Nostalgia: Then & Now · Betty Harlan · 22 July 2026

Substack adds tool that detects AI-generated content

Substack adds tool that detects AI-generated content

Substack adds tool that detects AI-generated content across posts, notes, replies, and comments. Powered by detection firm Pangram and announced Tuesday, it lets readers request a scan and see what percentage of text looks AI-generated, AI-assisted, or human—so expectations match reality. The rollout arrives as newsletter platforms face the same authenticity questions that once seemed distant from personal writing.

For years, a Substack inbox felt like a direct line to a human voice. That trust is harder to take for granted when AI can draft fluent prose in seconds. Substack's answer is not an automatic label on every post, but an on-demand scanner readers can run themselves—an update that fits the Nostalgia: Then & Now tension between old-school authorship and today's machine-assisted feeds.

Key Takeaways

What happened when Substack adds tool that detects AI writing?

According to Mashable's report, Substack announced the feature in a Tuesday blog post. Co-founder and CEO Chris Best framed the problem plainly: not everything made with AI is "slop," and not all slop is made with AI. The real damage, he argued, comes when a reader's expectation and reality diverge—especially when someone unwittingly invests attention in writing with no human thought on the other end.

That framing matters more than a splashy product name. The platform is treating provenance as a reader-trust problem, not a blanket ban. Detection runs only when someone asks for it. Pangram will not silently watermark every newsletter in your feed.

Availability is broad on the surface: posts, notes, replies, and comments can all be checked. On supported surfaces, the path is simple—tap or click the meatball ("...") menu, then select "Scan for AI text." A popup reports what percentage Pangram judged AI-generated, AI-assisted, or human.

How does the Pangram AI scan work on Substack?

Pangram itself uses AI to spot AI-written text, but not as a generative model. Mashable reports that it relies on a classifier neural network trained to distinguish human prose from machine output. Pangram says its human-authored training text came from sources published before 2021, before generative AI flooded the open web, to reduce the chance that synthetic writing contaminated the "human" baseline.

The scan is opt-in for readers. That design puts investigative effort on the audience: if you care whether a piece is machine-made, you have to open the menu and ask. Best also notes a clear blind spot—the tool cannot tell if a human used AI for research and then wrote the post themselves.

Authors get their own controls. Writers can add a statement explaining how they produced the work; that note appears when readers run detection on the post. They can run Pangram on drafts and submit a report if they think a classification is wrong. They can also remove the flag path entirely by disabling AI detection on their content, whether or not Pangram's call was accurate.

Best's own posts also offer an extra popup message: he is "still figuring all of this out," while insisting the words are his "for better or worse." It is a candid admission that the company is shipping a trust tool while still negotiating what fair disclosure looks like.

What limits should readers know before they scan?

The feature is not universal. Mashable notes two hard gates. First, passages shorter than 100 words cannot be scanned accurately; the popup explains there is not enough text. Second, only Substack text published after 8:30 a.m. PT / 11:30 a.m. on July 21 is eligible. Older posts trigger an ineligible message.

Platform coverage is also staggered. The scanner is available now on web and in the iOS app, with Android still to follow. Anyone reading primarily on Android will need to wait or switch surfaces for the same check.

Those limits reshape how useful the tool feels day to day. A short note, a quick reply, or an older archive piece may sit outside the net entirely. Readers hunting for a definitive "this is AI" stamp on every corner of Substack will not find one—and authors who disable detection can opt out of the reader-facing scan path altogether.

Why does AI detection matter for newsletters now?

Newsletters once sold intimacy: a person writing to people. The format still markets that promise, even as AI writing tools make fluent drafts cheap. Substack's move is less nostalgia theater than damage control for that promise. When Best talks about mismatched expectations, he is describing the gap between the classic "from my desk to yours" myth and a feed where machine text can pass as personal essay.

Mashable also cites a Graphite study finding that the number of online articles that are primarily AI-generated now equals the number written by humans. Publications have been called out for AI-written pieces laced with errors, sometimes attributed to fake AI "authors" or anonymous staff bylines. Against that backdrop, a reader-triggered detector on a major publishing network is a credibility signal as much as a gadget.

Best said Substack is weighing more AI-related features based on feedback, including recommendation preferences that could eventually let people filter out AI-generated content. If that arrives, today's manual scan may look like a first step from "check this post" toward "shape my whole feed."

Until then, the practical takeaway is modest and useful: Substack adds tool that helps curious readers pressure-test longer, newer text—without claiming omniscience, and without forcing a verdict on every sentence in your inbox.

← Open in blast feed