Slowing the Pace of AI Development: AI May Slow Down Even If the Technology Doesn’t

The argument over frontier model pacing overlooks the speed at which societies and enterprises can absorb what AI already does.

Dario Amodei’s essay We Must Pace the Frontier turned “pace” into the AI word of the week. He didn’t call for research to stop. He proposed giving safety work more time to catch up with capability, beginning with embedded independent evaluators at frontier labs, followed by coordination among companies in democracies and, eventually, some form of global coordination that includes China.

The reaction has been mixed. Much of the immediate coverage treated the proposal as a contest between catastrophic risk and geopolitical advantage. Amodei points to recent agent behavior, cyber and biological misuse, and the possibility that AI will help build its own successors faster than humans can understand or control them.

The White House, on the other hand, continues to frame AI as a race the United States must win. At the same time, Congress is considering responses that range from testing and reporting requirements for frontier systems to an outright ban on artificial superintelligence. As I argued in “Why AI Will Humble Regulators,” AI’s shifting and often emergent behaviors make it difficult to define when something like “superintelligence” has occurred—if it ever does.

The debate is important, but it isn’t as complex as it needs to be. It assumes that the only pace worth managing is the pace at which laboratories build more capable models.

There is another frontier: the rate at which businesses, public institutions, and communities can responsibly absorb the capabilities already available. That frontier is well behind the technical one. Several recent surveys suggest that AI governance lags behind implementation. A more measured release cycle could improve the odds that AI produces value because buyers would have time to redesign work, repair data, build governance, train people, and learn what the technology actually changes.

The United States does not control the clock

AI is global even if the largest frontier labs remain concentrated. The Stanford AI Index 2026 reports that the United States still produces more top-tier models, but the performance gap between leading US and Chinese models has nearly closed. Model production remains concentrated in the United States and China, while open-source development is broadening participation elsewhere.

That limits the effectiveness of unilateral pacing as a global governance mechanism. A US lab that slows while its competitors continue has made a business decision, not established a global standard of behavior. A US law would reach American companies, but it would not govern a Chinese lab, a European provider, or an open model adapted elsewhere without meaningful geopolitical negotiations. It could, however, still shape foreign behavior indirectly through dependencies on compute, capital, cloud, export controls, and market access.

Amodei’s proposal calls for cooperation with China while maintaining restrictions on advanced chips, semiconductor equipment, model distillation, and weight theft. From a US security perspective, those ideas can be presented as complementary. Chinese officials may view the package as asking Beijing to accept rules substantially shaped by US strategic and commercial interests. China’s Ministry of Foreign Affairs responded by calling the framing counterproductive and arguing that confrontation would disrupt global AI governance, according to the Associated Press.

Europe will define its own version of acceptable pace. The European Union already requires providers of general-purpose models with systemic risk to engage with the AI Office early in development and supports third-party involvement across the model lifecycle through its General Purpose AI Code of Practice. Europe may agree with more evaluation while declining to subordinate its regulatory calendar, sovereignty concerns, or risk definitions to Washington or Silicon Valley.

The United States also cannot assume that allies and trading partners will follow because American firms remain ahead. Brand Finance’s 2026 Global Soft Power Index (an indirect indicator of diplomatic receptivity, not a measure of AI policy alignment) still ranked the United States first but reported the largest year-over-year decline among the countries it measured. China moved ahead of the United States on the index’s reputation measure. One survey does not settle the state of American influence, but it does challenge any strategy that assumes automatic alignment by allies.

Because a globally enforceable technical pause would prove difficult to negotiate and verify, the more immediate and actionable question should be: how do organizations leverage uneven periods of stability when they occur?

Global pacing, therefore, begins as a diplomatic problem. It requires trust, reciprocity, and verification among governments that disagree about security, sovereignty, trade, and political legitimacy. Model capability is likely to continue to improve throughout those negotiations.

The enterprise frontier is already overloaded

The public debate says relatively little about the organizations expected to turn new models into economic value.

Enterprises do not implement a model. They change a process. That means understanding the current work and its full cost, designing the new work, integrating systems, revising policy, training people, defining decision rights, and measuring the result. A chatbot can respond in seconds. None of those activities take place at chatbot speeds.

In a recent interview, I described a public-sector customer-service implementation that had to be rewritten several times as the technology evolved. The initial system looked good for 2021. Better models then revealed that the solution had been suboptimized. New capabilities made more ambitious work possible, but they also forced the team to revisit architecture and implementation. Every version improved; every version consumed more transformation capacity.

That pattern is not an argument against improvement. It is an argument for accounting. If a new model reduces the time required for one task but creates more review, resubmission, auditing, security work or downstream correction, the organization must account for costs across the entire process. If a manager once trusted an employee’s output after a quick glance but now spends an hour verifying an AI-assisted draft, some of the apparent productivity gain has simply moved up the hierarchy.

I have made the same point in AI Use Is Not AI Value. Tokens, prompts, active users, and agent runs are telemetry. Value appears in changes to cycle time, quality, cost, rework, risk, resilience, customer experience, and employee capacity. Accounting becomes harder when model vendors change prices, features, interfaces, and limitations faster than buyers can establish a baseline.

Pacing could create a period of relative stability in which organizations can better plan and complete AI projects. They could push bounded use cases into production, observe them long enough to see failure patterns that might be masked by short evaluation windows, update knowledge and policy, and, most importantly, compare the new process with the old one and decide whether AI delivered on its promises. That would be more useful than adding another wave of pilots every time a benchmark moves. That benefit depends on the scope of the intervention; however, slowing frontier training or release alone may not stabilize the surrounding application and vendor ecosystem.

Pacing can improve the odds of success

The usual argument for slowing the frontier is that alignment, interpretability, and evaluation need time. Amodei makes that case directly. The enterprise case is parallel but distinct: organizational learning also needs time.

In Enterprise AI Insights from the Field, I argued that durable returns require narrow scopes, workflow integration, strong data foundations and governance. Those conditions are neither glamorous nor optional. They are also difficult to build while leaders are reacting to a fresh model, a new agent framework, and a revised vendor roadmap every few months.

A slower frontier would not guarantee better enterprise outcomes. Some organizations would use the time poorly. Others would treat a pause in headline releases as evidence that they could stop learning. The benefit depends on how buyers use the break.

It is important that organizations capture a baseline state, including labor, technology, service, and risk costs. They should redesign the process rather than automate the current steps. They should measure rework and transferred work, not only the time saved by the first user. They should make training continuous and require people to demonstrate AI competence in their work. They should also preserve the ability to change models or reverse an implementation when assumptions fail.

That work can create value even when attribution remains messy. AI may force an organization to confront a process it has not examined in years. The improved outcome may stem as much from managerial permission, process redesign, better data, and new knowledge practices as from the use of AI. That is not a failure of AI ROI. It is a more honest description of transformation.

The point is not to give slow projects unlimited time. An initiative that produces no useful results, no reusable knowledge, and no improvement in the organization’s ability to act is not succeeding. The evaluation window should be revised as part of an explicit learning process, not prematurely curtailed to protect a forecast.

The binding constraint may be social permission

From a scenario planning perspective, “pacing the frontier” combines several uncertainties that are usually discussed separately. The technology is not the constraint in this scenario. Models continue to improve. The binding constraints come from political legitimacy, social acceptance, infrastructure, and the capacity of institutions to absorb change.

Fear of catastrophic AI is one source of pressure, but it is not the only one. Public debate and local political disputes also center on jobs, privacy, fairness, concentrated corporate power, the environmental impact of AI infrastructure, and errors in consequential decisions. Those concerns do not require belief in human extinction to influence policy or purchasing.

Data centers make an abstract technology physically local. A nationally representative Annenberg Public Policy Center survey conducted in June and July 2026 found that 61 percent of US adults opposed new data centers in their area, up 12 percentage points from the spring. The same survey found that 68 percent believed the government had done too little to regulate AI. The constraints show up in local politics through zoning hearings, electricity rates, water use, construction delays, and demands that communities share in the benefits.

Congress is responding through competing proposals. Senators from both parties have discussed requirements for frontier developers to test and report catastrophic risks, while Senator Bernie Sanders and Representative Greg Casar announced a Ban Artificial Superintelligence Act that would pause advanced development until a federal regulator establishes safety rules and permanently prohibit systems that meet the bill’s definition of superintelligence. President Donald Trump has rejected new guardrails in recent remarks, while the administration’s AI Action Plan calls for faster innovation, expanded infrastructure and American leadership. These positions place acceleration, oversight, and local resistance inside the same political system.

The result may not be a coordinated slowdown. It may be uneven friction: one state imposes audits, another expedites a data center, one country restricts a model, another subsidizes an open alternative, and enterprises narrow projects because they cannot govern or justify them. That future resembles the Fragmented Intelligence scenario I have used in my State of AI reports. AI advances, but access, rules, economics, and adoption diverge by region and institution.

Pacing also raises questions about power

The current discussion must also ask who benefits from a coordinated pace. The Verge sharpens the question by asking whether a slowdown among the dominant AI companies would be a safety pact or a cartel. The useful answer is not either-or. Independent evaluation, incident reporting, and limits tied to measurable risk can improve accountability. Even if the warning of catastrophic outcomes is sincere, resulting arrangements may still protect incumbents if compliance costs keep smaller competitors out or if an agreement freezes the current order.

The Verge reports that many researchers see real value in the proposal while also worrying about safety-washing: auditors and paperwork could create the appearance of restraint without materially changing the race. Critics also question whether rules shaped by the largest labs would disadvantage new entrants or open-source developers and head off less industry-friendly safeguards. Those concerns do not establish that the labs are forming a cartel. They do establish that motive is the wrong test. The structure and effects of the agreement matter more than whether its sponsors describe themselves as frightened, responsible, or altruistic.

Anthropic and OpenAI are preparing for possible public listings. That does not invalidate their safety arguments, and commercial incentives do not prove that a warning is insincere. It requires examining competitive motivations and incentives. An evaluation of any pacing framework would need to identify who defines the thresholds, who selects and pays the evaluators, what evaluators may publish, what evidence would demonstrate that development actually slowed, how open models are treated, whether new entrants can comply, and how the arrangement changes competition and risk.

This is consistent with the argument I made in Why AI Will Humble Regulators. Audits help, but they certify a changing environment. Safety does not establish suitability for a particular use. Regulation can protect the public while also serving as a defense for companies able to absorb the costs and influence the details of the standard.

Pacing frontier labs also does not pace every application already in the market. Existing models can be connected to new data, tools, and authority. Open models can be modified and redistributed. Much of the next wave of risk may come from assembly and deployment rather than another jump in model intelligence. A credible framework must include approaches that govern the pace of training, release, diffusion, integration, and delegated action.

Several clocks will govern the future of AI

The frontier debate is asking the right questions too narrowly. AI capability may advance faster than developers can adequately evaluate, govern, and align it for consequential uses. It will also likely advance faster than governments can negotiate, communities can consent, infrastructure can expand, and enterprises can transform.

No single actor controls all of those clocks. The United States can influence access to chips, the behavior of domestic labs, and the terms on which the government buys AI. It cannot dictate China’s strategy or replace Europe’s regulatory institutions. We can no longer assume that American soft power will drive global coordination. Frontier companies can slow releases, but they cannot stop capable models already in circulation from being applied in new ways. Although enterprises can narrow deployments, they cannot make the vendor ecosystem stand still.

The debate can therefore be evaluated by a broader test: what becomes more understandable, governable, and dependable during the additional time. For frontier labs, the evidence would include evaluation, incident reporting, and operational discipline. For governments, it would include enforceable definitions, international verification, and clarity about responsibility. For communities, it would include documented participation in infrastructure decisions. For enterprises, it would include redesigned work, better data, trained people, and evidence that AI improves outcomes rather than dashboards.

If those things improve, a slower frontier could produce more successful AI adoption. If they do not, “pacing” will become another word for a temporary advantage held by whoever was ahead when the clock changed.

For more serious insights on AI, click here.

Did you find this analysis of the 2026 IBM CEO Study useful? If so, please like, share or comment. Thank you!

The cover image is AI-generated from the author’s prompt.