1 month ago
18
Yesterday, Substack announced that readers can now scan text on Substack to see how much of it was likely written with AI assistance. The tool is powered by Pangram, a leader in AI detection. Founder Chris Best writes, “We don’t want to wait until your Substack app turns into LinkedIn before we start to learn and make progress. So to start, we’re launching some new tools that we hope will increase transparency and give us a basis to ask for your feedback.”
Before I get into my hot take, a few important notes: First, the AI analysis is shown to readers only if requested. Second, creators can disable AI detection on their work. And finally, AI detection only works on material published after 8:30 a.m. Pacific Time on July 21 (and then only if it is longer than 100 words).
Substack is also adding tools that allow creators to add a statement called “How I make this” to explain their creative process and set expectations.
Substack emphasizes they are not against people using AI to assist their work, and there are no prohibitions on such work—their stated goal is transparency. Pangram is available to creators to run on drafts prior to publication, and report and remove any scans they believe mistaken, all in addition to disabling the AI detection feature for readers.
It’s fair to say that most writers are deeply unhappy about this development, not least because AI detection tools are not 100 percent accurate and they can be used punitively (in what people call “witch hunts”) and without human judgment and nuance. Others think this is a terrible purity test that doesn’t tell you anything about how much a human was involved. Both are fair critiques, but given how limited the feature currently is and that writers can turn it off, I’m not sure how much harm it can do.
In Substack’s defense (not a position I’m eager to take, but I’ll do it anyway), I do find it responsible of them—at least at this moment in time—to address the dramatic increase of AI-generated writing and offer readers some transparency about what they’re reading, even if the tools are flawed. (The tools are not as enormously flawed as people think—see Links of Interest for the latest study.)
I also appreciate that Substack isn’t expressing overconfidence or certitude (yet)—they say they don’t know the answers. This first step may help foster important discussion, even norms, around AI writing and publishing that I think are needed. If things somehow go off the rails, I expect Substack to adjust course. What I anticipate, however, is that the community will adapt. I believe that platforms today must do something to monitor the provenance of content and identify to the best of their ability what’s AI-generated or AI-assisted on their platform, for the exact purposes of preserving value around what is human.
