the firm’s responsible AI work. When there’s suspicion of image or video manipulation, detection can help answer basic questions like “Is this image doctored? Is it AI generated? Does this video depict something that really happened?” she says.
But since detectors are not always reliable, she cautions that they should be one input in a broader verification process.
In high-stakes situations—for example, a government-commissioned report—a detector result could prompt additional fact-checking, says Danya Henninger, editorial director at Technical.ly, a Philadelphia-based media company. She brought up a 2025 Deloitte report for the Australian government that included a fabricated quote from a federal court judgment and references to nonexistent academic papers.
Process check or bottleneck?
Technical.ly says it doesn’t use AI detectors, because the results aren’t reliable enough to inform editorial decisions.
“The detectors do not work. They are completely fallible,” Henninger says, noting that changing a sentence or even a few words can alter a detector’s result and make the tools easy to game. The publication instead relies on internal discussion and human review to check work before publication. Henninger is a former consultant at the tech startup OKhuman, which records writing activity to document human involvement rather than analyzing finished text after the fact.
PwC’s Blumenfeld says detection can be valuable in limited situations, including to “validate the authenticity of something that’s coming inbound.” PwC says it does not routinely use detectors but has not ruled out using them in limited cases where they have proven effective.
When used inside a workflow, detectors might send the wrong message when organizations are also encouraging employees to use AI, she adds.
Ultimately, employees are responsible for what they produce.
Not a judge
Experts say organizations should start by asking what they want detection tools to accomplish. Tuhin Chakrabarty, assistant professor of computer science at Stony Brook University, says a detector can offer organizations information about a text’s provenance. It’s up to them to decide how to act on the results.
“Whether you use it for moral policing, it’s on you,” says Chakrabarty.
AI copy may not be a concern with boilerplate text that has no point of view or argument, but for creative work, it can raise questions about authorship or credit, he says.
The value of those signals, however, depends on their accuracy. Low-quality detectors that flagged historical works as AI generated have undermined confidence in the category, Chakrabarty says.
Pangram says its rate of false positives is about one in 10,000, based on evaluations of hundreds of thousands of human-written documents. The company retrains its model every few weeks as new large language models are released, according to Max Spero, cofounder and CEO of Pangram Labs.
Chakrabarty says he considers Pangram highly accurate, though he acknowledges that any machine learning system can produce false positives.
Of course, an AI detector is only as useful as the policy governing it. Spero says organizations should spell out what kinds of AI use are acceptable.
“If you don’t have an AI policy today, then it’s basically just a policy that all AI use is fair game,” he says.
As AI-generated text becomes more common, questions around disclosure may also become more pressing. Companies may need to consider whether to disclose AI use in public-facing content. Nondisclosure is risky because it can lead to a “huge lack of trust,” Chakrabarty says. But including a disclaimer that content was AI-produced or assisted could present a dilemma of its own: People may reject the content if they know AI was used to create it.
“I’m very careful about calling [disclosure by companies using AI-generated content] the future because, from what I have seen, there is a huge rejection of AI-generated text on the internet,” he says.
“There is so much psychological bias against the use of AI in writing, [people] might not even engage with your content. They can say, ‘Oh, this is AI slop.’”
