Substack Launches AI Content Detection Tool for Posts and Comments

How It Works

Tap the “…” menu on any eligible piece of content and select “Scan for AI text.” Pangram’s classifier returns a breakdown: AI-generated, AI-assisted, or human. No guesswork, just percentages.

A few constraints worth knowing upfront:

• Only works on text longer than 100 words
• Only applies to content published after July 21, 2026
• Scanning is reader-initiated — nothing happens automatically

The detection is opt-in by design. Pangram doesn’t run in the background; a reader has to actively request it.

What Pangram Actually Does

Pangram isn’t using generative AI to catch generative AI. It deploys a classifier neural network trained on human-authored text from before 2021 — before LLM-generated content started saturating the internet. The idea is a clean training set, free of the very thing it’s trying to detect.

That said, the tool has real limits. It can’t tell if a writer used AI for research before drafting something themselves. As Substack CEO Chris Best put it, the problem isn’t AI assistance — it’s the mismatch between what readers expect and what they’re actually getting.

Controls for Creators

Writers aren’t just passive subjects here. They can:

• Add a disclosure statement explaining their writing process
• Run Pangram on their own drafts before publishing
• Flag incorrect classifications for review
• Disable AI detection on their content entirely

That last one is the interesting wrinkle. A creator can turn off detection regardless of what Pangram found. Substack is betting on transparency as a norm, not enforcing it as a rule.

Why This Matters Now

A recent study by digital marketing firm Graphite found that AI-generated articles now match human-written ones in volume online. That’s the context Substack is operating in — a platform built on the premise of authentic human voices, navigating a web where that’s increasingly hard to verify.

Best has signaled more AI-related features could follow, including reader preferences for the content Substack recommends — which could eventually mean filtering AI-generated content out of your feed altogether.

The Practical Takeaway

If you’re a reader, you now have a lightweight way to sanity-check what you’re reading on Substack — not perfect, but more than you had yesterday. If you’re a creator, adding a disclosure statement is a low-effort way to signal authenticity before anyone even reaches for the scan button.

Trust is easier to build proactively than to recover after a reader goes looking.