Substack readers no longer have to wonder whether a human wrote what they are reading. They can check. The platform announced on Tuesday 21 July that it has built AI detection from Pangram directly into its product, letting readers scan posts, notes, replies and comments longer than 100 words for an estimate of how much of the text was written by hand and how much with AI assistance.

The launch matters to B2B marketing teams because newsletters have become a standard channel for executive and brand thought leadership, and until this week checking a piece for AI required a separate tool and the will to use it. Building detection into the reading experience makes scrutiny the default: every subscriber, prospect, journalist and competitor now has a scan option sitting on every post. The operational question shifts from whether AI-assisted content will be noticed to what a publisher says when it is.

What can Substack readers scan, and what do writers get?

Substack’s scanner works on any post, note, reply or comment longer than 100 words published from 21 July onwards, and shows an analysis only to the reader who requests it. The feature is live on the web and in the iOS app, with Android support promised “soon” and no date given. Readers trigger a check from the menu on a post by choosing “Scan for AI text”, according to The Verge.

Writers get tools of their own. A new “How I make this” statement lets a publisher explain their process and set expectations, writers can run Pangram on drafts before publication, and they can report and remove scans of their own work they believe are mistaken.

Substack co-founder and chief executive Chris Best was careful to say the company is not policing AI use. “We’re not against people using AI to assist their work,” he wrote in the launch post, noting that Substack itself uses AI to write software and build product features. The target, he argued, is the mismatch between what a reader expects and what they get — content written by no one, passed off as human — which the post labels “Claudefishing”. His summary of the position: “people should know what they’re getting.”

How reliable is Pangram’s detection?

Independent testing puts Pangram ahead of rival detectors, though not beyond error. A peer-reviewed study from Vrije Universiteit Brussel, published in June 2026, tested Pangram, GPTZero, Turnitin and Copyleaks on 160 long academic papers and found Pangram was the only tool to produce satisfactory results: it caught 97.5 per cent of fully AI-generated papers and 95 per cent of “humanised” AI text, with zero false positives, including on essays by non-native English speakers, the group earlier detectors most often wrongly flagged. Research from the University of Chicago’s Becker Friedman Institute in 2025 likewise recorded near-zero false positive and false negative rates on medium and long passages.

The counterweight arrived in May, when The Atlantic reported that Pangram makes more mistakes than many of its users realise and warned against treating a tool with the power to damage reputations as proof of anything. Substack concedes the limits itself: the launch post says Pangram is “not perfect”, that it detects only whether AI was used to produce the text — not whether care went into the work, nor whether AI served as a research source — and that every result is an estimate for readers to weigh rather than a verdict.

That framing is the piece marketing teams should hold on to. A scan score is a probability signal, not an audit trail, and Substack has built in a route for writers to dispute results they believe are wrong.

Why has Substack moved now?

The volume of machine-written text on rival platforms gave Substack both a problem to head off and a position to claim. Pangram estimates that as much as 40 per cent of text on some social platforms is AI-generated, and a July analysis by its chief executive Max Spero ranked LinkedIn the most AI-saturated network, flagging 41 per cent of its long-form content, against roughly 10 per cent on Substack, which Spero called “an exception”. Best leaned on the same comparison, writing that Substack does not want to wait “until your Substack app turns into LinkedIn” before acting. The pressure is not new: a GPTZero analysis for Wired in 2024 found ten of Substack’s 100 biggest newsletters were using AI, seven of them heavily.

The reaction shows how contested the line is. Best opened comments on the announcement, and within a day the post had drawn more than 1,000 of them, including writers who use AI arguing that a detection score can show a tool was involved but cannot show who had the original thought or who stands behind the words.

For marketing teams the practical consequence is that disclosure has become a decision to take in advance, not a defence to improvise afterwards. LinkedIn already throttles the reach of generic AI-generated posts, and Substack is weighing whether to go further: Best listed possible tools letting readers set preferences about AI content in their communities and recommendations, which would move a detection score from information into distribution. Teams that use AI with genuine editorial oversight can now say so in the “How I make this” field and let the scan confirm the human hand. Teams publishing undisclosed volume content face a check button on every post, on a platform whose paid-subscription model depends on readers believing a person is behind the work — and with McCann finding 69 per cent of people drop brands they do not trust, guessing wrong carries a measurable cost.

Substack has given no date for Android support, and says the reader-preference and community controls remain under consideration, with what ships next depending on user feedback.