Substack is adding a new layer of “reader verification” to its platform—one that’s aimed less at fact-checking and more at authorship signals. Starting with a rollout across the web and the iOS app, Substack will let readers scan posts, notes, replies, and comments to get an estimate of how much of the text could be AI-generated or written with AI assistance. The feature is powered by Pangram, an AI detection company, and it’s designed to be used quickly: for content longer than 100 words, readers can select “Scan for AI text” from the three-dot menu in the top-right corner of a post.
On paper, this sounds like another tool in the growing ecosystem of AI detection. In practice, it raises a bigger question about what platforms should do when the “who wrote this?” line becomes harder to trust. Substack’s move is notable because it’s not just aimed at creators publishing polished essays; it also reaches the messy, conversational parts of the platform—replies and comments—where AI-assisted writing can blend into normal human discourse. That matters, because the most convincing AI content often isn’t a single dramatic paragraph. It’s the accumulation of small stylistic choices: the cadence of a reply, the confidence of a comment, the way a thread keeps moving without anyone quite knowing why.
The tool’s scope is also broader than many people might expect. Substack isn’t limiting scanning to standalone articles alone. It’s positioned as something readers can apply to posts, notes, replies, and comments. That suggests Substack is treating AI detection as a general reading aid rather than a specialized feature for long-form publishing. And because the scan is available through the interface where people already engage—right inside the conversation—its impact could be more immediate than tools that require separate uploads or external links.
Substack’s implementation details are straightforward, but they hint at the product thinking behind the scenes. The scan is available for content longer than 100 words, which implies the underlying detection model needs enough text to produce a meaningful estimate. Short messages are notoriously difficult for detection systems because there’s simply not enough linguistic signal. By setting a threshold, Substack is likely trying to reduce the number of low-confidence results that would frustrate readers or create noise in discussions. The feature is rolling out across the web and iOS first, with Android coming “soon,” which is typical for platform features that depend on app updates and integration work.
What Pangram is doing, at least at a high level, is estimating the probability that a given piece of text was generated by AI or produced with AI assistance. Substack frames the output as an estimate of how much text could be AI-generated or written with AI help. That wording is important. It doesn’t claim certainty, and it doesn’t present itself as a definitive verdict. Instead, it treats AI detection as probabilistic—something closer to a “risk indicator” than a courtroom exhibit.
That distinction may be the key to understanding why Substack is choosing this approach. In the last year, AI detectors have been criticized for being unreliable, especially across different writing styles, languages, and editing workflows. People have pointed out that detectors can flag human writing as AI-generated, and they can miss AI-written text depending on how it’s been edited or prompted. Substack’s decision to present the scan as an estimate rather than a binary label could be read as an attempt to avoid the worst outcomes of overconfident detection: public accusations, reputational harm, and the spread of “detector says so” logic.
Still, even an estimate can change behavior. Once readers have a tool that suggests AI involvement, they may start reading differently. They may interpret tone, structure, and phrasing as suspicious rather than simply persuasive. They may also use the scan strategically—either to challenge a writer or to defend them. In other words, the tool doesn’t just detect; it shapes the social dynamics around writing.
This is where Substack’s unique environment matters. Substack is built around newsletters and communities that often feel more personal than traditional media. Readers don’t just consume content; they follow writers, join conversations, and develop expectations about voice. When AI enters the picture, it can disrupt those expectations. A writer’s style might drift, or a community might suddenly see a surge of comments that sound polished in a way that doesn’t match the usual rhythm. Even if the writing is technically “good,” it can feel off—like a familiar face wearing a slightly wrong expression.
Substack’s scanner can be seen as a response to that discomfort. It gives readers a way to check their instincts with a tool. But it also introduces a new kind of uncertainty: if the scan returns a high likelihood of AI involvement, what does that mean for trust? Does it imply dishonesty? Not necessarily. Many creators use AI for brainstorming, editing, translation, or formatting. Some use it to improve clarity. Others use it to generate drafts and then rewrite heavily. A detector that estimates AI assistance can’t easily distinguish between “AI helped a little” and “AI wrote most of it.” Substack’s framing—“how much text could be AI-generated or written with AI assistance”—acknowledges that nuance, but the reader still has to interpret the result.
There’s also the question of incentives. If readers can scan content, writers may adapt. Some may avoid certain patterns that detectors flag. Others may lean into transparency, explicitly stating when AI was used. Communities might develop norms: “If you use AI, say so,” or “Don’t rely on detectors; focus on sources.” Alternatively, some writers may treat the tool as a threat and become defensive, leading to more friction in comment threads. The scanner could reduce confusion for some readers while increasing suspicion for others.
Substack’s choice to integrate the scan directly into the reading experience is likely intentional. External tools tend to be used by a subset of users who already care about AI detection. By embedding scanning into the platform’s interface, Substack makes it part of everyday engagement. That could normalize the idea that AI detection is simply another reading feature—like spellcheck, grammar hints, or content warnings. But normalization cuts both ways: it can also make detection feel like a default authority, even when the underlying science is imperfect.
To understand why this matters, it helps to consider what “AI-written” means in real-world writing. Most AI-assisted content isn’t produced in a single click. It’s drafted, revised, rephrased, and blended with human intent. A detector might pick up on statistical patterns that correlate with AI generation, but those patterns can shift depending on the model used, the prompt style, the amount of rewriting, and the editor’s own habits. That’s why detectors often struggle with adversarial cases—when someone knows how detectors work and tries to evade them—or with benign cases—when a human uses AI tools for editing and ends up with text that looks “machine-like” to the detector.
Substack’s tool, powered by Pangram, is entering this messy landscape. The company’s decision to roll it out now suggests confidence that the tool can provide useful signal for readers, at least in aggregate. It also suggests Substack believes the cost of not offering any guidance is higher than the risk of imperfect guidance. In a world where AI-generated content is increasingly common, leaving readers without any tool at all can feel like ignoring the problem. Even a flawed tool can be better than none, provided it’s presented responsibly.
Substack’s rollout strategy also indicates a desire to learn. Launching across the web and iOS first allows the company to gather feedback, monitor performance, and refine the user experience before expanding to Android. It also gives Substack time to observe how readers actually use the scan. Do they scan every post? Only controversial ones? Do they scan replies during heated debates? The answers will shape whether the feature becomes a quiet background utility or a flashpoint in community culture.
There’s another angle worth considering: Substack is not a generic social network. It’s a publishing platform with a strong emphasis on individual voices and recurring writers. That means the platform’s value proposition depends heavily on trust and identity. If readers can’t reliably infer who wrote what, the relationship between writer and audience becomes more fragile. AI detection tools are one attempt to shore up that relationship. But they also highlight a limitation: detection is about text, not identity. A detector can estimate AI involvement, but it can’t confirm authorship in the way a verified account or cryptographic signature might. It can’t tell you whether a writer is the same person they were last year, or whether a newsletter has been taken over. It can only analyze language patterns.
So the scanner should be viewed as one piece of a larger puzzle. In the long run, platforms may need multiple layers of defense and transparency: provenance standards, disclosure requirements, watermarking or metadata approaches, and stronger account verification. Text-based detection is likely to remain part of the toolkit, but it’s rarely sufficient on its own. Substack’s move may be best interpreted as an interim step—an immediate, user-facing feature that addresses a pressing concern while the industry figures out more robust solutions.
For readers, the practical question is how to use the scan without turning it into a blunt instrument. The most constructive approach is to treat the result as a prompt for curiosity rather than a final judgment. If a scan suggests AI involvement, readers can ask: Is the writing consistent with the author’s past work? Are there sources, links, or evidence that support the claims? Does the piece show original reporting or firsthand experience? Does the author respond thoughtfully to questions in the comments? AI can mimic style, but it struggles with genuine specificity—unless someone supplies the details. So the scan can be a starting point for deeper evaluation, not the end of it.
For creators, the scanner introduces a new kind of audience expectation. Even if a writer uses AI responsibly—say, to improve grammar or restructure paragraphs—the detector might still flag the text. That could lead to misunderstandings.