Tech
I Test-Drove Substack’s New AI Detection Tool, and It Mostly Worked
More and more, we want to know: Is it human, or is it AI? Is that image real? That song? That video? The publishing platform Substack is trying to add some clarity by partnering with the AI detector Pangram to help readers understand just how much AI is in all the content they consume on the publishing platform.
The tool is available to all Substack subscribers and works on posts, notes, comments and replies published on or after Tuesday that are longer than 100 words. Creators will also be able to scan their drafts with Pangram to see how much AI is detected and can also provide a statement to explain their writing process, Substack said.
Web and iOS users can try the Pangram features now, and Android users will get access later.
In a blog post announcing the new feature, Substack co-founder and CEO Chris Best said his platform aims to avoid “Claudefishing,” meaning when a reader consumes content without realizing it was generated by AI.
“When readers have to wonder if what they’re reading is real, it undermines trust in authorship and threatens the livelihood of writers — including those who use AI tools thoughtfully to produce work they believe in,” Best wrote.
From CNET: AI Slop Is Destroying the Internet. These Are the People Fighting to Save It
Pangram estimates that roughly 40% of text on some social media platforms is AI-generated, with a stunning two-thirds of that on the networking site LinkedIn. And AI growth agency Graphite recently published a white paper saying that AI is now writing as many online articles as humans. A Gartner survey found that 68% of people wonder if the content they see is real, and that half of consumers prefer brands that aren’t using AI in their marketing.
Testing it for myself
I’ve only recently joined Substack, but I wanted to check out the new Pangram AI detection tool. The tool works on posts, notes, comments and replies, so I decided to use posts as my makeshift testing lab. First, I logged in to my Substack account and published a new post. I copied and pasted text from an article I had written — by myself, no AI — into the Substack post and published it.
After a few minutes, I clicked on the live post. On the three-dot menu at the upper right of the page, I saw a “Scan for AI text” option. I clicked on it, and a small window popped up accurately declaring that my article was “Fully Human-Written.” A line appears below this that indicates the percentage of the text that is AI, AI-assisted, or human.
I next took the same post but added a long, fully AI-generated paragraph that I pasted from ChatGPT. When I performed the “Scan for AI text,” the tool reported that 10% of the article was written by AI, a roughly accurate estimate.
I next published a fully AI-generated post, and the scan correctly identified it as 100% AI-generated.
For another test, I added a couple of AI-generated sentences to a post that was otherwise purely human content, and the scan flagged it as 100% human-generated. Close, but not quite right. So, the Pangram tool worked pretty well for my crude experimentation.
Although I was fairly satisfied with how well Substack’s AI detection system worked on my own content, I was disappointed to learn that one of my favorite podcaster-bloggers apparently went full AI on his latest post. Other writers I checked out, however, used fully human-generated text.
What about false positives?
In his announcement, Best admitted that the AI detection tool is “not perfect,” but cited independent research that found Pangram’s software was 97.5% accurate in detecting fully generated AI text. The University of Maryland found that Pangram was 99.3% accurate in identifying “humanized AI-generated text,” which basically means a human has rewritten or paraphrased AI content.
Substack writers who do not use AI are understandably concerned about being incorrectly labeled as AI-assisted authors. As a result, the risk of false positives is a significant issue. Substack said creators can scan their content to see if it will be labeled as containing AI while it’s still in draft form. If Pangram incorrectly identifies AI text, the creator can report the error to Substack and have the false scan removed.
To promote greater transparency, publishers can now include a “How I make this” statement if they want readers to know how much, if any, AI they have used in their content. Best said Pangram can’t detect if AI was used to gather information or as a source for written pieces.
“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,” Best said. “But people should know what they’re getting.”
You must be logged in to post a comment Login