Who Decides, and Who Defends?A GliaNet Response to "We Must Pace the Frontier"

Dario Amodei's essay "We Must Pace the Frontier" has done something useful: it has forced an open argument about the speed of AI capability development, and it has put a frontier lab's own commitments on the table. His proposal has three parts: embedded third-party evaluators with employee-level access inside each lab, coordination among labs in democratic countries on shared safety standards (via regulation, or a narrow antitrust waiver in the meantime), and graduated global agreements running from bans on the most dangerous uses up to a possible speed limit on recursive self-improvement. Anthropic has committed unilaterally to the first step.

I am not going to argue here about whether pacing is right. Reasonable people are already doing that, and the honest answer is that the case turns on empirical claims about capability trajectories that none of us outside the labs can verify. The GliaNet Alliance has a different pair of concerns, and they hold whether or not the frontier slows down.

First: these should not be decisions made by a handful of companies, however well-intentioned.

Every serious proposal now on the table, including Dario's, originates inside the industry and is addressed, in the first instance, to the industry. The evaluators would be chosen and hosted by the labs. The standards would be negotiated among the labs. The antitrust waiver would be granted to the labs. Dario himself says that society "must have a say," and I take that as sincere. But a say is not the same as a seat. Whether the frontier goes faster or slower, the people living with the consequences have no structural role in the deciding.

This is not a complaint about any one company. It is a statement about the architecture of the conversation. Rules that are written by the governed for the governed, and that can be rewritten by the same parties when circumstances change, are not governance. They are a truce. Truces are valuable. They are not lasting.

Second: the whole environment would be safer if people could defend themselves.

Pacing addresses what gets built and how fast. It does not address the daily encounter between an ordinary person and an AI system that has been deployed by an institution to serve that institution's purposes. That encounter already happens millions of times a day: with insurers deciding claims, employers screening applicants, lenders setting terms, platforms shaping what you see. In each case the AI on the other side is capable, well-resourced, and loyal to someone else. The person arrives alone.

We call this the fidelity gap. Most AI governance, including the most careful lab-level safety work, aims to make systems less harmful. That is a duty of care, broadly understood. Far less attention goes to the next question: whose goals does the system actually serve? A model can be aligned to its developer's constitution, evaluated by embedded auditors, and released on a responsible schedule, and still be pointed at people rather than working for them.

The GliaNet answer is to give people their own agents. A Personal Augmentation Agent (PAA) is an AI system built to operate exclusively on behalf of the individual, across the domains of that person's life. It represents you when an institution's AI makes a consequential decision about you. It helps you contest systems that extract your data and attention. It strengthens your own judgment rather than replacing it. We have laid out a graduated path for such agents: a Self-Designated PAA that adopts the user-side mission, a Fiduciary PAA that embeds duties of care, fidelity, and loyalty into its architecture, and a certified Gliagent™ operated under the supervision of a Net Fiduciary™ that bears legal accountability for its conduct. The duties are carried by the institution behind the agent, through what we call the fiduciary tether, so that loyalty is a property of an accountable system rather than a claim made by a model.

A world of well-paced frontier models and unprotected individuals is safer than a world of racing frontier models and unprotected individuals. But it is not safe. A world in which each person also has a loyal, accountable agent working in their corner is safer still, and it is a world we can start building now, without waiting for antitrust waivers or treaties.

Third: the companies themselves should take on fiduciary duties, and build them into their governance.

This is the point I most want the frontier labs to hear. The GliaNet proposal is not only that people should have loyal agents. It is that each of the large AI companies should voluntarily act toward the public as a fiduciary, and should embed those duties deeply in its corporate governance, where they cannot be quietly unwound when competitive pressure returns.

Fiduciary law is the oldest technology we have for governing power exercised on behalf of others. Doctors, lawyers, trustees, and financial advisors operate under duties of care and loyalty that survive changes of management, market cycles, and shifts in public mood, because those duties are structural rather than promissory. The commitments now being offered by frontier labs, admirable as some of them are, do not have that character. An evaluator's access can be narrowed. A standard can be renegotiated. A safety team can be reorganized. Nothing in current corporate structure prevents any of this, and the history of the last several years shows how quickly it happens.

There is a sensible objection: a frontier lab has obligations to society and to long-term safety that may sometimes override what an individual user wants, so it cannot be a "pure" fiduciary to each user. I accept that. But it argues for a qualified and explicit set of duties, not for none. A lab could articulate publicly how conflicts among users, society, and the company are resolved. It could offer a fiduciary tier of service with real teeth and independent verification. It could serve as fiduciary infrastructure, giving certified Net Fiduciaries™ the access they need to build loyal agents on top of its models. It could seek certification under a recognized standard. Anthropic's status as a public benefit corporation gives it more room to do these things than most of its peers, and its own constitution for Claude already gestures toward them. The point is to make the gesture binding.

And there is a commercial case, not only a moral one. Fiduciary commitments build trust; trust brings the richer data and deeper engagement that make services better; better services build loyalty and durable revenue. Duty to users and duty to shareholders are not in conflict. They are the same strategy, seen from two sides.

AND, not OR.

Pace the frontier or don't; that debate will run its course. What should not be optional is the rest: decisions made with the public rather than merely on its behalf, loyal agents that let people meet institutional AI on something closer to equal terms, and fiduciary duties written into the governance of the companies building the most powerful systems in history. Absent those, there is little guarantee that anything produced in this process will last, or that it will be aimed at protecting and promoting the best interests of humanity. Dario says we owe it to humanity to try. So we do. Let us try on all of these fronts at once.

Want to take this further?

If you build, deploy, or govern AI systems and want to talk through what fiduciary duties would look like inside your organization, we'd like to hear from you.

Contact Us