Substack is shipping an AI detection feature that can estimate how much text you are reading may be AI-generated or written with AI assistance. The platform says the tool can scan posts, notes, replies, and comments, and it is powered by an AI detection company called Pangram, according to a blog post published on Tuesday.

Here is how readers use it. For content longer than 100 words, you can choose “Scan for AI text” from the three-dot menu in the top-right corner of a post. Substack also says the tool is rolling out across the web and its iOS app, with an Android launch coming “soon.”

On the surface, this is a product feature aimed at curiosity. But it lands in the middle of a bigger trust problem that has been getting worse as AI writing becomes easier. The internet did not just gain a new way to generate text. It gained a new uncertainty layer. When readers cannot tell whether a post is authored by a person, an audience starts reacting differently, even when the writing is good. Platforms are now being asked to do something they never used to: provide signals about authenticity, not just content.

Substack is not doing this alone. It is using a detection vendor, Pangram, which signals a key operational reality. AI detection is a moving target, and building or validating a detector in-house is hard work. By integrating Pangram, Substack is effectively outsourcing the “signal extraction” step. That lets Substack focus on distribution, UI, and scalability, while Pangram handles the underlying model and detection logic.

There is also a distribution and product strategy embedded in the rollout plan. Substack is launching across the web and the iOS app first, with Android later. That stagger matters because detection tools need feedback loops: readers will experiment, publishers will complain or adapt, and the platform will learn where false positives and false negatives show up in real usage. Starting with the web and iOS can be a way to reduce rollout risk while the team watches how people actually use the scanning option from the three-dot menu.

Then there is the compliance-adjacent angle. The source does not claim Substack is making a legal determination or enforcing rules based on the scan results. Still, tools like this tend to create pressure in ecosystems where “AI transparency” expectations are rising. Even if regulators are not explicitly requiring detectors in every case, the mere presence of an AI scan changes how users interpret posts, and how platforms may later justify moderation or labeling decisions. Once a platform offers a mechanism to estimate AI involvement, the next questions usually arrive fast: What does the estimate mean? Who is responsible for disclosures? How should publishers respond if readers suspect AI assistance?

Second-order effects are where the boardroom attention usually goes. First, adoption of scanning can influence creator behavior. Authors and newsletter operators may start adjusting their workflows, either to avoid triggering scans or to meet whatever disclosure norms audiences increasingly demand. Second, reader trust dynamics could shift. If readers treat the scan output as an authority, publishers might face reputational risk from misunderstandings or imperfect detection. Third, advertiser and partner conversations can change, because brand-safety and authenticity narratives often get more formal once platforms add detection tooling.

For decision-makers, the strategic stakes are straightforward: this is not just a new button in a menu. It is an attempt to add a credibility layer to user-generated writing at scale. Substack can now mediate the AI ambiguity problem for a segment of the internet that values long-form thinking and close reader-writer relationships. If it does this well, it improves the reading experience by giving audiences a way to self-check. If it does this poorly, it can create friction, generate distrust, or amplify disputes about authorship.

In other words, Substack is taking a position in the AI era: the platform will help readers determine whether content might be written by AI. The tool is live across web and iOS, scans items longer than 100 words, draws on Pangram’s detection capabilities, and is coming to Android “soon.” For everyone building media, community, or creator platforms, that is a signal worth noticing. The question is no longer whether AI text will spread. It is how platforms will help the rest of us navigate what we are reading.