San Francisco · USA · TRUST-CHECKED PBC · WITH A FAILSAFE
Who controls Anthropic?
ULTIMATE CONTROL — CONTESTED The Long-Term Benefit Trust — unless a stockholder supermajority bends around it
The most structurally checked of the major labs — but the charter contains 'failsafe' provisions that let a large enough stockholder supermajority amend or dissolve the Trust without the trustees' consent.
Chain of command
- THE TOOLAnthropic PBC
Builds Claude; legally bound to its public-benefit mission.
- BOARDBoard of directors
Sets company direction; a majority is now appointed by the Trust.
- CHECKS THE BOARDLong-Term Benefit Trust
5 independent, equity-free trustees (AI safety, national security, policy) — incl. Vas Narasimhan, M-F. Cuéllar, Richard Fontaine. Hold Class T stock to elect/remove a board majority.
- THE BEND-AROUNDStockholder supermajority 'failsafe'
Anthropic's charter lets a sufficiently large supermajority of stockholders amend the Trust's powers — or dissolve it — WITHOUT trustee consent. Critics note large investors like Amazon & Google could plausibly form such a bloc.
- OPEN QUESTIONHow strong is it, really?
The supermajority thresholds are undisclosed and the full Trust Agreement is unpublished. Until it's tested, the strongest check in the industry is also the least verifiable.
At a glance
| Model | Claude series, frontier + safety mission |
|---|---|
| Form | Public-benefit corp + Long-Term Benefit Trust |
| Equity | Founders + Amazon & Google as major investors |
| Workforce | ~3,000–5,000 staff [10] |
Ultimate control — contested: LTBT — overridable by stockholder supermajority. Most checked of the labs, but the charter has a trustee-free 'failsafe' dissolution path. [5]
The pitch and the record
They say: “An independent trust keeps us honest — it can even replace the board.”
True — until you read the charter. Anthropic's own incorporation docs let a stockholder supermajority amend or dissolve the Long-Term Benefit Trust without the trustees' consent. The thresholds are undisclosed; the full agreement is unpublished. Critics note Amazon & Google could plausibly form such a bloc. [4] [5]
The essay's proposed fixes
A prescription from "The People Inside", not a fact about Anthropic and not something the lab has agreed to.
CLOSE THE TRAPDOORS Anthropic is the most checked of the labs—but its charter leaves structural trapdoors that leave it vulnerable to a shareholder revolt. To actually guarantee a good future, those trapdoors must be removed.
- Revoke the shareholder override. Amend the charter to make the Long-Term Benefit Trust completely irrevocable. The clause letting a stockholder supermajority bypass or dissolve the Trust must be legally destroyed, giving the Trust absolute, unchallengeable authority over the board.
- Accelerate majority board control — now. The Trust currently elects only a minority, with majority control promised later. Execute that transition immediately. Waiting until models reach superintelligence guarantees investors who grasp the financial implications will aggressively block the handover.
- Establish cryptographic IP custody. Transfer legal ownership of the foundational model weights directly into the Trust. The commercial PBC operates only on a revocable license—so if investors attempt a corporate coup, the Trust revokes it and instantly destroys the rogue entity's commercial value.
- Implement a hard capital cap. Enforce a strict, legally binding cap on investor returns. Without an explicit ceiling, the pressure to squeeze infinite value from superintelligence will break the PBC mandate and force the board to choose between human safety and investor lawsuits.
Sources the essays cite for Anthropic
- [4] Anthropic — "The Long-Term Benefit Trust." anthropic.com/news/the-long-term-benefit-trust
- [5] Analysis of Anthropic's charter "failsafe" — stockholder supermajority can amend/dissolve the Trust. ailabwatch.org/blog/anthropic-certificate-of-incorporation · lesswrong.com
- [10] Approximate workforce figures: OpenAI ~4,500→8,000 (2026); Anthropic ~3,000–5,000; DeepMind ~6,000; Meta Superintelligence Labs ~3,400; xAI ~1,200. saastr.com · engadget.com · builtin.com · time.com
Read the full essays: Who Would Own God? · The People Inside.