Substack's new tool tells you who's been writing their newsletters with AI

Substack has integrated Pangram, an AI-writing detection tool, directly into its platform, giving readers and writers a way to see estimates of how much of a given post, note, comment, or reply was likely written by a human versus generated by AI — a feature the company is framing as a transparency tool rather than a policing mechanism.

How the detection tool works

The Pangram integration can scan any post, note, reply, or comment longer than 100 characters and returns an estimate of AI involvement in the text. Substack has been explicit that these are estimates rather than definitive verdicts, acknowledging the inherent uncertainty in AI-detection technology, which has struggled with false positives and false negatives across the industry more broadly.

Writers who do use AI in their process have the option to add an "AI author's note" disclosing that use voluntarily, and publishers can run the Pangram scan on their own drafts before hitting publish, effectively giving them a way to self-check before readers do it for them. Users who believe a scan produced an inaccurate result can report and request removal of that scan.

Encouraging disclosure, not banning AI

Substack CEO Chris Best was careful to frame the tool as pro-transparency rather than anti-AI. "This is good use of AI," he said, arguing that writers should focus on "the hard part" — original ideas and judgment — while leaving mechanical aspects of the writing process to software where it genuinely helps. The company is not proposing to penalize or ban AI-assisted writing outright, a notably different stance from platforms that have moved toward stricter content-authenticity rules.

Part of an industry-wide labeling trend

Substack's move fits into a broader pattern across major content platforms this year. YouTube, TikTok, Spotify, and Tidal have all introduced their own AI-content labeling or disclosure requirements in response to a rising volume of AI-generated material and growing audience demand to know what they're consuming.

For newsletter readers specifically, the tool addresses a trust question that has become increasingly relevant as AI writing tools have gotten harder to distinguish from human prose by eye alone: whether the personal voice a subscriber is paying for, in many cases through paid subscriptions, is genuinely the work of the writer they signed up to read, or substantially generated by a model on that writer's behalf.

Detection tools remain imperfect

Substack's own framing acknowledges the limits of the underlying technology, and independent testing of AI-detection tools across the industry has repeatedly shown they can misclassify heavily edited human writing as AI-generated, and polished AI output as human-written, particularly as language models have become better at mimicking individual stylistic quirks. By building in a reporting mechanism for disputed scans and stopping short of outright content removal or bans, Substack appears to be positioning Pangram as a directional signal for readers rather than a definitive authenticity verdict, an approach that hedges against the technology's known false-positive and false-negative rates while still giving the platform something concrete to show it is taking the AI-authenticity question seriously.

For Substack, which built its business model around paid subscriptions to individual writers rather than algorithmically distributed content, the authenticity question carries particular commercial weight: subscribers are effectively paying for a specific person's judgment and voice, and any erosion of confidence that the writing behind a paywall is genuinely that person's work could undermine the platform's core value proposition more directly than it would for ad-supported platforms where individual authorship matters comparatively less to the audience.