ItInsightsWorld Read: Pangram’s Anthropic Prospectus Scan Finds Mostly Human Writing With Some High-Confidence AI Signs

The scan’s core takeaway was that the text did not read as wholly machine-generated, yet it was not entirely free from automated fingerprints either: Pangram classified about 92% of a roughly 259,000-word filing as human-written while marking some passages with “high” confidence for possible AI traces.

That finding does not identify the exact lines produced by a person or a model, and it cannot confirm whether Anthropic used its own assistant Claude in drafting the prospectus.

Business Insider reported that an attempt to run the full document itself through a detector was blocked by the file’s size, before Max Spero, Pangram’s co-founder and chief executive officer, assisted with completing the whole check.

Spero characterized the result as suggesting light AI editing at the sentence level or localized AI use, rather than the presence of long sections created wholly by a model.

Among the flagged text was a passage describing how fast AI advances, with new model versions often arriving in months instead of years.

Business Insider added a caveat: AI writing detectors are not completely accurate, and outcomes can change depending on how AI is used to edit, proofread, or rephrase wording.

In a separate example involving the same detection tool, a Business Insider journalist fed published LinkedIn posts and articles to Pangram; the system identified one sentence as AI-written, though the reporter recalled writing it and later accepting a revision suggested by ChatGPT.

Anthropic did not reply to Business Insider’s request for comment, and the report noted that Anthropic’s own watermarking system was not used because it has not been opened to general users.

Katie Notopoulos argued that the outcome raises a compelling question about why an AI company such as Anthropic would not rely more on its own products for a prospectus. She also pointed out that the document faces review by the U.S. Securities and Exchange Commission, where each sentence can be examined, making human participation reasonable.

At ItInsightsWorld, the point read from the episode is that detection tools can describe textual signals without settling who wrote the document.