Substack adds tool that detects AI-generated content

Substack is introducing a new tool to help readers figure out whether content was written by a human or generated by AI. Announced in a blog post on Tuesday, the self-publishing platform's new feature is powered by AI detection tool Pangram, and is available on Substack posts, notes, replies, and comments.

"Not everything made with AI is slop, and not all slop is made with AI," wrote Substack co-founder and CEO Chris Best. "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."

Substack's AI detection tool is available now on web and in the iOS app, with Android to follow. To scan Substack content for AI contributions, simply click or tap on the "..." meatball menu, then select "Scan for AI text." A popup will then appear stating what percentage of the text Pangram has determined is AI-generated, AI-assisted, or human.

Substack's new AI detection feature does have some restrictions. Specifically, it will only work on Substack text passages longer than 100 words, and only those published after 8:30 a.m. PT / 11:30 a.m. on July 21. Attempting to scan Substack content published prior to this results in a popup stating that it's ineligible for AI detection. 

Trying to check passages that are shorter than 100 words also surfaces a popup explaining that there isn't enough text for accurate AI detection. The popup on Best's own posts also offers an additional message: "I'm still figuring all of this out. (In the meantime these words are mine, for better or worse.)"

Pengram will only scan text for any AI authorship if the reader requests it. As such, it's up to users to take the initiative and investigate the provenance of what they're reading to the best of their ability. Best notes that the tool cannot detect if, for example, a human used AI to do research before writing a post.

Substack's AI-detection tool adds some new features for authors as well. Writers are now able to add a statement explaining how they wrote content, which will appear when readers use the AI detection tool on their posts. Authors can run Pengram's AI detection tool on their drafts and submit a report if they believe it has classified their content incorrectly. Interestingly, writers can also remove the flag themselves by completely disabling AI detection on their content, regardless of whether or not Pengram's detection was correct.

Pangram itself uses AI to detect AI-generated text. However, rather than generative AI, it deploys a classifier neural network model trained to recognise and distinguish between human and AI-generated text. The company states that its algorithm's dataset of human authored text was gathered from sources published prior to 2021, before generative AI was widely used on the internet, to ensure no AI-generated text snuck its way in.

Best stated that Substack is looking into adding more AI-focused features in the future, depending upon user interest and feedback. This includes allowing users to set preferences for the content Substack recommends to them, suggesting that we may eventually see the option to filter out AI-generated content altogether.

A recent study by digital marketing firm Graphite found that the number of online articles that are primarily AI-generated is now equal to the number written by humans. Multiple publications have been called out for using AI to generate articles over the last few years, several of which were laden with errors. In some instances, these AI-generated articles were attributed to fake, AI-generated authors or the publication's "staff."