Substack Adds Pangram AI Detection to Posts, Notes, and Comments
Substack’s new Pangram integration gives readers and creators a built-in way to estimate how much text was written by a person or with AI assistance. The feature is available for posts and Notes published on or after July 21, 2026, and can also scan individual comments and replies on Notes through Substack’s Reader and iOS app, according to Substack’s current support documentation.
For creators, the practical takeaway is simple: treat Pangram as a transparency signal, not as proof of authorship. You can add a public statement explaining how you make content, disable detection on individual posts or Notes, and report a suspected error. Readers should use the result as one piece of context alongside the writer’s disclosure, reporting history, sources, and editorial process.
What Substack announced
The launch moves AI-writing detection from a separate browser tool into the platform’s own reading and publishing flow. Substack users can open the three-dot menu on an eligible post or Note and select “Scan for AI text”; comments and replies can be scanned individually. In the iOS app, the same option appears from the content menu, while Substack says Android support is coming later. These availability details are listed in the company’s help center, updated July 21, 2026.
The feature was presented as a response to a trust problem rather than as a blanket rejection of AI. Coverage of the announcement quotes Substack CEO Chris Best describing the issue as a mismatch between what readers expect and how a piece was produced, while also noting that Pangram cannot determine every way AI may have participated in the writing process. That distinction matters: a detector can estimate signals in the finished text, but it cannot reconstruct the author’s complete workflow.
The timing is significant for the creator economy. Newsletters, paid posts, and Notes depend on a perceived relationship between a specific writer and an audience. When readers cannot tell whether an opinion, explanation, or personal-sounding comment came from that writer, the value of the subscription can weaken even when the text is grammatically polished.
What the result actually tells you
Pangram’s output should be read as a classification of writing patterns, not as a forensic record. Pangram describes its system as a classifier that maps writing according to many learned patterns and estimates whether a passage is AI-generated. It says no single feature, such as an em dash or a formal sentence structure, determines the result; longer text gives the model more material for a more stable estimate, as explained in Pangram’s explanation of its detection method.
That model is different from a tool that asks whether each next word looks statistically predictable. Pangram says it compares broader stylistic signals and can identify passages that appear mixed or ambiguous. The important editorial implication is that a percentage is not the same as a statement such as “the author used ChatGPT for 42% of the article.” It is an estimate about the text that the system received.
There is also a difference between generation and assistance. Pangram’s product documentation describes categories ranging from fully human-written to lightly AI-assisted, moderately AI-assisted, and fully AI-generated. Its examples of light assistance include grammar corrections, translation, phrasing, and readability edits; heavier assistance may include restructuring, adding details, or rewriting substantial sections. Those categories are useful for disclosure, but the boundary between them remains interpretive rather than a universal industry standard.
Where the feature works—and where it does not
Substack’s rollout is narrower than the phrase “AI detection on Substack” might suggest. The company currently lists the Substack Reader on the web, the iOS app, Notes, and individual comments or replies on Notes as supported surfaces. A Note with too little text may return “not enough text” instead of an analysis.
The transparency tool is not available for video or audio posts, posts viewed on standalone Substack sites or custom domains, or posts delivered by email. This creates an important gap between the platform’s central reading experience and the ways many subscribers actually consume newsletters. A reader opening a publication through an email or a custom domain should not assume that the same scan is available there.
Publication date is another limitation. The support page says the system works on posts and Notes published on or after July 21, 2026. Older archives therefore do not automatically become verified. That does not mean older writing is human or AI-assisted; it means the new in-product workflow does not cover it under the stated launch rules.
How creators can prepare their publications
The most useful response is to publish a clear process statement before readers have to infer one from a score. Substack lets a publication add a note under Settings, in the “How I make this” section. That statement can appear when readers scan a post, Note, or reply, giving the creator an opportunity to explain whether AI is used for research, translation, editing, outlining, drafting, or not at all.
A good disclosure should be specific enough to reduce ambiguity without becoming a legal disclaimer. For example, a creator might distinguish between correcting spelling and asking a model to produce a first draft. If the workflow changes by format—such as human-written essays but AI-assisted translations—say so. Readers usually need to know the role of the tool, not just whether a tool was present somewhere in the process.
Creators can also analyze a draft from the Publish page and disable detection from the report menu. On Notes, the equivalent control is available from the composer after selecting “Check for AI.” Once disabled, readers see an “AI detection unavailable” message rather than a Pangram analysis. This is a meaningful choice: it removes the automated estimate, but it does not replace it with a positive claim that the work was written entirely by a human.
Why a detector score should not decide a dispute
The central risk is false certainty. Pangram publishes strong performance claims for its own models, including a reported 99.98% accuracy figure for AI-generated text in its 2025 product update, but those figures come from the company’s own evaluation framework and should not be treated as a universal guarantee across every language, genre, length, editing history, or model.
Independent research also shows why evaluation conditions matter. A 2025 Findings of NAACL paper tested multiple AI-text detectors on unfamiliar models and domains and reported that performance could be poor at very low false-positive rates. The broader lesson is not that every detector fails equally; it is that a benchmark result must be connected to the exact decision being made.
For a Substack reader, the decision may be whether to investigate a suspicious post. For a publication editor, it may be whether to ask an author for clarification. Neither decision should rely on a single percentage. A scan can justify a conversation or closer review, but it cannot establish intent, plagiarism, factual accuracy, or ownership of an idea.
How readers should interpret a scan
Start by reading the label as a probability-like signal about the submitted text. Then look for corroborating evidence that does not depend on the detector: a creator’s disclosure, links to source material, consistency with the author’s established expertise, revisions over time, and willingness to answer specific questions.
Be especially cautious with short comments. Substack itself warns that text may be too short for an analysis, and short-form writing contains fewer signals for any statistical classifier. A polished sentence, a list, or conventional punctuation is not enough evidence by itself. Nor should a high estimate be used to publicly accuse a writer without giving them a chance to explain their process.
When a result seems wrong, Substack provides a “Report detection error” option inside the AI detection report. That feedback is intended to help improve detection quality, but it is not an appeals court and does not automatically resolve a dispute. The sensible reader workflow is to preserve the original context, contact the creator privately when appropriate, and avoid presenting an estimate as a confirmed fact.
What this changes for paid newsletters
For paid publications, the integration may make workflow disclosure part of the product experience. A reader is not only buying information; they are often buying judgment, taste, access, and a continuing relationship with a recognizable author. If AI is used to accelerate production, the commercial question becomes whether the creator has preserved the value that subscribers expected.
This does not create one correct policy for every newsletter. A market briefing may reasonably use AI for translation, formatting, or first-pass research while keeping analysis and verification human-led. A personal column may choose a stricter standard because the author’s voice is the primary product. The useful distinction is between assistance that supports the creator’s responsibility and automation that obscures who is accountable for the final work.
Creators should therefore document their workflow privately even if they do not publish every detail. Keep drafts, source notes, fact-checking records, and major revisions. These materials will not make a detector more accurate, but they provide a better basis for answering reader questions when an automated result conflicts with the creator’s explanation.
Common mistakes to avoid
The first mistake is treating “AI-assisted” as synonymous with “low quality.” Assistance can cover a spelling correction, a translation, a structural rewrite, or a fully machine-generated draft. Those are materially different workflows, and collapsing them into one label makes disclosure less useful.
The second mistake is assuming that disabling the scan proves human authorship. Substack’s interface makes the outcome explicit: readers see that detection is unavailable. That is a privacy or presentation choice, not an independent verification.
The third mistake is using the tool as a moderation shortcut. A platform or publication that removes content solely because of a detector result risks confusing authorship with misinformation, plagiarism, spam, or bad editorial judgment. Each of those problems needs its own evidence and policy.
Finally, do not overreact to a single false positive or false negative. Detector behavior can change as models, writing habits, languages, and editing tools change. Recheck the platform’s current documentation and Pangram’s published methodology before designing a permanent editorial rule around the launch version.
The practical next step for creators
As of July 22, 2026, the best operating policy is to combine disclosure, human accountability, and cautious interpretation. Add a concise “How I make this” statement, decide which uses of AI are acceptable for your publication, and make sure a named editor or author remains responsible for claims, sources, and the final voice.
- Write down your allowed AI uses, separating editing and translation from drafting and substantive rewriting.
- Add the policy to the publication’s transparency statement.
- Review eligible posts and Notes in the Substack Reader or iOS app to understand how the new workflow appears to readers.
- Use “Report detection error” when a result is plainly inconsistent with the text or your documented process.
- Handle disputes through evidence and dialogue, never through the score alone.
Substack’s Pangram integration makes authorship more visible, but visibility is not the same as certainty. The creators most likely to benefit are those who use the feature to explain their process and strengthen reader trust—not those who present an automated estimate as the final word on who wrote a piece.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.