About 92% of Anthropic’s 335-page IPO registration statement was written by humans, according to an analysis by AI-detection service Pangram. The service flagged a small number of sentences and passages as potentially involving AI, offering a glimpse into how the AI company may have used its own technology in a key investor filing. But detection results cannot establish who actually wrote or edited the text.
Pangram analyzed roughly 259,000 words
Anthropic filed its S-1 on October 26, 2026. The document runs to 335 pages and contains about 259,031 English words. Pangram co-founder and CEO Max Spero helped analyze the full filing after technical issues arose in an initial attempt to process the large document.
Pangram classified about 92% of the filing as human-written. It flagged some sentences and passages with relatively high confidence as potentially involving AI. One example described the rapid pace of AI development, with new models appearing on cycles measured in months rather than years.
Spero said the results were more consistent with light AI editing or the use of AI on individual sentences than with large sections written entirely by AI. Anthropic had not responded to the findings at the time.
Detection does not establish authorship
Pangram’s analysis identifies features that may be associated with AI-written text; it is not a record of how the document was produced. Detection is not perfectly accurate, and human-written material can be flagged if AI tools were used to polish, rewrite or proofread it.
Journalist Aki Ito previously submitted her LinkedIn posts and published articles to Pangram for analysis. The service flagged one sentence in just one article. On review, Ito found that ChatGPT had suggested rewriting a sentence she had drafted, and she had accepted the change. The example illustrates how AI-assisted editing can make it difficult for a detector to distinguish human writing from machine involvement.
Anthropic also has its own AI watermark-detection system, but the tool was not publicly available at the time, so it could not be used to independently verify Pangram’s findings on the S-1.
Human review remains central to IPO disclosures
An S-1 is a central disclosure document for investors and regulators, setting out a company’s business and financial information for review by the U.S. Securities and Exchange Commission. Anthropic’s filing covers its plans and current financial position. In documents of this kind, accurate wording and complete disclosure shape how investors assess the business and its risks.
The analysis raises a practical question: how much should Anthropic rely on its own models when preparing such filings? S-1 documents contain extensive standardized and technical language that could, in principle, be drafted or edited with AI. But if Anthropic relied mainly on human authors, that would underscore the continued role of human review and accountability in high-stakes disclosures.
Pangram’s conclusion is that most of the filing was written by people, with a small amount potentially receiving AI assistance. The detector’s limitations, together with Anthropic’s lack of comment, mean the analysis cannot establish which passages involved AI or how extensively it was used.