Substack’s Pangram Rollout Turns AI Detection Into a Reader Feature

Substack turned on an AI content detector across posts, notes, replies, and comments on July 21, 2026. Any reader can now check whether a piece of writing longer than 100 words was likely produced by a human, by AI, or by some mix of the two. The feature runs on technology from Pangram, an AI detection startup founded in 2024, and Substack co-founder and CEO Chris Best announced it in a post titled “Against Claudefishing.” It is one of the first times a major publishing platform has built AI-authorship scoring directly into the reading experience instead of leaving the job to browser extensions and third-party tools.

The timing is not subtle. Best’s post cites Pangram’s own estimate that as much as 40 percent of text on some social platforms is now AI-generated, and frames the new detector as an attempt to keep Substack from following the same path.

What Substack Actually Shipped

The mechanics are simple. From the three-dot menu on any post, note, reply, or comment longer than 100 words, a reader can choose “Scan for AI text” to get an estimate of how much of that writing was produced by hand versus with AI assistance, The Verge reported. The feature is live now on the web and iOS, with Android support coming soon.

Substack built this as an opt-in lookup rather than a label stamped on every piece of content by default. Best’s post says the analysis is shown “only to those who request it,” so a reader has to actively ask the question before seeing a score. That single design choice does a lot of work: it turns the detector into a tool a curious reader can reach for, rather than a public score attached to every writer’s byline by default.

Creators get a matching set of tools. They can scan their own drafts with Pangram before publishing, add a new “How I make this” statement to explain their process and set expectations up front, and report or remove any scan on their own work they believe is wrong.

The Problem Substack Says It Is Solving

Best’s post coins a term for the specific failure mode Substack is targeting: “Claudefishing,” the experience of investing attention in something that reads as a human perspective but has no human thought behind it. He is careful to separate that from AI use in general. “The core problem is not people using AI, or the quality of its output,” he writes. “The problem is when there is a mismatch between a reader’s expectation and reality, especially when they unwittingly invest their attention in something with no human thought on the other end.”

To make the point, Best quotes writer Freddie deBoer: “I access human-made art because I know there’s a human behind it and that’s what I’m looking for, other humans, showing me in art what they hide in their selves. Fooling me in that process is just a con.”

The post also takes a swipe at LinkedIn. Best writes that Substack does not want to wait until its own app “turns into LinkedIn” before acting, an implicit jab at how AI-generated posts have already reshaped that platform’s feed, and warns, “Platforms that reward fakeness will create a race to the bottom.” Substack’s argument ties back to its own business model: the company describes itself as “building an economic engine for culture,” a network of paid subscriptions that only works if readers trust there is a real person behind the writing they are paying for.

How Pangram’s Detector Actually Works

Pangram was founded in Brooklyn, New York, in 2024 by former Google machine learning engineer Max Spero and former Tesla machine learning scientist Bradley Emi. Pangram’s own site says its detector is built on natural language processing models trained across large sets of human and AI writing, using diverse datasets, hard negative mining, and active learning, rather than the older perplexity and burstiness scoring methods the company says “often fail in practice.” Pangram says its models recognize output from ChatGPT, Gemini, Grok, Llama, and Claude, cover more than 20 languages, and can flag text that has been run through “humanizer” tools built to evade AI detectors.

What the Accuracy Numbers Do and Do Not Cover

Pangram advertises detection accuracy of 99.98 percent or higher and says the figure has been verified by researchers at the University of Chicago and the University of Maryland, a third-party verification claim that has not been independently reviewed for this article. A professor’s customer testimonial published on Pangram’s own site, describing use in an academic-integrity setting, cites a false-positive rate of 1 in 10,000. That figure comes from a customer quote, not an audited company statistic listed alongside Pangram’s own headline accuracy claim.

Best’s own post is more careful than the marketing copy, and it is worth reading closely: “Pangram can only detect whether AI was used to make the text, not whether great human care went into creating it, nor whether AI tools were used as a source.” That is the load-bearing caveat in the entire launch. A detection score speaks to the statistical likelihood of AI involvement in production. It says nothing about whether a piece of writing is good, careless, honest, or worth reading, and Substack’s own framing goes out of its way not to conflate the two.

The Design Choices That Matter More Than the Accuracy Number

Any classifier deployed at platform scale runs into the same math problem: even a very low false-positive rate produces a meaningful number of wrongly flagged posts once it runs across millions of pieces of writing. Substack’s mitigations here are mostly structural rather than statistical. Scores are opt-in to view, not pushed to every reader by default. Writers can pre-screen their own drafts before anyone else sees a score. Writers can also dispute a scan on already-published work. None of that changes the underlying detector’s error rate, but it changes what happens when the detector is wrong, turning a mistaken flag into something a writer can contest rather than a silent, permanent mark.

Best is explicit that this is a first step, not a finished system. Substack says it is considering AI-content preferences inside its Reply Rules, more ways for creators to signal how they use AI, reader controls over what gets recommended to them, and continued work on the separate systems that fight spam, bots, and scams. He also draws a clear line against banning AI-assisted writing outright: “We’re not against people using AI to assist their work, and we think people should be free to choose which tools they use to express themselves.” Substack says it uses AI internally for writing software, doing research, and building product features like its clipping and translation tools.

Why This Matters Beyond Substack

Every platform that lets people publish text is going to run into the trust problem Substack just put a name on, and Substack is unlikely to be the last one to bolt a detector directly into the product instead of leaving the question to browser extensions and manual moderation. For teams building anything that scores user content with a probabilistic classifier, this launch is also a useful reference point for shipping one responsibly: state plainly what the score can and cannot tell you, make it something people opt into rather than something imposed on them, and give the people being scored a real way to contest a wrong call.

The harder problem is durability. Pangram’s own pitch includes catching text run through “humanizer” tools built specifically to evade detectors, which is itself an admission that detection and evasion are already in an arms race. Whatever accuracy figure holds true in July 2026 is not guaranteed to hold a year from now, so any product decision that leans on a detector score, including Substack’s own roadmap, needs a plan for what happens as that number moves. Best closed his own post more simply: “Claude has a lot to offer! But when I want Claude’s opinion, I’ll ask Claude.”