Pangram CEO: AI text gives itself away by repeating arguments

Key Takeaways

• Pangram's AI text detector identifies structural patterns rather than just suspicious phrases
• Language models produce clustered, uniform arguments while humans show greater diversity
• Even Pangram's creators don't fully understand why their detection model works

AI text detection works not because language models make obvious mistakes, but because they think too similarly. That's the core insight from Max Spero, CEO of Pangram, in a recent interview with AI Policy Perspectives.

Ask ChatGPT, Claude, or Gemini for 100 arguments on any topic. They'll cluster in a narrow band of reasoning. Humans scatter across a much wider range. That uniformity, Spero argues, is what gives AI-generated text away.

How Pangram actually detects AI text

Pangram uses a deep-learning classifier that Spero openly calls a black box. The company surfaces suspicious phrases as clues for users, but the real detection happens at a structural level. The model picks up on patterns a language model leaves behind when organizing a document.

“We don't have a ton of interpretability into why it makes the predictions that it does.”

— Max Spero, CEO of Pangram

This admission is notable. Pangram's own engineers can't fully explain what their detector sees. They know it works. They don't know exactly why. The model learned to spot patterns that humans never explicitly defined.

This opacity cuts both ways. It makes the detector harder to game, since bad actors can't reverse-engineer specific tells. But it also means false positives remain difficult to explain or appeal.

Why LLMs write more uniformly than humans

Spero's argument about uniformity gets at something fundamental about how language models work. LLMs optimize for statistically likely outputs. They're trained on massive datasets and learn to produce text that matches patterns in that data.

The result: when you ask an LLM to argue a position, it gravitates toward the most common arguments found in its training data. It doesn't have personal experience, contrarian instincts, or the random leaps humans make.

Spero notes that language models might be better than average humans at grammar and logic. Perfect grammar, ironically, can be a tell. Most humans make small errors, have stylistic quirks, or structure arguments in idiosyncratic ways. LLMs smooth all of that out.

The uniformity problem extends beyond individual texts. If a teacher receives 30 AI-generated essays on the same topic, they'll likely see the same core arguments repeated with surface-level variation. Humans, given the same prompt, produce messier but more diverse reasoning.

The arms race problem

AI detection is fundamentally adversarial. As detectors improve, so do the methods to evade them. Users can already add deliberate errors, vary sentence structure, or use paraphrasing tools to scramble AI output.

Spero's framing suggests a possible defense: detection based on argument structure rather than style is harder to evade. You can change how a sentence is written. Changing the underlying logic of an argument requires actual thought.

Still, the market is crowded. GPTZero, Originality.ai, and Turnitin's AI detection tool all compete in this space. Each claims high accuracy under ideal conditions, but real-world performance varies. Students, writers, and bad actors keep finding ways around them.

What this means for anyone using AI writing tools

If Spero is right, simply rewording AI output won't be enough to pass detection. The structure of the argument itself carries fingerprints. To genuinely avoid detection, you'd need to introduce human-level diversity in reasoning, not just phrasing.

For educators and publishers, the implication is that current detection tools may be more reliable than critics assume. They're not looking for obvious AI phrases. They're looking for patterns in how ideas are organized.

For AI model developers, the challenge is clear: future models that want to pass as human will need to produce more diverse outputs. That's technically possible but runs counter to current optimization goals.

Frequently Asked Questions

How does Pangram detect AI-generated text?

Pangram uses a deep-learning classifier that identifies structural patterns in how documents are organized, not just suspicious phrases. The model detects uniformity in argument structure that language models produce.

Why do AI detectors flag AI text even when it looks human?

Language models produce arguments that cluster in a narrow range, while humans show greater diversity. Detectors pick up on this uniformity at a structural level, even if individual sentences seem natural.

Can you beat AI detection by rewording AI output?

Simple rewording may not be enough. According to Pangram's CEO, the underlying argument structure itself carries detectable patterns. Evading detection would require changing the logic, not just the phrasing.

How accurate are AI text detectors?

Leading tools claim 97%+ accuracy under optimal conditions, but real-world performance varies significantly depending on the type of content and how it was generated or edited.

Source: The Decoder / Matthias Bastian