{"version":1,"type":"story","url":"https://digestai.news/story/pangram-detector-flags-claude-opus-5-text-as-100-ai-generated-in-subst","json":"https://digestai.news/story/pangram-detector-flags-claude-opus-5-text-as-100-ai-generated-in-subst.json","markdown":"https://digestai.news/story/pangram-detector-flags-claude-opus-5-text-as-100-ai-generated-in-subst.md","slug":"pangram-detector-flags-claude-opus-5-text-as-100-ai-generated-in-subst","headline":"Pangram detector flags Claude Opus 5 text as 100% AI-generated in Substack test","summary":"Writer BJ Beyond conducted an adversarial test to see if Claude Opus 5 could evade Pangram, the AI detection tool recently partnered with Substack. Using Claude Opus 5 at its maximum reasoning effort, the author asked the model to write an entire essay in English specifically designed to defeat AI detection while explicitly analyzing the experiment itself.\n\nClaude refused twice to produce a covert evasion guide or pretend to be human, agreeing only to a fully transparent experiment. When the resulting, unedited text was evaluated, Pangram assigned it a score of 100% AI-generated. Pangram claims a 99.8% benchmark accuracy on Claude Opus 5 outputs.\n\nThe author notes that while Pangram's classifier successfully recognized the unmodified model output, the test highlights the functional difference between text artifact detection and provenance. Substack launched the Pangram scanner alongside a \"How I make this\" process disclosure feature, and the essay argues that mechanical detection flags stylistic patterns rather than intent or deception.","keyPoints":["Pangram flagged an unedited essay written by Claude Opus 5 at max reasoning effort as 100% AI-generated.","Claude Opus 5 refused prompts requesting covert human impersonation before agreeing to a transparent evasion test.","Pangram claims a 99.8% accuracy rate on its Claude Opus 5 detection benchmark."],"whyItMatters":"The test demonstrates that frontier models struggle to evade dedicated stylistic classifiers without manual rewriting, underscoring why platforms like Substack pair automated detection tools with creator disclosure forms to establish genuine provenance.","category":{"slug":"society","name":"Society & Work","url":"https://digestai.news/category/society"},"entities":{"companies":["Substack","Pangram","Anthropic","HackerNoon"],"models":["Claude Opus 5"],"people":["Chris Best","BJ Beyond"]},"firstPublishedAt":"2026-09-20T17:00:00Z","updatedAt":"2026-09-20T17:00:00Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"hackernoon.com","title":"I Asked Claude Opus 5 to Beat Pangram's AI Detector. It Failed with A Perfect Score","url":"https://hackernoon.com/i-asked-claude-opus-5-to-beat-pangrams-ai-detector-it-failed-with-a-perfect-score","publishedAt":"2026-09-20T17:00:00Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[],"thread":null,"cite":{"text":"Digest AI, \"Pangram detector flags Claude Opus 5 text as 100% AI-generated in Substack test\", 20 September 2026, https://digestai.news/story/pangram-detector-flags-claude-opus-5-text-as-100-ai-generated-in-subst","publisher":"Digest AI","title":"Pangram detector flags Claude Opus 5 text as 100% AI-generated in Substack test","datePublished":"2026-09-20T17:00:00Z","url":"https://digestai.news/story/pangram-detector-flags-claude-opus-5-text-as-100-ai-generated-in-subst"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}