AI Detection as a Transparency Layer, Not a Punishment System
Substack’s new AI detection tool is a reader-invoked feature that estimates how much of a post was written or assisted by AI, while letting creators disclose or even disable detection on their own work, turning AI transparency into an option rather than a mandate for everyone on the platform.
That choice-first design is the real story. Substack partnered with Pangram to scan posts, notes, replies, and comments so readers can see what share of a text appears AI-generated, AI-assisted, or human-authored. The tool works on content longer than 100 words published after a specific launch time, and it’s available on the web and iOS, with Android support on the way. In an ecosystem drowning in AI-generated content, this is less about policing writers and more about correcting a broken expectation: readers no longer know whether a post carries human thought on the other end. Substack’s stance is direct: “people should know what they’re getting.”
How Substack’s AI Detection Tools Work in Practice
In practical terms, Substack’s AI detection feature lives in the same place readers already manage everything else: the three-dot menu. A reader taps “Scan for AI text,” and a popup reports what percentage of that specific passage Pangram thinks is AI-generated, AI-assisted, or human-written. There’s no automatic scanning: Pangram only analyzes a post when a reader asks for it, making AI provenance a choice rather than a constant nag.
There are constraints, and they matter. The classifier only runs on passages longer than 100 words and only on content published after the feature’s launch window, because shorter or older text doesn’t give enough signal for a meaningful result. This is where the tool’s limitations intersect with its ambition. Even Substack admits that no AI detection tools can guarantee accuracy, and independent scrutiny has already shown that Pangram, like peers, can misclassify both human and AI work. Yet Substack shipped it anyway — not as a final verdict, but as one more signal readers can weigh when they decide what deserves their attention.
Putting Control in Creators’ Hands Instead of Enforcing Blanket Rules
The most controversial Substack feature isn’t the scanner; it’s the off switch. Creators can now publish an explicit statement explaining how they used (or did not use) AI in their writing, which appears alongside the detection result. They can also run Pangram on drafts and contest misclassifications. And if they think the tool gets it wrong — or simply don’t want it on their work — they can completely disable AI detection on their posts, removing the flag regardless of what Pangram concluded.
That will upset people who want creator platforms to enforce strict rules against AI-generated content, but it’s a smarter long-term bet. Forced detection regimes inevitably punish edge cases: writers who draft in AI and revise heavily, non-native speakers, or anyone whose style happens to resemble a model’s training data. Substack’s approach shifts the emphasis from control to consent: readers have the power to scan when they care; writers decide how transparent they want to be. In a world where blanket policies are easy PR but blunt instruments, letting humans decide when and how AI is disclosed is a rare show of trust in the actual participants.
Why Creator Platforms Need Transparency, Not AI Witch Hunts
The backdrop to all of this is an internet where AI-generated content is no longer a fringe case. One recent study found that the number of online articles that are primarily AI-generated is now equal to the number written by humans. Several outlets have been called out for publishing AI-written articles, sometimes credited to fake authors and riddled with errors. Substack’s co-founder frames the problem clearly: not everything made with AI is “slop,” but readers feel cheated when they invest attention in something with no human thought on the other end.
Substack’s AI detection tools are an attempt to fix that mismatch of expectations without demonizing AI itself. Pangram uses a classifier neural network to distinguish between human and AI text, trained on pre-2021 human writing to avoid contamination by generative outputs. It’s far from bulletproof, and Substack admits there are limits to what it can detect. Still, giving readers a visible estimate of AI involvement — even an imperfect one — pulls AI out of the shadows and into the disclosure layer where it belongs. The risk isn’t that this becomes a witch hunt; the risk is that platforms shrug and leave audiences guessing.
The Next Wave: From Detection to Preference-Driven Feeds
Substack’s AI detection feature is clearly a first step, not the finish line. The company is already talking about additional Substack features tied to AI content and reader preferences. Future plans on the table include letting users set how much AI-generated content they want in their recommendations and potentially allowing them to filter out AI-heavy posts entirely. That turns detection from a passive label into a signal that shapes what people see in their feeds.
This is part of a wider trend: creator platforms are adding AI detection tools not as gatekeepers, but as plumbing for transparency. Pangram itself is AI detecting AI, a meta-layer where a classifier keeps track of what generative models might have touched. The bigger shift is cultural. If platforms keep building preference-based filters — human-only streams, AI-assisted discovery modes, or hybrid feeds — the question won’t be “Is this content pure?” but “Does this mix match what this reader wants?” The platforms that win will be the ones that stop pretending they can outlaw AI and instead give people knobs, dials, and clear labels. Substack’s new tools are imperfect, but they move the debate where it belongs: away from purity tests and toward informed choice.
