Substack's Answer to AI Slop: Pangram Detection and Transparency Labels
Substack is adding Pangram-powered AI text detection and transparency labels to fight AI slop. Here's how the system works and why it matters.
What is Substack doing about AI slop?
Substack has built an integration with Pangram, an AI text detection company, directly into its app. Readers and writers can now scan a piece of long-form writing and get an estimate of whether it was generated by a large language model. The feature doesn’t block, flag, or punish anything by default. It’s opt-in: you have to actively run the scan to see the result. Substack is also adding a “how I make this” option so writers can voluntarily explain their process, whether that means dictating a first draft, editing with AI tools, or writing everything by hand. CEO Chris Best described it as a first step toward transparency, not a policy that dictates what tools creators can or can’t use.
TL;DR
- Substack has integrated Pangram’s AI detection tool into its app, letting anyone scan long-form text and see an estimate of whether it was generated by an LLM.
- The feature is opt-in and non-punitive: there’s no automatic labeling, ranking penalty, or ban tied to a high AI-detection score.
- Chris Best draws the line at intent, distinguishing between people who use AI as a tool within their own thinking and people who mass-produce content to game a system without engaging with it at all.
- Pangram found that roughly 40% of long-form writing on LinkedIn showed signs of being fully AI-generated, a stat Best cited as the wake-up call behind Substack’s move.
- A new “how I make this” field lets writers voluntarily disclose their process, from dictation tools to AI editing passes to fully manual writing.
- Detection tools like Pangram work by identifying statistical patterns common to LLM output, which means they can be gamed by fine-tuned models designed to evade detection.
- Best frames the effort as protecting the “public square”, arguing that unchecked slop functions like a denial-of-service attack on trust between readers and writers.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Why is AI slop suddenly a platform problem?
Slop, in Best’s framing, isn’t a technical category. It’s content nobody actually believes in: thoughtless, generic, or optimized purely for reach rather than communicating something. That definition covers old-school spam and clickbait as much as it covers AI-generated filler. What’s changed is scale. Generating a thousand plausible-sounding blog posts or LinkedIn updates used to take real effort. Now it takes a prompt.
Best pointed to Pangram’s own measurement of long-form writing on LinkedIn, where the company estimated that about 40% of it was fully AI-generated. He was careful to note that this isn’t a clean proxy for quality: not all AI-assisted writing is slop, and not all slop involves AI. But a rapid shift in the ratio of human-written to machine-written text across major platforms is, in his view, a structural change that every platform will eventually have to confront, whether or not they want to.
The effect shows up as a kind of trust tax on readers. Best described people who’ve stopped reading new writers or stopped engaging in comment sections because they can no longer tell if they’re talking to a person. That uncertainty erodes the willingness to discover anything new, which is a direct threat to a platform built on writers building audiences from scratch.
How does Pangram’s detection actually work?
Pangram and similar AI-text detectors work by identifying statistical fingerprints in how large language models generate text. Because of how these models are pretrained and then fine-tuned with reinforcement learning, they tend to converge on particular patterns of word choice, sentence rhythm, and phrasing, sometimes called an LLM’s stylistic “peak.” Detectors are trained to recognize that signature.
This approach has a known weakness: it’s steerable. Fine-tuned models built specifically to sound more human can push detection scores down, and prompting techniques can push output away from the most recognizable LLM patterns. Best acknowledged this directly, noting that detection accuracy is “reasonably good” rather than definitive. Pangram’s tool doesn’t claim to know whether a piece of writing was made with care or whether AI was involved in some smaller way, like research or light editing. It answers a narrower question: does this text show the statistical signature of having been generated by an LLM as it was written?
That narrowness is deliberate. Substack isn’t using the tool to make a final judgment about quality or authenticity. It’s surfacing one data point and letting readers and writers interpret it themselves.
Why draw the line at intent instead of tool use?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Best repeatedly rejected the idea of policing which tools people use to write. His dividing line is different: whether someone is using AI within their own thinking and editorial judgment, or using it to skip the step of having anything to say at all. The extreme example he gave is telling an AI assistant to generate a large batch of viral-style posts with no oversight and no real point of view behind them. That kind of output, in his framing, functions less like writing and more like a denial-of-service attack on a platform’s discourse, because it forces every reader to spend energy figuring out whether anything they’re reading is genuine.
By contrast, using AI as part of a writing process, dictating a rough draft, refining phrasing, checking research, still counts as legitimate work if there’s a person driving intent and thinking behind the result. Best described his own workflow as an example: producing long unbroken dictation sessions to capture an initial idea, then iterating with AI tools to sharpen clarity and cut generic phrasing, while the underlying thesis and structure come from his own judgment throughout.
Is AI detection enough to fix the problem?
Detection alone doesn’t solve slop, and Best didn’t claim it would. He was explicit that this is a first step, not a comprehensive policy. The goal is to open a conversation rather than close one down: give people a way to check what they’re reading, give writers a way to explain their process, and let norms develop from there rather than dictating rules from the top down.
There’s an implicit bet in this approach. Best suggested that responsible AI use should be able to withstand transparency, meaning writers who are comfortable disclosing their process have nothing to hide, while those gaming the system for reach have an incentive to avoid disclosure altogether. Whether that social pressure is enough to hold back a flood of low-effort AI content remains an open question, especially since detection tools can be evaded by anyone motivated enough to fine-tune around them.
The stakes, in Best’s view, are less about any single feature and more about avoiding a tipping point. He raised the possibility that a platform could “wake up” one day resembling the parts of the internet already saturated with generic AI content, and that once that shift happens, it’s much harder to rebuild reader trust than to protect it early.
Frequently Asked Questions
What is AI slop?
AI slop generally refers to low-effort, generic, or mass-produced content, often but not always generated by AI, that isn’t backed by real thought or intent. The term covers everything from AI-written spam to clickbait designed purely to game engagement or search algorithms.
How does Substack’s Pangram integration work?
Users can run a scan on a piece of long-form text within the Substack app to get an estimate, based on Pangram’s detection model, of whether the text appears to have been generated by a large language model. The scan is opt-in and doesn’t automatically label or penalize content.
Can AI detection tools be fooled?
Yes. Detection tools like Pangram identify statistical patterns common to LLM-generated text, but models fine-tuned specifically to write in a more human-sounding style can lower their detection scores. Accuracy is described as good but not perfect.
Does Substack ban or penalize AI-generated writing?
No. The Pangram integration doesn’t restrict what tools writers can use. It’s designed to add transparency, giving readers a way to check content and writers a way to disclose their process voluntarily.
Why does Substack care about this specifically?
One coffee. One working app.
You bring the idea. Remy manages the project.
Substack’s business model depends on writers building trust and paid audiences over time. If readers can’t tell whether content is genuinely written by a person, that trust erodes, which threatens the platform’s core value proposition as a home for authentic, independent writing.
