Substack is rolling out a new reader-facing feature that aims to answer a question many people have started asking—sometimes quietly, sometimes angrily—about the newsletters they consume: how much of this was written by a human, and how much was produced with the help of AI?
The tool, reported as AISubstack, is designed to let readers estimate the portion of a newsletter that may have been written using AI. In other words, it’s not just another “disclosure” checkbox or a vague promise that a creator “used tools.” Instead, it attempts to provide a more measurable signal, giving audiences a way to calibrate trust and expectations before they invest time in a post.
This matters because the newsletter ecosystem has always been built on a particular kind of intimacy. Readers don’t just subscribe to topics; they subscribe to voices. They expect a certain texture—idiosyncratic phrasing, selective emphasis, the occasional digression that reveals how a person thinks. When AI enters the picture, even when it’s used responsibly, it can subtly change that texture. The shift isn’t always obvious at first glance. That’s part of why platforms are now moving toward transparency mechanisms that go beyond self-reporting.
AISubstack represents a broader trend: platforms trying to mediate the AI question for their users. For years, the debate has largely been framed as a creator responsibility issue—“label it if you used AI,” “don’t mislead,” “be honest about assistance.” But as AI writing becomes easier to deploy and harder to detect reliably, the burden of disclosure increasingly collides with a practical reality: many readers don’t know what to look for, and many creators don’t want to be forced into binary categories that don’t reflect how writing actually happens.
In practice, writing with AI often looks like a spectrum. Some creators use models to brainstorm outlines, refine tone, or translate drafts. Others use AI to generate entire sections and then heavily edit. Still others rely on AI as a first draft engine and do minimal rewriting. The result is that “AI-assisted” can mean anything from light support to near-total generation. A tool like AISubstack is essentially an attempt to convert that messy spectrum into something closer to a readable signal for audiences.
What makes the move notable is that Substack is positioning the feature as transparency rather than enforcement. There’s a difference between a platform saying, “Here’s what we think is likely AI-generated,” versus a platform saying, “This violates our rules.” The former is informational; the latter is punitive. By framing AISubstack as an estimation tool, Substack is leaning into the idea that readers should have more context, not that creators should be automatically judged.
That framing also hints at a strategic balancing act. Substack’s business model depends on creators, and creators depend on trust. If the platform were to treat AI detection as a strict compliance mechanism, it could quickly become adversarial—especially given the known limitations of detection systems. Even the best approaches can produce false positives, and the stakes for creators are reputational. A reader-facing estimator can still influence perception, but it does so without immediately turning into a legalistic verdict.
Still, the feature raises immediate questions about how such estimates will be interpreted. Readers may treat the output as a rough indicator of authenticity, but authenticity is not the same thing as authorship. A newsletter can be “human-written” in the sense that a person authored the final text, while still being shaped by AI suggestions. Conversely, a newsletter can be “AI-assisted” while remaining deeply original in reporting, analysis, and perspective. The tool’s value will depend on whether readers understand it as a probabilistic signal rather than a definitive audit.
There’s also the question of what exactly the estimator is measuring. AI writing detection is notoriously difficult because modern language models can mimic human patterns extremely well, and because human writing itself varies widely across individuals and contexts. Some newsletters are short and punchy; others are long and essayistic. Some writers use rhetorical repetition; others write in clean, structured prose. If AISubstack is based on stylistic cues, it may correlate with certain writing behaviors that are not uniquely AI-related. If it’s based on metadata or submission patterns, it may be more accurate but also more complex to implement and explain.
Even without knowing the underlying method, the direction is clear: Substack is acknowledging that the AI question is now part of the reading experience. That’s a meaningful cultural shift. For most of the internet’s history, readers have had to infer credibility from content itself—sources, reasoning, consistency, and track record. Now, platforms are adding a new layer of “credibility instrumentation,” where the interface itself provides a hint about how the text might have been produced.
This is where the tool becomes more than a single feature. It’s a signal that the next phase of AI governance may be less about policing and more about surfacing. Instead of waiting for legislation or relying solely on creator disclosures, platforms may start embedding interpretive aids directly into the product. Think of it as the evolution of content labeling: first came “sponsored” tags, then “fact check” modules, then “read time” and “content warnings,” and now potentially “AI assistance estimates.”
For readers, that could be empowering. Many people don’t want to become amateur forensic analysts. They want a quick, understandable cue that helps them decide whether to read critically, skim, or trust. But there’s also a risk: a tool like this could encourage a new kind of superficial judgment. If readers see a high estimated AI portion, they might dismiss the piece entirely—even if the reporting is solid and the argument is thoughtful. Meanwhile, if the estimate is low, readers might over-trust the content, assuming “human” automatically means “truthful.”
The deeper issue is that AI assistance doesn’t inherently determine truthfulness. It can affect clarity, coherence, and style, but it doesn’t guarantee accuracy. A human can write misinformation just as easily as an AI can. A model can help structure an argument, but it can also hallucinate facts. The real question for readers remains: what evidence supports the claims, and how well does the author reason from that evidence? AISubstack can’t replace that critical thinking. It can only add another dimension to the decision-making process.
For creators, the feature introduces a new kind of pressure: not just to be transparent, but to anticipate how transparency will be interpreted. Some creators may welcome the tool because it could validate their honesty—if they use AI lightly and the estimate reflects that, they can point to the signal as proof of responsible practice. Others may worry that the estimator will penalize certain writing styles or workflows. If a creator uses AI to polish language, the tool might interpret that polishing as a larger share of AI involvement than the creator intended.
There’s also a creative dimension. Newsletter writing is partly about voice, and voice is partly about habits. If AI assistance changes those habits—even slightly—then the estimator could become a feedback loop. Creators might adjust their writing to “look more human” according to whatever the tool measures. That could lead to a subtle homogenization of style, where writers optimize for detection metrics rather than for authentic expression. The irony is that a transparency tool meant to clarify authorship could inadvertently shape it.
At the same time, the existence of AISubstack could push the industry toward more nuanced disclosure norms. Instead of a binary “AI used / AI not used,” creators may begin to describe how they used tools: brainstorming, editing, translation, summarization, fact-checking support, or drafting. Even if the platform’s estimator is imperfect, it could encourage better conversations about process. Readers might ask not only “was AI involved?” but “how was it involved?” That would be a healthier direction for the ecosystem.
Another angle worth considering is how this affects the economics of attention. Newsletters compete for subscriptions, and subscription decisions are emotional as well as rational. A reader might subscribe because they feel a writer is “one of us,” someone who thinks clearly and speaks honestly. If AI assistance becomes a visible factor, some readers may feel betrayed if they believe the voice is synthetic. Others may feel relieved if the tool makes the situation clearer and reduces uncertainty. Either way, the feature changes the emotional calculus of subscribing.
It also changes how newsletters are compared. Previously, two newsletters on similar topics might be judged primarily on content quality and perspective. With AISubstack, readers may start comparing not just arguments but production signals. That could create a new competitive advantage for creators who either avoid AI or use it in ways that the estimator interprets as minimal. But it could also create a disadvantage for creators who use AI for accessibility—such as rewriting for clarity, translating into other languages, or adapting for readers with disabilities. If the estimator doesn’t account for those legitimate uses, it could unintentionally punish inclusive practices.
Substack’s move can be read as a response to a growing expectation among audiences that AI should be disclosed in some form. But it’s also a response to a more practical problem: disclosure alone is often unverifiable. If a creator says “I didn’t use AI,” readers have no way to confirm. If a creator says “I used AI,” readers may not know whether that means “light editing” or “full generation.” An estimator tries to bridge that gap, even if it can’t eliminate uncertainty.
The key word here is “estimate.” Any system that claims to measure AI involvement is operating under uncertainty. The best it can do is provide a probability-like signal. That means the tool’s design and communication will be crucial. If Substack presents the output as a percentage or a score, it must also communicate what that number means and what it doesn’t. Readers need to understand that the tool is not a lie detector. It’s a heuristic.
If Substack gets that messaging right, AISubstack could become a useful part of the reading interface—like a “nutrition label” for writing process. If it gets it wrong, it could become a weapon