What happened
On July 21, Substack CEO Chris Best published a post announcing a new Pangram integration inside the Substack app. The feature lets readers scan text and see an estimate of how much was likely written by hand or with AI assistance. Best framed the problem as a mismatch between what readers expect and how something was actually made.
Substack's July 23 help documentation gives the operational details. The Scan for AI text tool works on posts and notes published on or after July 21, 2026, and uses Pangram to assign percentages for human-written and AI-assisted text. The help page says it is available in the Substack Reader on the web, the iOS app, notes, comments, and replies, with Android coming soon.
There are clear limits. Substack says the tool is unavailable for video or audio posts, posts viewed on standalone Substack sites or custom domains, and emails. The same help page says creators can add a How I make this statement, scan drafts before publishing, disable detection on a post or note, and report a detection error.
Axios covered the rollout on July 23 and corroborated the core product move: Substack is partnering with Pangram so users can estimate how much text on the platform was made with AI. Pangram's own explainer, published July 17, describes its detector as a classifier that estimates whether a segment of writing is AI-generated based on learned text patterns.
What this is not
This is not a universal rule for every publisher, and it is not a ban on AI-assisted writing. Substack's announcement explicitly leaves room for people who use AI thoughtfully and still stand behind the work.
It is also not proof that a detection score can answer the authorship question by itself. Pangram says its model estimates whether text is AI-generated. Substack's announcement says Pangram cannot tell whether human care went into creating a piece or whether AI was used as a source. That distinction matters.
A detector looks at the finished text. It does not see the interview, the notes, the draft history, the source material, the editorial judgment, the rejected claims, or the approval conversation. In expert communication, those are often the parts that matter most.
REC's read: the process note is the real product signal
The most useful part of Substack's rollout may not be the percentage. It may be the How I make this field.
That field points to a more durable trust pattern. Readers do not only want to know whether AI touched a sentence. They want to know what role AI played, what the named person contributed, and whether the final claim still belongs to someone accountable.
For a researcher, founder, consultant, educator, or creator, a useful process note can be simple: AI helped gather background, organize questions, transcribe the recording, suggest clips, or tighten a draft. The human supplied the examples, caveats, interpretation, approval, and responsibility for the published claim.
That is stronger than pretending AI was absent when it was present. It is also stronger than treating any AI assistance as disqualifying. The issue is role clarity. Did AI prepare and organize the work, or did it impersonate the author's judgment?
Why video interviews change the trust equation
Substack's current detection tool is text-only. That makes the REC angle more specific, not less. As platforms add labels and detectors for different media types, teams still need a source that explains where the claim came from before it was packaged.
A research-guided video interview creates that source of record. The person answers on camera. The transcript captures the words. The editor can trace a clip, article paragraph, caption, or newsletter section back to the original answer. If AI helped prepare the question or organize the result, that role can be described without making AI the author.
This matters because short public assets lose context quickly. A clip may travel without the full article. A quote may be pasted into a post. A summary may be rewritten for another channel. The internal source trail protects the meaning when the format changes.
The practical advantage is not that video is magically more authentic. People can perform on camera, and video can be edited badly. The advantage is that a recorded answer gives the team a primary source to inspect, approve, and revisit. That is harder to get from a polished text draft that arrived with no visible path behind it.
A practical process disclosure
If your team uses AI in expert publishing, write the process note before you need one. Keep it factual and boring.
Start with source material. What did the human actually provide: a recorded answer, a transcript, a document, a product note, a research link, a firsthand example, or an approved position? If the central claim cannot be traced to any of those, it is not ready for publication.
Then name the AI role. Did AI research context, draft questions, summarize notes, suggest structure, identify clips, create captions, or help with editing? The more specific the role, the easier it is for a reader, editor, or stakeholder to understand what happened.
Finally, name the human review. Who checked the claim, the implication, the sources, and the final wording? Approval should cover what the audience is likely to believe, not only whether a sentence sounds acceptable.
That kind of disclosure will not satisfy every reader, and it does not remove the need for good work. But it answers the question a detector cannot answer: who shaped this, from what source, and who stands behind it now?
The takeaway
Substack's Pangram rollout is a timely signal because it treats authorship as something platforms now have to design around. The reported feature is specific: text scanning, creator statements, draft checks, opt-outs, and error reports.
REC's analysis is that detection is only one layer. For expert-led publishing, trust depends on the visible chain from source to claim. A percentage can describe a text pattern. A process note can describe the work. A recorded interview can preserve the human judgment that the work is supposed to carry.
The practical standard is simple: use AI where it helps prepare, organize, and adapt. Keep the person's real answer upstream of the public claim. Then make the process plain enough that readers do not have to guess what they are getting.