Substack’s AI author’s note and Pangram: a publisher workflow

Substack’s optional AI author’s note and its Pangram scan do different jobs. The note explains your process; the scan estimates whether text is human-written or AI-assisted. Neither replaces editorial verification.

For a publisher, the practical response is to disclose actual tool use, decide whether scanning should remain enabled, and retain enough drafting evidence to investigate a disputed result. Substack told TechCrunch at launch on July 22, 2026 that the feature was intended to encourage transparency, not prohibit or penalize AI-assisted writing.

What readers can see

Substack calls its author’s note “How I make this.” It is a publication-level statement that appears when readers scan your posts, notes, or replies—not a disclosure automatically inserted into every email. The scan itself produces an estimated percentage of human-written or AI-assisted text, according to Substack’s help guide, updated August 9, 2026.

That guide says scanning works on posts and notes published on or after July 21, 2026. On the web, readers open a post through Subscriptions in the Substack Reader, select the three-dot menu, and choose Scan for AI text. Notes, comments, and replies also have scanning support, though notes with insufficient text return a “not enough text” message.

The same guide excludes audio and video posts, emails, and posts viewed on standalone publication sites or custom domains. It also lists Android as “coming soon,” despite Substack’s launch follow-up saying Android was available. Check the actual app before promising readers a particular mobile workflow.

Add a note that describes the work

To configure the statement, go to your publication’s Settings → Details, click Edit beside How I make this, enter the text, and select Save statement. These are the steps in Substack’s help guide.

A useful statement answers three questions:

• Where does AI enter? Research assistance, transcription, outlining, drafting, translation, or editing?
• What does the author or editor do? Choose the argument, check sources, rewrite passages, and approve publication?
• Who owns corrections? Name the person or editorial role responsible for errors.

Hypothetical example—use only if it matches your workflow:

I write the initial draft and use AI to suggest structural edits and clearer phrasing. I check factual claims against the cited sources and approve the final text. Errors and corrections remain my responsibility.

If a model produces the initial draft, say that instead. “AI helps with editing” is not an accurate description of AI drafting followed by human editing. If practices vary substantially between articles, add a short disclosure in the affected article as well, so email and standalone-site readers receive the relevant context.

For a repeatable production process, pair disclosure with a controlled AI-assisted publishing workflow: sourcing, verification, and final approval should remain explicit tasks.

Decide whether to leave scanning enabled

Substack lets publishers scan drafts before publication and disable scanning per post. Its follow-up announcement says publishers can disable post scanning without first running a scan or interacting with Pangram. The help guide directs web publishers to Disable AI detection on the draft’s Publish page.

Disabling detection does not display a “human-written” result. Readers attempting a scan see “AI detection unavailable.” The control must be applied individually; do not treat it as a publication-wide default. For notes, the guide directs publishers to open the composer’s three-dot menu, select Scan for AI text, then open the analysis menu and select Disable detection.

A proposed operating policy:

• Set a consistent scanning preference rather than changing it to obtain a favorable score.
• Use a draft scan, if desired, to anticipate reader questions—not as a publication gate.
• Require source checks and editorial approval regardless of the result.

Pangram describes its system as a classifier making statistical predictions from text alone. A scan is therefore not a record of who supplied an idea, which sources were checked, or how a draft changed. Do not rewrite accurate prose merely to pursue a more reassuring label.

Handle errors and privacy separately

For a disputed result, preserve the draft history, source notes, and scan output. Substack’s documented reporting route is Report detection error on the analysis report, followed by feedback. Publishers can also disable detection. Explain the process to readers without presenting either the score or your disagreement as conclusive proof.

Substack says neither it nor Pangram uses publisher content to train generative AI models. Pangram’s privacy explanation separately says submitted data is not used to train or improve its models, but describes keeping customer history until deletion is requested. That general customer-history policy does not establish how long Substack-integrated scans are retained. “Not used for training” should not be read as “never stored.” Keep confidential source material out of optional scans unless its processing has been approved.