What does “AI detection unavailable” mean?

If a writer disables AI detection on a post or note, readers who use Scan for AI text in the Reader see that message, according to the help page. It is a setting, not a score, and it has to be set manually on each post published on or after 26 July 2026.

Does Substack train generative models on my posts?

The help page states that neither Substack nor Pangram uses publisher content to train generative AI models, and that a Block AI training toggle in Settings signals to external AI crawlers not to use your content for training. That is Substack’s wording as of the 26 October 2026 update, not a guarantee about every crawler on the open web.

TimeBinder

Start free

Substack’s Scan for AI text: what writers should keep

Inna Hall

November 12, 2026

Guides

Flat illustration of abstract document pages and two small meter chips on a cream card, beside a clock and a teal check-mark badge. No people, text or logos.

Short answer: Substack’s help page “How can I detect AI on Substack?”, updated 26 October 2026 and read on 12 November 2026, describes Scan for AI text, a tool that uses Pangram to estimate how much of a post or note is human-written or AI-assisted. It works on posts and notes published on or after 21 July 2026. Writers can add a “How I make this” statement, and can disable detection on a post or note so readers see “AI detection unavailable.” The tool is unavailable for video or audio posts, for posts viewed on standalone Substack sites or custom domains, and for emails. People have been searching and talking about this feature during the last two weeks; this page is about that current demand and the official wording, not a claim that Substack launched the scanner in November. Keep the dated draft you actually submitted and a screenshot of the scan or the unavailable message as it stood. A TimeBind cannot detect AI writing, change a Pangram percentage, or prove that a detector is right or wrong.

Last checked 12 November 2026. General information for writers and newsletter authors, not legal, editorial or platform-policy advice. The Substack help article cited below was updated 26 October 2026 and read on 12 November 2026; it is just outside the last-14-days news window and is cited as the current official page, not as a November launch. A detector score is not a finding that a person did or did not write a piece, and a dated file does not change that score.

What Substack’s help page says
Why dated drafts and scan records matter
Practical checklist for writers and publishers
What a TimeBind does and does not show
Limits
Tools that can help
Frequently Asked Questions
Sources checked

What Substack’s help page says

The official help article, “How can I detect AI on Substack?”, said when we fetched it on 12 November 2026 that it was updated 26 October 2026 at 18:11. That update sits just outside a strict 29 October to 12 November news window, so we treat the article as the current official wording, not as a November launch. The questions people are typing now include the “AI detection unavailable” message.

According to that help page, Scan for AI text estimates whether text was written by a human or AI-assisted. It works on posts and notes published on or after 21 July 2026, and uses Pangram to assign a percentage of how much of the text is human-written or AI-assisted. Substack states that neither Substack nor Pangram uses publisher content to train generative AI models, and that Substack offers a Block AI training toggle in publication Settings that signals to external AI crawlers not to use the content for training models.

The tool can be used in the Substack Reader on the web (open a post from Subscriptions, then the three-dot menu), in the Substack iOS app (Android coming soon, as of that page), on notes in the feed, and on individual comments and replies on notes. Notes with too little text show a “not enough text” message instead of an analysis. The help page also says the AI transparency tool is unavailable for video or audio posts, or for posts viewed on standalone Substack sites (such as on.substack.com or custom domains) or emails.

Writers have controls as well as readers. You can add a transparency note: in publication Settings, under Details, edit “How I make this.” That statement appears if readers scan one of your posts, notes or replies. You can also disable AI detection on a draft post from the Publish page; readers who then use Scan for AI text in the Reader see “AI detection unavailable.” The help page says the tool has to be disabled manually for each individual post published on or after 26 July 2026. Detection can also be disabled on a note from the composer, and you can report a detection error from the analysis report.

An optional reader setting, under Content preferences, can put a badge on posts and notes by other people when Pangram detects their text as 100% AI-generated, and a badge when the author has turned AI detection off. That is a reader preference, not a site-wide verdict, and it is limited to the 100% case described on the help page.

Why dated drafts and scan records matter

A Pangram percentage is a score from a vendor tool on a particular version of a text, in a particular surface, at a particular time. It can disagree with how you actually wrote, and it can change if the text or the tool changes. If a reader, an editor or a collaborator asks about a scan, it helps to still have the exact draft you published, the “How I make this” statement as it stood, and a screenshot of the analysis or of the “AI detection unavailable” message. None of those files proves the detector was right or wrong. They show what was submitted and what the tool displayed.

Demand in the last two weeks has been clear even though the official help page was last updated on 26 October. Google autocomplete in the United States on 12 November 2026 suggested “substack ai detector,” “substack new ai detector,” “ai detection unavailable substack,” “is substack ai detector accurate,” and several close variants. An in-window X search found 24 posts, with impressions summing to 16,351, including @jessesingal on 31 October (6,052 impressions), @JenCommon on 30 October (4,090) and @EliudTapia on 12 November (3,512). The YouTube videos we dated for this query were all older than 29 October and are supporting context only.

This is a different problem from being accused, off-platform, of using AI in a book or a prize entry, and from how text watermarks work. See how to show your writing history when AI is suspected, a manuscript workflow with drafts plus a dated record, and what AI text watermarks mean for writers. Those pages do not change a Substack scan.

Practical checklist for writers and publishers

Re-read the current Substack help article before you change a setting or argue with a score. The page we fetched was updated 26 October 2026. Save the wording you relied on, with the date.
Keep the exact draft you published, not only the live page. Scan for AI text, according to the help page, works on posts and notes published on or after 21 July 2026, and it is unavailable on standalone sites, custom domains, emails, and video or audio posts. The file you dated should be the text that was actually submitted.
If you use “How I make this,” keep a dated copy of that statement as it stood. You can edit it later. A screenshot or an export from Settings is easier to show than a memory of the wording.
If you disable detection on a post or a note, keep a screenshot of the “AI detection unavailable” message and a note of which post you disabled it on. The help page says this has to be done manually for each post published on or after 26 July 2026.
If you or a reader runs a scan, capture the analysis as it appeared, including the date and the surface (Reader, app, note or comment). A later rescan may not match. A screenshot is a record of a display, not a finding that the score is correct.
If you report a detection error, keep a copy of what you submitted. The help page invites feedback to improve detection quality. That report is not the same as a dated draft.
Do not treat a 100% badge, or the absence of a badge, as proof of how a piece was written. The optional feed label applies when Pangram detects text as 100% AI-generated, or when the author has turned detection off. Other percentages will not get that badge.
Date the pack you would actually send to an editor: the submitted draft, the statement or disable-detection note, and the scan screenshot. A TimeBind of that zip records that exact pack. It does not detect AI and it does not override Pangram.

What a TimeBind does and does not show

A TimeBind is a dated, checkable record that an exact file existed no later than a recorded time and still matches. The SHA-256 fingerprint is made in your browser and the file is never uploaded. The record is anchored to Bitcoin (a Merkle root in an OP_RETURN). Your first Pooled TimeBind is free; after that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. See how it works.

Here it is useful for a manuscript or newsletter draft, a Settings export of “How I make this,” and a screenshot of a scan or an unavailable message.

A TimeBind does not detect AI-assisted writing, does not change a Pangram percentage, does not prove a detector error, and does not replace Substack’s own controls. It is one piece of a timeline of the files you kept.

Limits

This article summarises Substack’s help page as updated 26 October 2026 and read on 12 November 2026, plus in-window search and X demand. It is not a review of Pangram’s accuracy, not a Substack policy change, and not a finding about any author. Feature availability (including Android) can change. We are not certifying any post as human-written or AI-assisted. This page is not legal or editorial advice.

Tools that can help

Use the bulk SHA-256 manifest for a folder of drafts, statements and screenshots, and keep a dated copy of the exact set you would show an editor.

Your file’s fingerprint, anchored to a global network secured by real electricity and cryptography.

Editing a file—including with AI—changes its fingerprint, not its existing network record.

How this works

Your file’s SHA-256 fingerprint, a short code worked out from its exact contents, is made in your browser, and the file itself stays on your device. A cryptographic commitment to that fingerprint, which may be combined with others, is then anchored in a block on the Bitcoin blockchain. Bitcoin is secured by proof of work: computers around the world spend real electricity to add each block, and changing an earlier block would mean redoing that work faster than the rest of the network. Until that block is confirmed, your proof is pending. Once the Bitcoin anchoring is confirmed, you can verify the proof against that recorded commitment. The confirmed record can then be checked; Pooled TimeBinds are checked through TimeBinder, while ASAP TimeBinds can also be checked independently. If you check the file later and its fingerprint still matches, that shows the file existed no later than the recorded time. The record doesn’t show who made the file, whether AI was used, or that its contents are true.

Your filestays on your device
FingerprintSHA-256
Network timestampon Bitcoin

Frequently Asked Questions

Did Substack launch Scan for AI text in November 2026?

No. The help page we read on 12 November 2026 was updated 26 October 2026, and it says the tool works on posts and notes published on or after 21 July 2026. This article is about current demand and the official wording, not a November launch.
Where will the scan not run?

The help page says the tool is unavailable for video or audio posts, for posts viewed on standalone Substack sites or custom domains, and for emails. Notes with too little text show a “not enough text” message.
What does “AI detection unavailable” mean?

If a writer disables AI detection on a post or note, readers who use Scan for AI text in the Reader see that message, according to the help page. It is a setting, not a score, and it has to be set manually on each post published on or after 26 July 2026.
Does a TimeBind prove I wrote the piece, or change the Pangram percentage?

No. A TimeBind is a dated, checkable record that an exact file existed no later than a recorded time and still matches. It does not detect AI writing and it does not alter a detector score.
Does Substack train generative models on my posts?

The help page states that neither Substack nor Pangram uses publisher content to train generative AI models, and that a Block AI training toggle in Settings signals to external AI crawlers not to use your content for training. That is Substack’s wording as of the 26 October 2026 update, not a guarantee about every crawler on the open web.

Sources checked

Substack help, “How can I detect AI on Substack?,” updated 26 October 2026, 18:11, https://support.substack.com/hc/en-us/articles/50891130623508-How-can-I-detect-AI-on-Substack (read 12 November 2026)
Google autocomplete, 12 November 2026 (US): “substack ai detector,” “substack new ai detector,” “ai detection unavailable substack,” “is substack ai detector accurate,” and close variants
X in-window search, 29 October–12 November 2026: 24 posts, sum 16,351 impressions, including @jessesingal (31 October, 6,052), @JenCommon (30 October, 4,090) and @EliudTapia (12 November, 3,512)

Date the draft and scan record you keep yourself

Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.

Inna Hall

Inna Hall is an Australian lawyer admitted to the Supreme Court of Queensland and the High Court of Australia in 2013. She has worked across government and private legal practice and is the director of TimeBinder, focusing on document authenticity, evidentiary risk, and blockchain-based proof of existence.