✕ BANNED Anthropic PBC T-48: 🔴 STRONG SHORT
[DOCUMENTED · HIGH RISK]
Bartz v. Anthropic — $1.5B copyright settlement (2025, USDC ND Cal, Judge William Alsup).
Anthropic downloaded approximately 500,000 copyrighted works from shadow libraries including LibGen, Books3, and Pirate Library Mirror.
Court held: LLM training = "exceedingly transformative" fair use; creation of a permanent pirated central research library = direct copyright infringement.
Settlement: ~$3,000 per work × ~500,000 works = $1.5 billion. Largest copyright class action settlement in US history.
Bartz v. Anthropic Order (Judge Alsup 2025) · Wolters Kluwer Copyright Blog · iplawgroup.com [Ref. 1]
[DOCUMENTED · HIGH RISK]
SEC Rule 21F-17(a) NDA violation.
Anthropic exit agreements required prior corporate consent before employees could communicate with the SEC, required waiver of whistleblower financial awards, and bundled non-disparagement clauses under threat of vested equity clawback.
Anonymous whistleblowers filed formal complaint with SEC July 2024. Directly violates Dodd-Frank Act Rule 21F-17(a).
SEC whistleblower complaint Jul 2024 · Dodd-Frank Act · SEC Rule 21F-17(a) [Ref. 2]
[DOCUMENTED · ACTIVE]
FTC Section 6(b) inquiry — Federal Trade Commission actively auditing exclusive Anthropic-Amazon-Google data partner arrangements to determine how integrated architectures distort market competition and resource allocation.
FTC p246201 AI Partnerships 6(b) report · FTC press release Jan 2024 [Ref. 3]
[ACADEMIC · DOCUMENTED]
Machine unlearning: computationally near-impossible.
Exact unlearning of data from trained LLM weights without inducing catastrophic forgetting is mathematically infeasible for massive neural networks. Constitutional AI safety guardrails are superficial patches over retained restricted data — including the pirated works described above.
arXiv:2406.10952 · arXiv:2410.16454 · arXiv:2510.25117 [Ref. 4]