Substack rolls out AI text scanner for posts and comments
The Pangram-powered feature lets readers check eligible Substack text and see estimates for AI-generated, AI-assisted or human writing.
By Bianca Rossi · Entertainment Editor
3 min read
Substack readers can now run an AI check on posts, notes, replies and comments, with the platform adding a Pangram-powered scanner that estimates how much of a passage was written by a person or generated with AI.
The feature was announced Tuesday in a blog post by Substack co-founder and CEO Chris Best. He framed the tool as a way to close the gap between what readers think they are getting and what they are actually reading.
“Not everything made with AI is slop, and not all slop is made with AI,” Best wrote. “The problem is when there is a mismatch between a reader’s expectation and reality, especially when they unwittingly invest their attention in something with no human thought on the other end.”
The scanner is live on the web and in Substack’s iOS app, according to the company, with Android support set to follow. To use it, readers open the three-dot menu on eligible Substack content and choose “Scan for AI text.”
A pop-up then shows the percentage of text that Pangram has classified as AI-generated, AI-assisted or human. The check only happens when a reader asks for it, according to Substack.
What the scanner will and won’t check
Substack says the tool has limits. It only works on text passages longer than 100 words and on content published after 8:30 a.m. PT on July 21.
Older content triggers a notice saying it is not eligible for AI detection. Shorter passages bring up a message saying there is not enough text for an accurate scan.
Best said the detector also cannot tell whether a human used AI for background research before writing a post. In other words, it is aimed at the text itself, not every step behind the finished piece.
On Best’s own posts, the pop-up includes an extra note: “I'm still figuring all of this out. (In the meantime these words are mine, for better or worse.)”
Writers get controls, too
The rollout also gives Substack authors new tools around disclosure and disputes. Writers can add a statement about how they created a piece, and that statement appears when readers run the AI scan on their posts.
Authors can also test drafts with Pangram’s scanner before publishing. If they believe Pangram has classified their work incorrectly, they can submit a report.
Substack’s support guidance says writers can disable AI detection on their content entirely. Doing so lets them remove the flag, regardless of whether Pangram’s classification was accurate.
Pangram uses AI to detect AI-written text, but the company says its system is a classifier neural network rather than a generative AI model. Pangram says its dataset of human-written material was gathered from sources published before 2021, before generative AI became common online, to reduce the chance of AI text being included in its training data.
Best wrote that Substack is considering more AI-related features based on user feedback. One possibility he mentioned is giving readers preferences for the content Substack recommends, which could eventually include filtering out AI-generated material.
The launch lands amid growing anxiety over machine-written material online. A recent study by digital marketing firm Graphite found that online articles that are primarily AI-generated now match the number written by humans. Several publications have also been criticized in recent years for AI-generated articles, some containing errors and some credited to fake AI-made authors or to “staff.”
This story draws on original reporting from Mashable.