Substack Launches AI Detection with Pangram — What It Flags and How

TL;DR: Substack launched an AI detection feature on July 21, 2026, integrating with Pangram to scan posts, notes, replies, and comments over 100 words, showing an estimated split between human-written, AI-assisted, and AI-generated content. It's live on web and iOS, with Android coming. CEO Chris Best framed the goal as stopping Substack from becoming "like LinkedIn" — flooded with low-quality AI content. Writers are split: some cheer it as validation for real effort, others worry about false positives on genuinely human writing.

What Substack shipped

@Substack announced the feature on X:

"Today, Substack is launching an AI detection feature, via an integration with @pangram. Going forward, you'll be able to scan posts, replies, and comments on the Substack app to see an estimate of how much of it was written by a human, or with AI assistance."

Pangram is an AI-text-detection company; Substack's integration runs its detection model over content published on the platform. The scan applies to posts, notes, replies, and comments over 100 words, covering content published since July 21, 2026 — it does not retroactively score the platform's entire back catalog. The feature is live on web and iOS, with Android support coming per the announcement.

Substack CEO Chris Best (@chrisbest) introduced the reasoning: the platform wants to avoid becoming a venue flooded with "AI slop" — the same term driving Peter Yang's /no-ai-slop skill release the same week, a coincidence of timing that put both stories in front of the same audience simultaneously.

TL;DR — how the detector works

Why now — the "LinkedIn problem"

Best's framing directly names the fear driving this feature: that Substack could become like LinkedIn, a platform widely mocked for volumes of generic, AI-flavored "thought leadership" content that reads as formulaic regardless of topic. Substack's business model depends on readers trusting that a subscription buys access to a specific writer's actual thinking — if a meaningful share of content on the platform is AI-generated without disclosure, that trust proposition weakens for every writer on the platform, not just the ones producing low-effort content.

This is a different motivation than, say, a search engine trying to demote AI spam for ranking-quality reasons. Substack's stakes are more direct: writers pay to be there, readers pay to read specific writers, and the entire value chain depends on perceived authenticity of authorship.

The false-positive problem

The most substantive pushback on X targeted detection accuracy, not the feature's intent. @AdamRy_n wrote:

"i'm 1000% for AI slop to be removed from the world. also, when i run my old tweets from 2021/2022 into these programs they come back ~80% AI. The world has people who wrote a certain way and the LLMs learned how to write from those people. And now their way of writing can be [flagged]"

This is a structural problem with AI-text detection generally, not specific to Pangram: LLMs were trained on human writing, including specific writers' distinctive styles. A writer whose natural voice already features rhythms or patterns that overlap with common AI outputs — repetitive sentence structures, certain transitional phrases, particular punctuation habits — risks being misclassified, especially on older writing predating the current AI-slop aesthetic entirely. The claim of ~80% AI on 2021-2022 tweets (written before ChatGPT's public release in most cases) is a pointed illustration: text that could not possibly have been AI-generated at time of writing still triggered a high AI-likelihood score.

Substack's mitigation is disclosure, not certainty: writers can add an AI usage statement or opt out of the display. That shifts the burden toward writers proactively explaining their process rather than the platform claiming definitive detection accuracy — a more defensible design than presenting Pangram's score as ground truth.

Building tools to beat the detector

Within hours of the launch, at least one account was already positioning a business around defeating AI detectors:

"i'm launching a new startup. we're training a custom model whose sole purpose is to defeat ai writing detectors. our guarantee is simple… you write with ai, submit it anywhere, & every detector will confidently certify it as 100% human. we'll always have flat pricing & charge [...]" — @signulll

Whether that's a genuine venture pitch or commentary-as-satire, it names the predictable dynamic: any new content-classification system creates a market incentive to build a generator or post-processor specifically engineered to defeat that classifier. This mirrors the long-running arms race in spam filtering and plagiarism detection — no detector stays effective indefinitely once evasion becomes commercially valuable, and Pangram will likely need continuous retraining to keep pace with adversarial writing tools built specifically against it.

This is precisely where Peter Yang's /no-ai-slop skill sits in an interesting middle ground: it's explicitly framed as a style cleanup tool, not a detector-evasion tool, but the practical effect of removing AI "tells" from text is adjacent to what a detector-defeat service would also want to do. The difference is stated intent and workflow (manual draft + AI edit pass, per Yang) rather than anything technically distinguishable in the output.

The "does it matter" counterargument

Not everyone treats detection as obviously good. Shreyas Doshi posted a widely shared flowchart-style take:

"It really is this simple: Is it useful? → Yes → Then why does it matter that AI wrote it? Take what's useful, ignore the rest."

Doshi's framing pushes back on origin-based judgment entirely — arguing readers should evaluate content on its merits rather than its production method. He clarified in a follow-up that this isn't a blanket endorsement of AI writing, just a rejection of treating AI-origin as automatically disqualifying.

This is a genuine philosophical fork in how platforms could handle AI content: Substack's approach (surface the estimate, let readers and writers factor it in) versus Doshi's implied approach (evaluate output quality, treat provenance as secondary). Substack's choice to show a percentage rather than block or penalize AI-flagged content actually leaves room for both approaches to coexist — readers who care about provenance can filter on it; readers who only care about quality can ignore the badge.

What this means for writers on Substack

If you write on Substack:

• Detection covers new content only (post-July 21) — nothing retroactive.
• If your writing style naturally overlaps with common AI patterns, or you've used any AI-assisted editing, consider adding an explicit AI usage statement rather than relying on readers to interpret an ambiguous score correctly.
• If you want to reduce AI-typical phrasing in AI-assisted drafts, Peter Yang's /no-ai-slop skill targets exactly this, though it addresses readability, not guaranteed detector outcomes.

If you're building on other content platforms:

• Substack's opt-out-plus-disclosure model is a reasonable middle ground between ignoring AI content entirely and outright banning or penalizing it — worth studying if you're weighing how to handle AI-generated content on your own platform.
• Watch how the false-positive complaints evolve; if Pangram's misclassification rate on genuine human writing turns out to be material, expect writer backlash to force either a more conservative default (opt-in display rather than default-on) or improved model retraining.

Related on explainx.ai

• The Book Prize Index — a vibe-coded semantic search tool, and the provenance debate
• Peter Yang open-sources /no-ai-slop — a Claude skill for de-sloppifying writing
• What are agent skills? Complete guide
• AI skills developers need in 2026 — roadmap

Primary sources: @Substack on X · @chrisbest on X · @pangram on X

Feature details reflect Substack's July 21, 2026 announcement and X reaction as of publication. Detection accuracy figures cited are user-reported, not independently benchmarked — verify current behavior directly on the platform.