Pangram Finds Anthropic IPO Filing Mostly Human-Written

Anthropic’s S-1 registration statement for a planned initial public offering was largely written by humans, according to an analysis using Pangram, an AI-detection service. A 335-page, 259,031-word version of the filing was submitted to Pangram, whose co-founder and chief executive Max Spero reported that about 92 percent of the document appeared human-written. Pangram flagged a small number of sentences and passages with high confidence as AI-generated, but Spero said the result indicated light AI editing or sentence-level assistance rather than substantial sections being fully produced by a model. The analysis did not use Anthropic’s watermarking tool, which is not widely available, and AI detectors can mistake lightly edited or AI-assisted text for synthetic writing. Anthropic did not respond to a request for comment. The finding arrived as the filing detailed Anthropic’s financial position, business plans and risk disclosures, including concerns about AI’s rapid development and broader societal consequences. For an AI company preparing to raise public capital, the result carries an awkward implication: Anthropic did not appear to rely heavily on its own models for one of the most mechanical and technical genres in corporate finance. That could reflect caution, because securities filings are scrutinized by the SEC and investors, and inaccurate or awkward language can carry legal and reputational consequences. It also suggests the company still treated the document as a high-stakes legal instrument requiring human judgment. At the same time, the largely human composition may reassure regulators and investors. If a filing of this importance were substantially generated by a model, concerns could arise over accuracy, accountability and the company’s own confidence in the reliability of its technology. The broader lesson is likely less about deception than about the cautious, hybrid way high-profile organizations are using AI in sensitive settings: models may assist with phrasing or proofreading, but humans remain responsible for the final text and its meaning.