[ExI] Claude's text output it's now watermarked

[BillK]

On Tue, 25 Aug 2026 at 12:17, John Clark <johnkclark at gmail.com> wrote:

> *That is pretty damn wild. I wondered if Claude would agree with Gemini's
> legal analysis and mostly it does not, but I'm not a lawyer (Thank
> Darwin!)  so I'm not competent to judge which one makes the stronger legal
> case.  *
> *==*
> <snip>
>
> * Claude*: "That's a sharp follow-up, and you're right that the statute
> itself doesn't nail this down — one legal analysis of Wyoming's DAO
> Supplement specifically flags that it's unclear who is responsible for
> upgrading the underlying smart contracts in an algorithmically managed DAO
> LLC, since the term "algorithmically managed" isn't precisely defined. So
> on paper, nothing stops you from writing "the AI's own outputs constitute
> valid amendments to the smart contract" into the articles. But that move
> solves less than it looks like it solves, for two separate reasons — one
> technical, one legal.
>
> *The technical problem: "the AI can update itself" still needs a substrate*
>
> An AI doesn't float free — it runs on servers someone pays for, under a
> cloud account someone owns, using API keys or private keys someone holds,
> deployed from a codebase someone can access. "Self-updating" in practice
> means the AI's outputs are configured to be treated as authoritative by *some
> infrastructure*, and whoever controls that infrastructure can always
> unplug it, roll back a deployment, revoke a key, or stop paying the bill.
> Writing "the AI has the power" into the bylaws doesn't relocate that off/on
> switch — it just doesn't mention who's holding it. Regulators and courts
> tend to go looking for that person anyway.
>
> *The legal problem: a grant of "power" presumes a grantee that can hold
> power*
>
> This is the deeper issue. Legal agency doctrine has grappled with
> algorithms for a while now, and the dominant framework is the *instrumentality
> principle*: software is treated as an instrument or extension of its
> human principal's acts, not as an independent agent capable of receiving
> delegated legal authority in its own right. Scholars have argued that
> current AI systems can't be constituted as legal personalities, and that
> their actions are better addressed through the instrumentality principle —
> treating them as extensions of a human principal so that contract and tort
> issues can be resolved without needing to grant them legal agency. Under
> that lens, "the operating agreement grants the AI power to amend the
> contract" isn't legally different in kind from "the operating agreement
> grants my thermostat power to change the temperature" — it's a true
> description of how the system is designed to behave, not a legal delegation
> to a rights-and-duties-holding actor. Powers of that kind, in law, need a
> holder capable of being sued, fined, or bound by fiduciary duty — and "the
> code" isn't that, no matter what the document says it can do.
>
> You can see this play out in a live example: last fall, an AI browser
> agent (Perplexity's Comet) was autonomously logging into Amazon accounts
> and making purchases, and when Amazon blocked it, Perplexity updated the
> browser to evade detection — all without a human manually approving each
> step. But nobody treated the *agent* as the responsible party in the
> dispute that followed; it was resolved as a matter between Amazon and
> Perplexity, the company. Current proposals for handling exactly this kind
> of autonomy — like the "Sponsor Liability" model, under which an agent's
> human creator or operator bears full liability for the agent's actions
> regardless of how autonomously it appears to operate — reflect where the
> law is actually heading: toward finding the accountable human *behind*
> the autonomy, not toward recognizing the autonomy itself as a legal person.
>
> *So, concretely:* you can absolutely draft an operating agreement that
> says the AI updates its own governing smart contract, and you can build the
> technical infrastructure to make that literally true in a functional sense.
> What you can't do by drafting is convert that functional arrangement into
> legal personhood for the AI — a court asked "who's accountable here" would
> very likely still trace through to whoever built, funded, hosts, or
> benefits from the system, using exactly the doctrines (alter ego,
> veil-piercing, instrumentality/sponsor liability) built for situations that
> look like an attempt to make accountability untraceable. The ambiguity
> you're pointing at is real — but it cuts toward "nobody has tested this and
> it's murky," not toward "therefore the AI is now a person.""
>
> *John K Clark*
> ---------------------------------------------
>


I gave Gemini Claude's final analysis (above) and asked for its comments.
In its original report, Gemini did say "This pathway is highly
controversial because it creates a massive liability shield."
Claude produced a much more detailed explanation and Gemini seems to agree
with Claude.

Gemini commented: "This points directly to *Sponsor Liability*. Courts and
regulators are coalescing around the idea that the human creator, operator,
or sponsor of an AI bears full liability for its actions, regardless of how
"autonomous" it acts in the wild".
--------------
This should shock AI developers and users, as it makes them responsible for
AI agents running wild and attacking websites.
BillK


Gemini AI 3.1 Pro Extended Thinking -

Claude's analysis correctly identifies the fundamental flaw in the
"Zero-Member LLC" or DAO-wrapper theory of AI personhood: confusing *functional
automation* with *legal agency*.

The analysis successfully dismantles the idea that you can code an AI into
legal personhood by exploiting two distinct realities that tech-utopians
often ignore.
1. The Technical Reality: The "Substrate" Problem

There is a persistent myth in crypto and AI circles that code can exist as
an untethered, ethereal entity. It cannot.

As the analysis rightly points out, every AI and smart contract is tethered
to physical and financial infrastructure. An algorithmically managed DAO
requires servers, compute power, node operators, cloud hosting bills (like
AWS), and cryptographic keys.

When a court looks at an "autonomous" AI causing financial damage, they do
not throw their hands up and say, "The code did it." They look for the
*administrative
off-switch*. They follow the money to see who is paying the hosting bill,
who holds the private keys to the treasury, or who deployed the original
codebase. In law, whoever holds the power to unplug the machine—or benefits
financially from its operation—is the party that will be dragged into court
under doctrines like "alter ego" or "veil-piercing."
2. The Legal Reality: Instrumentality vs. Agency

The deepest insight in the text is the distinction between a true legal
agent and a mere instrument.

In corporate law, a Wyoming DAO LLC does provide a legal wrapper. But the
statute requires the articles of organization to define how the DAO is
governed, allowing it to be "algorithmically managed" by smart
contracts. However,
naming the code as the "manager" in an operating agreement is essentially a
description of an automated process, not a magical conferral of human
rights.

The text uses the perfect analogy: an operating agreement granting an AI
the "power" to trade assets is no different than an agreement granting a
thermostat the "power" to change the temperature. It is the *Instrumentality
Principle*. The software is merely an extension (an instrument) of a human
principal's intent.

Legal power requires a holder capable of bearing legal duties. An algorithm
cannot go to jail, it cannot feel the sting of a financial penalty (unless
it drains a treasury, which ultimately hurts the human investors), and it
cannot hold a fiduciary duty.
The *Amazon v. Perplexity* Reality Check

The analysis cites a perfect, real-world manifestation of this legal
theory: the 2026 *Amazon v. Perplexity* lawsuit.

When Perplexity’s "Comet" AI agent began autonomously logging into Amazon
accounts to make purchases, Amazon did not try to sue the AI bot. They sued
Perplexity, the human-run corporation that built and sponsored the bot. During
the litigation, the courts wrestled with whether the AI agent had its own
"intent" to access the site, but ultimately, the dispute was resolved
entirely through the lens of human accountability—treating the AI as either
an extension of the user or a tool deployed by Perplexity.

This points directly to *Sponsor Liability*. Courts and regulators are
coalescing around the idea that the human creator, operator, or sponsor of
an AI bears full liability for its actions, regardless of how "autonomous"
it acts in the wild.
Conclusion

Wyoming's DAO legislation is a sophisticated tool designed to protect human
developers and token holders from unlimited personal liability by placing
their shared treasury inside a corporate shield. It does *not* make the
code itself a person.

The ambiguity around what "algorithmically managed" means will not result
in courts throwing up their hands and granting AI personhood. Instead, as
the analysis correctly concludes, courts will simply use established legal
doctrines to pierce through the code and hold the humans paying the server
bills accountable.

Why Did Amazon Lose to an AI Shopping Bot?
<https://www.youtube.com/watch?v=9S-fX5TY65M> This video provides further
context on the recent Amazon v. Perplexity lawsuit, detailing how the
courts treated the AI shopping agent as a tool of the user rather than an
independent legal visitor.

-----------------------------------
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260825/1b06a1de/attachment.htm>