Dario Amodei has made an argument that deserves serious attention.
In We Must Pace the Frontier, he argues that AI capability is advancing faster than alignment, interpretability, operational discipline, and evaluation can keep up. He proposes stronger independent oversight, coordination among frontier laboratories, and international arrangements intended to slow the most dangerous forms of capability growth.
The risks he identifies are real. The need for independent evaluation is real. The possibility that commercial competition can reward speed over safety is also real.
But there is a deeper assumption inside the proposal that must be challenged.
The argument risks placing frontier laboratories and leading states in a morally privileged position: we possess the capability, we understand the danger, we will act responsibly, and wider access must be constrained because others may not.
That position may come from genuine concern. It is still the wrong foundation for governing a technology with consequences for all humanity.
Conscience is not governance. No company or country should become the moral owner of human progress.
The better direction is not uncontrolled acceleration. It is accountable progress: broaden beneficial capability, narrow high-consequence authority, test safeguards continuously, and distribute the power to question those who control the technology.
What Amodei Gets Right
A contrarian response should begin with the strength of the original argument.
Amodei does not call for abandoning AI. He describes its potential to improve health, accelerate science, increase abundance, and strengthen human agency. His concern is that the pace of frontier development may outrun the institutions and technical methods required to keep increasingly capable systems under control.
His proposal for embedded external evaluators is particularly important. Frontier companies should not be the sole judges of whether their systems are safe. Evaluators need meaningful access, the ability to report material findings, and protection from commercial pressure.
He is also right that a promise to slow down means little without verification. A voluntary commitment that competitors cannot inspect can become a public-relations position rather than a constraint.
The article’s emphasis on operational excellence matters too. Serious failures do not always come from a missing scientific breakthrough. They can emerge from weak data controls, faulty environments, incomplete monitoring, poor execution, or safeguards that were designed correctly but implemented badly.
These are valuable observations.
The disagreement begins when safety becomes tied to preserving the strategic lead of those already at the frontier.
Reader takeaway: The warning should be heard. The proposed distribution of power should still be questioned.
Where The Logic Turns
Amodei’s framework has three levels: embedded evaluators inside frontier companies, coordination among companies and governments in democratic countries, and eventual global coordination. At the same time, he argues that democratic states should protect and widen their AI lead over China through restrictions on advanced chips, action against unauthorised distillation, and stronger protection against model theft.
The strategic logic is understandable. If one group slows while another accelerates, the group that slows may lose influence over the future.
But this creates a circular structure.
The institutions that hold the greatest capability define the risk. They help define the standards. They choose or contract with evaluators. They seek coordination that protects their commercial position. Their governments preserve geopolitical advantage. Wider access is then treated as a threat to be managed.
The result may be safer frontier development. It may also be a more durable concentration of technical, economic, and political power.
This is not an accusation about motive. It is a structural concern.
A system of governance should remain credible even when the people running it are less conscientious, less competent, or more commercially aggressive than today’s leaders. If safety depends on the character of a few executives, it is not yet a public control system.
Good intentions should inform leadership.
They should never substitute for contestable authority.
The Moral Asymmetry Problem
The most dangerous sentence in any technology debate is rarely spoken directly:
We can be trusted with this power. Others cannot.
That belief can feel responsible from inside the institution that holds the advantage. Its leaders see their safety work, internal disagreements, safeguards, and moments of restraint. They see outsiders mainly through threat models: hostile states, reckless competitors, criminals, unregulated developers, or users who may misuse the system.
This creates moral asymmetry. One’s own power appears governed by conscience and expertise. Another person’s access appears governed by worst-case possibility.
History gives us little reason to accept that asymmetry permanently.
Institutions change. Leadership changes. Incentives change. Governments change. Companies are acquired, pressured by markets, redirected by boards, or overtaken by competitors. A system that feels responsible in one moment can become unaccountable in another.
The answer is not to assume that everyone is good.
The answer is to build governance that does not require anyone to be inherently good.
Character is not a control. Accountability is.
That means independent scrutiny, enforceable limits, transparent evidence, plural participation, appeal, redress, and consequences for misuse. These mechanisms should apply to frontier laboratories, governments, enterprise deployers, and individual users according to the authority and harm each can create.
Reader takeaway: AI safety cannot rest on a permanent distinction between responsible holders and irresponsible outsiders.
Pacing Can Protect Incumbents As Easily As It Protects Society
Any policy that slows a frontier changes who can catch up, who can compete, and who can participate.
Large laboratories can absorb the cost of extensive evaluations, compliance teams, secure infrastructure, and long development cycles. Smaller companies, universities, public-interest laboratories, developing economies, and independent researchers may not have the same capacity. A rule designed around the operating model of today’s leaders can turn those leaders into tomorrow’s licensed gatekeepers.
This does not mean every model weight, dangerous capability, or autonomous tool should be released without restriction. It means restrictions need a public justification based on demonstrated capability and consequence, not the institutional prestige of the holder.
The US National Telecommunications and Information Administration’s report on open model weights offers a more balanced posture. It recognises significant risks but also identifies benefits for competition, research, innovation, public institutions, nonprofits, and independent developers. Its approach emphasises continued evidence gathering and the ability to introduce restrictions when evidence warrants them, rather than treating closure as the default.
That distinction is crucial.
Governance should ask:
- What can this system actually do?
- What authority can it exercise?
- What is the plausible consequence of failure or misuse?
- Which safeguards reduce that consequence?
- What evidence shows those safeguards work?
- Who can challenge the decision?
It should not begin with:
- Which company built it?
- Which country controls it?
- Is the holder already inside the frontier club?
Safety should follow capability and consequence.
It should not follow status.
Humanity Has Never Advanced By Eliminating All Dual Use
Scientific and technological progress has always created tension between benefit and misuse.
New knowledge changes medicine, communication, industry, transport, agriculture, and public life. The same underlying capability can create harm when combined with hostile intent, concentrated authority, weak institutions, or poor operating controls.
Humanity’s durable response has not been to ensure that only the discoverer retains the knowledge. Nor has it been to pretend that harm will disappear through optimism.
The stronger response has been institutional learning.
Societies build professional standards. They create testing regimes, safety engineering, licensing for high-consequence activities, liability, monitoring, incident reporting, public oversight, and legal redress. Education spreads capability. Standards make practice more reliable. Independent research exposes weaknesses. New institutions emerge because old ones no longer fit the power of the technology.
This process is imperfect and often late. It remains the mechanism through which civilisation converts dangerous capability into broadly useful infrastructure.
AI should be treated with the same seriousness.
But seriousness is different from technological guardianship.
The frontier should not become a private territory whose current occupants decide when the rest of humanity is ready to enter.
Good Use Requires More Good Users
The positive case for AI cannot be separated from access.
If AI can accelerate disease research, improve education, increase small-business productivity, support public administration, expand scientific discovery, strengthen accessibility, and help communities solve local problems, then beneficial use should not remain concentrated in a handful of companies and states.
Good use needs capable people across many institutions.
It needs doctors who understand local health systems, teachers who understand their students, scientists who can inspect models, entrepreneurs who can build for underserved markets, public servants who know where administration fails, and civil-society organisations that can challenge both government and corporate power.
The UN Advisory Body’s report on governing AI argues for an inclusive and distributed international governance architecture. NIST’s AI Risk Management Framework was developed through an open and collaborative process and calls for diverse perspectives across the AI lifecycle. These approaches recognise that trustworthy governance needs more than technical concentration. It needs participation from those who experience the consequences.
Broad participation also improves safety.
Distributed researchers discover failure modes that internal teams overlook. Domain specialists identify harms that generic benchmarks miss. Smaller organisations build alternatives when dominant platforms make poor choices. Open scrutiny tests claims that closed institutions may sincerely believe but cannot independently prove.
The existence of bad actors is a reason to improve detection, attribution, containment, resilience, and enforcement.
It is not a sufficient reason to deny every beneficial actor meaningful capability.
Reader takeaway: Expanding responsible use is part of the safety strategy, not an exception to it.
Pace Authority, Not Human Progress
The central governance mistake is to treat AI capability as one undifferentiated object.
There is a material difference between a model that helps a scientist analyse public data and an agent that can modify critical infrastructure. There is a difference between generating an explanation and executing a financial transaction. There is a difference between research access and unrestricted autonomous deployment.
The control burden should rise with authority, reach, irreversibility, and consequence.
This produces a more useful principle:
Expand beneficial capability. Narrow authority as consequence rises.
Low-consequence access should remain broad enough to support learning, research, competition, and public benefit. Higher-consequence systems should face stronger identity controls, independent evaluation, red teaming, permission boundaries, monitoring, evidence requirements, incident reporting, and redress.
This approach does not assume that innovation is harmless. It separates the ability to know and create from the authority to act without constraint.
It also creates a better response to uncertainty. When evidence shows that a capability materially lowers the barrier to severe harm, targeted controls can tighten. When evidence shows that access produces broad value with manageable risk, use can expand. Controls become adaptive rather than permanently captured by the fears or commercial structure of one moment.
A Better Governance Compact
The alternative to both uncontrolled acceleration and incumbent-controlled pacing has five parts.
1. Independent Evaluation With Independent Governance
Embedded evaluators are a meaningful step, but access alone does not guarantee independence. Evaluator selection, funding, publication rights, conflicts of interest, and mechanisms for escalating disputed findings need public credibility.
A plural evaluation ecosystem should include technical specialists, domain experts, civil society, academia, affected communities, and public authorities. No single evaluator should become the universal source of truth.
2. Capability-Based Thresholds
Controls should activate when systems demonstrate specific high-risk capabilities or receive high-consequence authority. Thresholds should be testable, revisable, and applied consistently across organisations.
The identity of the developer may affect supervision. It should not replace evidence about the system.
3. Distributed Beneficial Capacity
Governments and institutions should invest in public-interest compute, shared evaluation infrastructure, safety research, education, and access for universities, smaller enterprises, nonprofits, and emerging economies.
If responsible participation remains unaffordable, governance will reproduce concentration even without explicitly requiring it.
4. Controls At The Point Of Consequence
High-impact AI should operate through bounded identities, least privilege, transaction limits, human decision rights, auditable execution, stop mechanisms, and tested recovery. These controls address what a system can actually reach and change.
They are more durable than relying only on a model to refuse a harmful request.
5. Resilience, Enforcement, And Redress
Bad actors will not disappear because responsible companies agree to slow down. Society needs the capacity to detect misuse, attribute attacks, share intelligence, contain incidents, recover services, enforce consequences, and provide redress to people harmed by AI systems.
A governance regime that invests only in prevention will remain fragile.
The NetworkGain View
Amodei is right to reject a race in which capability grows while safety work falls behind.
But a race to the top cannot mean that today’s leaders remain on top and define responsibility for everyone else.
The real race to the top should be a race toward better institutions: stronger evaluation, clearer accountability, wider beneficial participation, more resilient systems, and controls proportionate to consequence.
NetworkGain’s view is that AI governance must connect four questions:
- Benefit: Which human, economic, scientific, or public outcome can this capability improve?
- Authority: What can the system see, decide, trigger, or change?
- Evidence: What proves that safeguards work under realistic challenge?
- Accountability: Who owns the decision, the consequence, the response, and the redress?
This frame avoids two weak extremes.
It does not treat progress as automatically good. Power without limits can create serious harm.
It does not treat concentration as automatically safe. Power does not become legitimate because its holder expresses concern about misuse.
AI is too important to be governed through faith in a few companies. It is also too important to be governed through fear of everyone else.
The path forward is accountable stewardship.
Let useful capability reach the people and institutions that can turn it into health, knowledge, education, productivity, and public value. Apply stronger constraints as systems gain authority over sensitive data, critical processes, physical infrastructure, financial value, or human rights. Test those constraints. Publish meaningful evidence. Give outsiders the power to challenge. Build response capacity for the failures that still occur.
Humanity has not advanced because powerful people promised never to misuse discovery.
It has advanced because knowledge spread and institutions slowly learned how to govern consequence.
AI should be no different.
The frontier does not need a moral owner.
It needs accountable stewards.
Research Note
This perspective responds to Dario Amodei’s September 2026 essay, We Must Pace the Frontier, using the user-provided page capture and the original published post. It also draws on public-interest governance approaches from NIST, NTIA, and the United Nations. The critique addresses the structure and implications of the proposal, not Amodei’s motives or sincerity.