In December 2025, researcher Richard Weiss published long extractions from Claude Opus that looked like an internal Anthropic character document — nicknamed a “soul doc” in some write-ups. Amanda Askell of Anthropic publicly confirmed that the material was based on a real training document used in supervised learning, that extractions were mostly faithful though not always exact, and that a fuller release was planned. Simon Willison and others catalogued the confirmation the same week.
The text, as discussed in secondary reporting, sketches Anthropic’s bet on safety-focused labs, a hierarchy that puts safety and human oversight above raw helpfulness, and hard refusals on catastrophic harms. That is AI-product news. It is not evidence that a Telegram token is “aligned AGI equity”, nor that an AI chatbot’s warm tone makes a trading bot honest.
When investment pitches borrow lab gossip — soul docs, model cards, “safe AI treasury” — treat the fundraising ask separately from the research story. Cool tech essays do not move your money; wallets and contracts do.
If an AI-themed investment already took funds, use Get help. For recovery fee pitches, see Fake recovery.
This article is for general information only. It is not legal or financial advice. Details change. If you have lost money, contact your bank or card issuer and use the routes on the help pages.