Substack has introduced AI text detection; authors criticize the tool as inaccurate
Substack has launched detection of AI-generated text in collaboration with Pangram, which assigns percentage scores to posts. Some authors criticize the tool as unreliable and a risk to their reputation; the company responds that it is about transparency, not penalties.
According to its statement, Substack has introduced new features for detecting text generated by artificial intelligence, in collaboration with Pangram. The tool is now built directly into the platform and allows any user to have text scanned, generating a percentage score estimating how much of the text was written by a machine and how much by a human. According to the head of Substack, the company is taking this step in response to the growing difficulty of distinguishing what was created by a human on the internet. Authors can also add a “How I make this" statement to their posts, describing their creative process and any use of AI, disable detection on their own posts, and report or remove scores they consider incorrect.
Some authors on the platform have objected to the new feature, calling it a “witch hunt". Critics, including ghostwriter Alice Lemee, say they consider AI text detectors notoriously inaccurate and that a single false accusation can irreversibly damage an author's reputation. In this context, Professor Sam Illingworth warns that, in his view, similar tools generate false alarms particularly for native speakers of other languages and neurodivergent authors, and that simply by starting from suspicion, they hinder dialogue between authors and readers. Max Spero, CEO of Pangram, says the company estimates its false positive rate at roughly one case in ten thousand. A spokesperson for Substack said the new tools are intended to give readers more context and increase transparency, are not intended to ban or punish AI-assisted creation, and do not affect whether content on the platform is recommended to other users.
The introduction of detection was also welcomed by some readers who, according to the article, are tired of unlabeled AI content on the internet. The text links this situation to similar efforts on other platforms, where users themselves are creating tools to detect unlabeled AI content. According to the author, the dispute shows that creating AI content is easy, while the question of how to handle it on platforms remains unresolved.
Why it matters
The case shows the practical risk of automatically labeling content as AI-generated: even with a low reported error rate, individual authors face reputational damage from a false accusation, without a clear way to defend themselves effectively. It is also a concrete example of the compromise platforms choose between transparency for readers and the risk of unfairly penalizing creators.
Two audiences, two different impacts
What this means
For individuals
Anyone publishing on Substack may have their text scanned by Pangram and face the risk of being incorrectly labeled as AI-generated content; they can address this by completing a “How I make this" statement and reporting incorrect scores on their own posts.
For a business
For companies considering deploying AI detection tools as part of content moderation, the case illustrates the reputational risk arising from probabilistic labeling even with a low reported error rate, and the need to ensure users can challenge and remove scores.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.