I watched Max Spero’s TechCrunch Equity interview twice. I watched Max Spero’s TechCrunch Equity interview twice. Not because I couldn’t follow it the first time — but because the way he framed AI detection changed how I think about the whole problem Pangram research page explains their methodology..
Spero, CEO of the AI detection startup Pangram, kept repeating a phrase that stuck with me: “It is not a problem that gets solved once.” Every improvement in generation makes detection harder. You don’t ship a detector and walk away. You keep training, keep testing, keep chasing the models that keep getting better at sounding human.
That’s the part most people miss. They think of AI detection like a spam filter — you train it, it works, done. But AI-generated text isn’t spam. It isn’t a fixed category with a fixed signature. It’s a moving target that gets redefined every time a frontier model ships.
Here’s what I learned from digging into how Pangram actually works — and why the old “perplexity and burstiness” approach that so many detectors still use is basically useless.
The Wrong Way: Perplexity and Burstiness
If you’ve ever seen an AI detector claim that the Declaration of Independence is 100% AI-generated, you’ve seen perplexity-based detection fail in the wild.
Here’s why. Perplexity analysis works by asking: “How surprised would an LLM be to see this word next in this sentence?” If the text consistently follows the model’s expected word patterns, the detector flags it as AI-written. It sounds reasonable on paper.
The problem is that LLMs read the Declaration of Independence thousands of times during training. Every word in it is maximally unsurprising to the model. A perplexity detector sees that predictability and calls it AI-generated. It’s confusing “familiar text” with “AI-generated text” — and that’s a fundamental category error.
Burstiness — the idea that human writing varies more in sentence length and structure than AI writing — has the same weakness. Good human writing can be consistent. Bad AI writing can be chaotic. The signal is nowhere near reliable enough for high-stakes detection.
Pangram’s answer to this is straightforward: stop reverse-engineering how LLMs produce text and start asking who the text sounds like.
The Right Way: Stylistic Mapping and Synthetic Mirroring
Pangram’s core insight is that AI detection is fundamentally a stylistic problem, not a statistical one. You don’t ask “what would an LLM write here?” You ask “who does this sound like?”
The company trained its model on tens of millions of known human-written documents drawn from 2021 and earlier — before ChatGPT launched and polluted the internet with AI-generated content. That data hygiene matters. You can’t train a detector on data that might already contain AI writing and expect clean results.
Then comes the clever part: for every human document in the training set, Pangram generates a “synthetic mirror” — an AI-written piece that matches the original’s topic, length, and tone as closely as possible. The model learns by comparing human-and-AI pairs that are as similar as possible except for one thing: who wrote them.
Think of it like training to tell apart two musicians who play the same song in the same key at the same tempo. The notes are identical. The style is what gives them away. Pangram’s model learns the stylistic fingerprints that distinguish human choice from AI probability — the subtle patterns in word choice, rhythm, and structure that persist even when the surface content looks the same.
This is why Pangram claims over 99% accuracy on AI-assisted writing and mixed human-AI content — and why it can catch text that’s been run through AI humanizer tools. The stylistic signature doesn’t disappear just because someone asked a model to “make this sound more human.”
The Spectrum Problem Nobody Talks About
Here’s the thing that makes AI detection genuinely hard, and the part Spero’s interview drove home for me: the question is almost never “was this written by AI or not?” It’s almost always “how much AI was involved, and what kind?”
A writer might draft a paragraph themselves, then ask an LLM to clean up the grammar. Another might write an outline and have AI expand it. Someone else might generate a full draft and rewrite it section by section. A fourth might paste in a fully AI-generated piece with two swapped words.
These are not the same thing. They’re not even on the same axis. And yet most detectors give you a single percentage and call it a day.
Pangram is conservative by design. According to their published data, human essays that have been heavily rewritten by AI still come back as human-written roughly 41% of the time WIRED got an early look at Pangram detection claims.. That might sound like a weakness, but it’s an intentional choice: Pangram would rather miss some AI-assisted writing than falsely accuse a human writer. In a world where detection results can affect everything from academic Standing to publishing opportunities, that tradeoff matters.
The company also published internal benchmarks every time an AI lab updates their models. That’s a level of transparency most detection tools don’t offer — and it matters, because the models keep changing, which means the detection baselines keep changing too.
What the Substack Integration Actually Means
In July 2026, Substack integrated Pangram’s technology into its platform. Readers can now scan posts, comments, and replies to see an estimate of how much was written by a human versus AI. Writers can attach an optional AI author’s note to disclose their process.
Substack CEO Chris Best framed it plainly in a public chat with Spero: software should handle everything except the hard part — having something worth saying. The detection tool, in that framing, isn’t a policing mechanism. It’s a transparency layer. Readers get context. Writers get to explain their process. Nobody gets surprised.
That distinction matters more than the technology itself. The tool doesn’t penalize AI-assisted writing. It doesn’t auto-takedown flagged posts. It gives readers information they can use to decide how much trust to invest — and gives writers a way to be honest about how they work.
That’s a meaningfully different model from the “detector as gatekeeper” approach that tanked the reputation of earlier tools. It treats AI assistance as a spectrum to be disclosed, not a binary to be enforced.
Why This Is Harder Every Year
The uncomfortable truth Spero keeps coming back to: AI models are getting better at sounding human. Every major release pushes the boundary a little further. The stylistic differences that Pangram’s model learned to detect in 2023 are narrower in 2026. They’ll be narrower still in 2027.
That’s not a criticism of Pangram or any other detector. It’s the nature of the problem. You’re building a classifier whose negative class — “human writing” — is being actively mimicked by tools that improve every quarter. Even the people building these models admit they don’t have them fully under control. The target moves.
Pangram addresses this by retraining continuously and by being careful about data drift. The human corpus they trained on is from before the AI flood, which gives them a clean baseline. But language itself evolves, and a detector trained on pre-2022 writing will eventually struggle to recognize human writing from 2027. Keeping the training data current without contaminating it with AI-generated text is an open problem.
There’s also the adversarial angle. AI detection is not just a technical challenge — it’s an arms race. Humanizers, paraphrasing tools, and prompt engineering techniques that specifically target detector weaknesses are all part of the landscape. A detector that can’t handle deliberately evasive prompts is a detector with a hole in it. AI-powered attacks are already running autonomously — and detection tools are part of the same race.
What This Means for You
If you’re reading content online — newsletters, blog posts, product reviews, job applications — and you want to know whether AI was involved, here’s the practical reality:
No single detector is definitive. Even the best models make mistakes. A 99%-accurate detector still misclassifies one in a hundred documents. Building your own local AI text detector is possible — but it faces the same fundamental tradeoffs. At internet scale, that’s a lot of false positives.
Look for the context, not just the score. A Substack post labeled as partially AI-assisted tells you something different from an anonymous product review with no author history. The detection score is one signal among several — author reputation, editorial process, topic complexity, and whether the writing has the kind of uneven texture that human writing tends to have.
Understand what the percentage means. A “30% AI-assisted” label doesn’t mean 30% of the words came from an AI. It means the model detected stylistic patterns consistent with AI involvement in some portion of the text. The actual breakdown — how much was generated, how much was edited, how much was written from scratch — is almost never knowable from the outside.
Check the tool’s methodology before trusting its verdict. Perplexity-based detectors will tell you the Declaration of Independence is AI-written. Detectors that don’t publish benchmarks or explain their training data are asking you to take their accuracy on faith. The same trust problem applies to companies building AI security tools. In a domain this hard, faith is not a verification strategy.
Assume the problem is getting harder. The models you’re trying to detect are improving faster than the detectors are. That’s not pessimism — it’s the observed trend. Any detection tool you use today will be less effective against next year’s models than it is against this year’s. Plan accordingly.
The Honest Take
The internet has a trust problem, and AI-generated content is part of it — but the solution isn’t a perfect detector. It can’t be, because the thing you’re detecting keeps getting better at pretending not to be detected.
Some AI leaders have called for pacing the frontier — but the transparency layer emerging in practice is more concrete than any policy proposal. What’s actually emerging is something more modest and more useful: a transparency layer. What AI CEOs mean when they call for slowing down is a question this layered approach answers better than any memo. Readers get signals. Writers get to disclose. Platforms get a way to show their work. Nobody pretends the problem is solved.
That’s probably the best we can hope for. And honestly, it’s enough. Knowing whether the thing you’re reading came from a person or a model — or some mix of both — is worth something. It changes how you read it. It changes what you trust. It doesn’t give you certainty, but it gives you context. And in a world where the alternative is reading everything with your guard up all the time, context is a pretty good deal.
Max Spero calls himself the “slop janitor” of the internet. I think of it differently. He’s building the closest thing we have to a way to tell, at a glance, what’s real and what isn’t — and then being honest about how imperfect that way still is. That’s not a janitor. That’s a starting point.