Substack’s new AI-detection tool has some critics worried about false flags
As artificial intelligence continues to evolve, it’s getting harder to determine what’s been created by a human and what’s been generated by a machine. To try to combat this confusion, the publishing platform Substack announced it’s partnering with an AI detection software company to flag AI-crafted content on its site.
Substack announced the new partnership on Tuesday with Pangram. This collaboration enables platform users to analyze posts, comments and replies to evaluate the extent of AI-generated content, if present.
The feature will be available in Substack’s app for any content that is above 100 characters. The company also said the partnership allows writers to include an optional AI author’s note, which lets creators disclose their use of AI, according to TechCrunch.
How does Pangram work?
Pangram said it created its AI detector by training it on human text written before 2021. The company states that its latest model, Pangram 3.3, has a 0.01% false positive rate. But outside studies of similar programs that advertised similar results found that the rates were underestimated. For example, Turnitin, a similar AI detection program, said its false positive rate was less than 1%, but a Washington Post report from 2023 found that it was more than 50%.
Pangram has said that its software performed the best in a review of 30 free AI detectors, with multiple universities also finding it the most reliable.
But that doesn’t mean there aren’t issues with it, something Substack acknowledged.
“It is not perfect, but independent research suggests that it detects AI-generated text with a high degree of accuracy,” Substack CEO Chris Best said in a post. “We believe it is beneficial for people to have access to tools like this to help make their own judgments.”
Pangram can only detect whether a writer used AI to produce the text, not whether AI was used as a research tool. In May, The Atlantic published an article criticizing Pangram. They said AI detectors as a whole had issues and that since Pangram is considered the “gold standard” in AI detection software, it could lead to more headaches.
“While Pangram is accumulating the power to end reputations and careers, the tool does make mistakes, perhaps to a greater extent than is currently understood,” The Atlantic’s Matteo Wong wrote. “In turn, AI accusations could very quickly spiral into a witch hunt.”
Another issue is that since the company trained Pangram on data from 2021 and before, critics worry that newer AI models can deceive the program. But the company’s CEO, Max Spero, says they’re confident that the program they’ve created is capable of detecting AI-written content from human-written content.
What are Substack users saying?
After the announcement, several users reacted to the partnership. Many were excited about the change, but others were skeptical about Pangram’s abilities, saying there would be workarounds to stop it.
One user who replied to Substack’s X post questioned how Pangram would respond to Substack content written by neurodivergent authors. In his reply, the user cited a study published in 2025 that found that some AI detectors more often flagged content written by neurodivergent people than neurotypical people. The study didn’t specifically test Pangram but did test OpenAI’s GPT-2 output detection model, and while it didn’t find a straightforward answer to why the software more frequently flagged text written by neurodivergent people, it did call for more scrutiny for these programs.
While it’s still too early to see if Substack’s partnership with Pangram will result in correct detection, Substack is following other companies’ footsteps. Many other online platforms feature some form of AI detection software. Several social media groups already feature programs that automatically detect if a photo or video was likely generated using AI. Music streaming platforms have also begun rolling out features that detect and even penalize AI-generated content, TechCrunch reports.
But those detectors also carry flaws, with Reuters reporting that Meta’s AI detection feature failed to identify more than half of images created using Meta’s own image generation software.
Round out your reading
- From Obama to Trump to Biden: When networks said ‘no’ to presidential speeches.
- ‘The end of American science as we know it.’ White House funding plan sparks alarm.
- AI data centers are popping up everywhere. What exactly do they do?
- The Trump administration is accused of breaking a law built to target the KKK.
- One Republican senator is dead, another is still alive. Both are the center of conspiracy theories.
Ella Rae Greene, Editor In Chief
