I Test-Drove Substack’s New AI Detection Tool, and It Mostly Worked

Substack's new AI detection tool, powered by Pangram, works—but not perfectly. A CNET test drive found the system correctly flagged most AI-generated text yet missed some sophisticated outputs, raising questions about the arms race between detection and generation.

Substack, the newsletter platform hosting millions of writers, has integrated Pangram's AI detection engine directly into its editor. The move aims to help readers and publishers identify content potentially written by large language models like ChatGPT or Claude. As generative AI floods the web, platforms face pressure to maintain trust and transparency. Substack's integration is among the first by a major publishing tool to bake detection into the writing workflow, not just as a post-hoc check.

The tool appears as a simple button in Substack's editor. When clicked, it scans the selected text and returns a percentage likelihood of AI generation. In testing, it correctly identified 4 out of 5 clear AI-written passages, but flagged a human-written sample as 'possible AI' once. False positives—accusing human authors of using AI—remain a concern for writers who rely on AI assistants for editing or brainstorming. Pangram claims its model is trained on patterns specific to AI writing, including repetitive phrasing and overly balanced sentence structures.

CNET's test involved five samples: two purely human-written op-eds, two fully AI-generated articles, and one hybrid. The tool gave scores of 0%, 12%, 87%, 94%, and 63% respectively. The 63% on a hybrid mix suggests it can detect blended content but with lower confidence. Substack CEO Chris Best stated the integration 'gives readers more context about what they're reading,' while acknowledging no detector is 100% accurate. Competitors like Originality.ai and GPTZero already offer similar services, but Substack's native integration could set a new standard for platform-level transparency.

Industry observers note that AI detection tools often struggle with evolving models—what works today may fail tomorrow. The Pangram engine uses a combination of perplexity scoring and burstiness analysis, metrics that measure randomness and variance in sentence length. However, as generative models improve, they produce more human-like text, potentially rendering detection obsolete. For now, Substack's tool is a useful guardrail, but not a silver bullet.

Substack plans to refine the tool based on user feedback and false-positive reports. Writers and subscribers should watch for updates to Pangram's model and perhaps broader adoption across the platform. If detection becomes standard on Substack, other publishing tools like Medium or WordPress may follow, accelerating a shift toward mandatory AI labeling. The next milestone: a public accuracy audit by third-party researchers.