Anthropic S-1: Pangram Estimates 92% of Human Text

Key Takeaways

1Pangram analyzed the entire S-1 of Anthropic and estimates that 92% of the text is of human origin.

2Some sentences were flagged with a high degree of confidence as generated by AI, with no entire sections produced by AI according to Pangram.

3Anthropic's AI tagging system is not public, and the company has not commented on these findings.

Why it matters : The use of external detection tools highlights the presence of spot editing by AI in official documents, while internal tools are not accessible.

The entire S-1 filing of Anthropic has been submitted to Pangram, which concludes that it is predominantly human-written. However, some sentences are flagged with a high degree of confidence as being AI-assisted. Since Anthropic's marking tool is not public, the company has not commented on these results.

No official marking available and no comment from Anthropic

Anthropic's official AI marking system is not yet accessible to the general public and has therefore not been used to verify the S-1. Anthropic was approached for a response to the results from another tool but did not provide an answer. Information about the company's current plans and finances was made public today.

Pangram's quantified verdict on the entire S-1

According to Pangram, 92% of Anthropic's IPO filing, which contains 259,031 words, is written by humans. However, the tool flagged certain sentences and passages with a high degree of confidence as containing AI. For example, a sentence about the speed of AI development, mentioning cycles measured in months rather than years, was marked with high confidence as generated by AI. Pangram notes that these results cover the entire analyzed file.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

An analysis conducted by Pangram after a file size failure

The Pangram service checked for signs of AI writing in Anthropic's S-1 filing. An initial attempt at local analysis failed due to the file size. Max Spero, CEO and co-founder of Pangram, then executed the analysis of the complete file, which consists of 335 pages.

Signals of AI editing and limitations of detectors

The accuracy of a tool like Pangram is not 100%, and the use of AI assistants for editing or proofreading can influence detections. According to Max Spero, the footprint detected in the S-1 suggests slight AI editing or use at the sentence level, and it does not appear that significant sections are entirely generated by AI. A separate experiment by Aki Ito illustrates this gray area: after submitting her publications to Pangram, a sentence from an article was flagged as written by AI, and she traced its origin to a rephrasing suggested by ChatGPT that she had accepted. Generally, tools like Grammarly, powered by AI, can also offer rephrasings, and accepting these suggestions can lead to reports of being written by AI.