What Can the Pangram AI Detector on Substack Actually Tell You?

Authorship Is a process. Slop Is a systems problem.

In May 2024, Meta began to label images that it identified as having been “imagined with AI.” Shortly after the rollout, Australian photographer Peter Yan posted that a photo he’d taken of Mt. Fuji had been branded. Yan’s image wasn’t synthetic. It was a human-taken photograph, lightly edited in Photoshop using Adobe’s Generative Fill to remove a trash bin from the scene. But because the edit left AI metadata, Meta’s systems labeled the photograph: “Made with AI.” Yan’s photograph received the same label Meta had rolled out to address entirely generated images.

Meta had implemented the labeling program not because users were getting fed up with AI slop—though many were—but as part of an effort to address concerns about AI-generated deception. People were worried about bad actors generating images and videos to manipulate political discourse, or using the mere existence of synthetic content to undermine confidence in what was real.

The label on Yan’s photograph was technically correct: AI had been used somewhere in the editing process. But that wasn’t the question viewers thought the label was addressing. Most people interpreted “Made with AI” to mean an image had been generated. Meta was disclosing a true fact, but inviting a misleading inference.

I’ve been thinking about that incident this week as Substack rolled out Pangram’s AI detector on the platform.

Substack calls the problem it is trying to address “Claudefishing”: a mismatch between a reader’s expectation and reality, particularly when someone invests attention—or money—in work that involved little meaningful human thought. As generative AI has become more prevalent, many readers genuinely want to know whether the person they chose to read meaningfully wrote the thing in front of them.

That question is not without merit. Writing is an expressive act, and delegating a substantial portion of it to a machine genuinely matters to some people. Some people are concerned about copyright, or supporting what they see as unethical companies. Others simply want a pure human product—the heirloom tomato of content. These are authorship norms rather than firm integrity standards. But on a platform built around connecting writers and audiences, they are worth taking seriously.

Pangram is Substack’s attempt to translate a cultural norm into platform infrastructure. This is a familiar problem: a community develops a messy expectation about authenticity, integrity, authorship, civility, or truth, and a platform goes looking for something concrete it can measure. But the measurable signal is not always 1:1 with the value itself. In this case, detecting whether something is “Made with AI” is being used as a proxy for meaningful human authorship.

Labels are attractive because they collapse complex judgments into simple signals. Platforms often can’t reliably determine whether an image is deceptive, whether a publication is low-effort spam, or whether a writer meaningfully authored a piece. But they can attempt to detect whether AI was involved. The result is that we end up surfacing what is easiest to detect rather than addressing the deeper underlying concern.

What Are We Actually Measuring?

Pangram offers an estimate of how much of a text appears to have been written by hand, written with AI assistance, or generated by AI. Substack’s implementation is more nuanced than a badge automatically stamped on every post: a reader can request an analysis of text longer than 100 words published after the feature launched, and the result is shown only to the person who requested it (and who can, of course, screenshot and share it).

Accuracy matters here. A false result can damage a writer’s reputation, particularly if they have built a brand around artisanal human content. Pangram has plainly tried to address the failures associated with earlier AI detectors. It publishes model cards and says its latest system has a very low false-positive rate, including on long-form creative writing and text written by non-native English speakers. Those are company-reported results, and I am not auditing them in this post; I have personally experienced false positives, and anecdotally so have many Substackers. But Pangram is not ignoring the most prominent accuracy and fairness critiques of AI detection.

Still, even a perfectly accurate detector would not necessarily answer the question readers care about.

People bring at least two distinct kinds of concern to AI-generated content. One is about the artifact itself: Is it deceptive? Impersonating someone? Mass-produced? The other is about the relationship between creator and audience: Did the person whose work I chose to read meaningfully produce it?

Both are legitimate concerns. But they are not the same question, and a declaration that “AI was involved” does not address either.

A fantastical AI meme (“eating the pets”) and an altered documentary photograph (the arrest image of Nekima Levy Armstrong) use the same tools while posing entirely different risks of. A human-written content farm may be every bit as manipulative, repetitive, and low-effort as an automated one. The method of production tells us surprisingly little about what the artifact is doing, how it will be understood, or whether it deserves trust.

That was also the lesson of the Mt. Fuji photograph. Meta’s system detected the thing it could measure: AI metadata. Viewers processed the label as telling them something much broader about the authenticity of the image. The technical signal and the social meaning were discordant.

Pangram on Substack will likely create the same underlying problem in a different domain.

Authorship Is a Process

Text is more difficult to detect than other forms of generated content, and ‘AI-generated writing’ is not always a discrete category. LLMs are increasingly tools within the writing process: brainstorming, outlining, editing, finding a better verb, formatting a citation. I ask mine to interrogate my thinking, to check grammar or fact-check, and to suggest titles (I suck at titles). A finished essay may now contain contributions from a writer, an editor, a research assistant, software tools, and several different models. But professional published essays and journalism are rarely the output solely of the bylined writer, even if there is no AI involved. Ghostwriters, too, have existed for a very long time.

The AI detector evaluates the artifact produced at the end. It doesn’t actually know whether AI went into the research, or the outlining. Does the community norm of avoidance among those who dislike AI extend to those steps? It’s unclear. The point is, there are going to be false negatives as well as false positives.

Substack has, to its credit, launched more than the detector. It also now added a “How I make this” statement, where creators can explain their process and set expectations for readers. Writers can run Pangram on drafts before publication and report or remove scans they believe are mistaken.

In other words, Substack has paired a statement from the person who knows how the work was made with a numerical estimate from a system that does not.

But a number—say, 56% AI generated—feels like hard evidence while an author statement is a self-declared claim. “Meaningful authorship” isn’t an inherently discernible property of finished prose, it is a claim about a process. The author can describe how a tool was used, what decisions remained theirs, and whether they stand behind the result—but the adoption of “Made with AI” detectors creates something of a stigma all by itself, creating an incentive to minimize. The detector, meanwhile, cannot observe the writing process, know where the ideas originated, or discern whether the person whose name at the top exercised meaningful creative judgement but it has an aura of forensic precision.

Because the analysis appears only when a reader requests it, Pangram is not a public warning label. Major platforms have gone through several iterations in how they handle assertive provenance and public labeling at this point—two years after Mt. Fuji, even Meta is still figuring this out. These are hard questions! But Pangram will be abused: the person most likely to run it on someone else’s work is likely to be someone who has already decided that a writer cheated. The result of the scan can be screenshotted and circulated as quasi-forensic evidence about honesty, effort, or quality.

Substack is not wrong to build workflows around an authorship norm that unites a vocal percentage of its writers and readers. But it does remaining comparatively hands-off about the accuracy, toxicity, and political norms that divide them. That is not necessarily inconsistent. But it reveals which form of trust the platform believes is central to its product.

Slop is a Systems Problem

My main concern is that Pangram risks collapsing two different issues—authorship and slopishness—into a single technical signal.

I am neither an AI purist nor an unqualified cheerleader. I got researcher access to GPT-3 in 2020 and co-authored an article with it in the Atlantic about how LLMs were going to make it astonishingly easy to produce copious amounts of well-inflected propaganda. It’s aged well! Today I use Claude Code constantly, and have plenty of dedicated use cases for Gemini and Grok as well. These tools are useful. But, as Shrimp Jesus and AI slop have shown us, there is a meaningful difference between using machines to create something for people and producing content because other machines will reward it.

Slop is a systems problem. It’s what happens when cheap production meets distribution and monetization systems that reward volume, repetition, and engagement. It’s pollution. (My friend Aidan Walker disagrees and has argued that it’s folk art; his theories of slop are interesting.) We will continue to see low-quality content proliferate so long as it’s curated, amplified, and financially incentivized.

Pangram may catch some slop, but it’s a multi-faceted issue. Provenance systems and content credentials can provide information about origin, although they remain incomplete or vulnerable to manipulation. Author disclosures can help establish expectations around meaningful human authorship. Spam detection, ranking controls, and changes to monetization programs (hello, X) can reduce the pollution.

Substack is responding to a real demand. Readers want to know what they are paying for, and writers want a way to distinguish their work from content produced with little meaningful human thought. The Pangram integration is an attempt to translate a messy cultural norm into platform infrastructure. But by focusing on individual posts rather than the systemic incentives that reward volume over meaning, detectors address the symptom while leaving the disease untouched. Relying on detection to solve a slop issue doesn’t just fail to bridge the trust gap; it risks institutionalizing a ‘gotcha’ culture that devalues the very relationship between writer and reader it purports to serve.