Anthropic Google Leaders Ask Washington to Put Brakes on Automated AI
Anthropic Google leaders joined more than 1,200 frontier AI workers in an unusual request: prepare a coordinated way to slow automated AI development. The signatories include senior researchers from OpenAI, Meta, Thinking Machines, Microsoft, Mistral, and other laboratories. Their concern is not ordinary chatbot adoption. They fear AI systems will soon help design stronger successors faster than people can evaluate them.
The July statement, called “Pacing the Frontier,” does not demand an immediate pause. It asks the US government to support an international effort for controlling the pace of advanced AI development. The organizers say companies and countries cannot safely slow alone because every participant faces pressure to keep racing.
That distinction matters. These researchers are not asking Washington to stop a particular model release today. They want institutions, technical controls, and international agreements ready before automated AI research creates an emergency. The request pits collective risk management against the competitive logic governing OpenAI, Anthropic, Google, Meta, and their international rivals.
Anthropic Google Researchers Join a Cross-Industry Warning
The statement turns private concern inside frontier laboratories into a public request for government coordination.
The live frontier statement listed 1,293 verified employees when accessed on July 30, 2026. That count can continue changing because eligible workers may still sign. Organizers require a corporate email or other employment evidence before verifying a participant.
The signatories span companies that compete for researchers, customers, computing capacity, and technical leadership. They include Anthropic CEO Dario Amodei and co-founders Jared Kaplan, Jack Clark, Chris Olah, and Benjamin Mann. OpenAI chief scientist Jakub Pachocki and chief research officer Mark Chen also signed.
Google representation includes DeepMind co-founder Shane Legg, strategy chief Jasjeet Sekhon, and AI safety vice president Anca Dragan. Meta chief scientist Shengjia Zhao and AI research vice president Dawn Song appear as well. Thinking Machines chief scientist John Schulman and Safe Superintelligence CEO Ilya Sutskever add further institutional weight.
The presence of executives and research leaders makes the document harder to dismiss as a protest by employees far from model development. However, it is still a self-selected group. The count does not represent a workforce vote at Anthropic, Google, OpenAI, Meta, or any other company.
Comments on the statement are explicitly personal and do not necessarily reflect employers’ positions. That qualification separates individual participation from formal corporate policy. It also prevents the signatures from becoming a binding commitment by any laboratory.
The statement makes one central request. It asks Washington to support an international effort that develops technical and governance tools for deliberately pacing automated frontier AI. “Pacing” means retaining the ability to slow capability development when evidence shows that oversight cannot keep up.
Its target is narrower than AI development as a whole. Automated AI research means using AI systems to perform substantial parts of the work required to create better AI. That work can include coding experiments, designing evaluations, finding software flaws, interpreting results, and proposing model improvements.
The organizers argue that this feedback loop presents a distinct risk. An AI system that accelerates ordinary office work changes productivity. A system that accelerates its own field can shorten the time between major capability advances.
This is why the letter is more consequential than another broad call for responsible AI. It identifies a potential transition point and asks governments to prepare before crossing it. The signatories believe preparation cannot wait until a dangerous system is already running.
The request also arrives amid a wider shift in Washington. The US government has moved from general AI principles toward direct access, testing, and national security reviews. That creates a possible institutional foundation, but not yet an international pacing mechanism.
Why Automated AI Research Changes the Risk
Automating AI research compresses the time available for testing, public debate, and government response.
Current frontier models already assist researchers with programming, literature analysis, mathematical reasoning, and experiment design. These tools remain inconsistent, and human researchers still direct important work. The statement does not establish that a self-improving system already exists.
Instead, it focuses on what happens when those capabilities become reliable enough to automate a large share of development. A laboratory could use one generation of models to improve training software or identify new research directions. The resulting model could then contribute more effectively to the next generation.
That process does not require a machine to rewrite itself without human involvement. A company can preserve human approval while still accelerating each research cycle. Faster cycles leave evaluators, security teams, and policymakers less time to study unfamiliar behavior.
The letter says leading companies believe they are approaching the automation of AI research. It also acknowledges uncertainty about how strongly that transition would accelerate progress. That uncertainty is central, not incidental, to its case.
If automation produces only incremental productivity gains, aggressive pacing rules could impose unnecessary costs. They might delay useful medical, scientific, accessibility, and cybersecurity applications. They could also concentrate influence among companies already wealthy enough to satisfy complex requirements.
If automation sharply accelerates capabilities, waiting for certainty becomes its own gamble. Governments may discover that their evaluation methods, legal powers, and international channels take years to establish. Model development can move much faster.
This timing mismatch explains the request to “buy time.” The phrase does not describe delay as an end in itself. It describes a temporary interval for investigating emerging risks, improving security, and strengthening oversight.
John Schulman said his signature was intended to create shared awareness about the possible need for coordination. He also argued that laboratories should start designing mechanisms voluntarily before the US government intervenes. His comment points toward preparation rather than an unconditional halt.
Ilya Sutskever offered a sharper warning. He said future AI will require unprecedented measures and stressed that international implementation must be done well. A poorly designed system, he warned, can make the situation worse.
Their positions expose two separate uncertainties. Researchers do not know precisely when automated research will become transformative. Policymakers also do not know whether institutions built to control it will remain effective, legitimate, and internationally credible.
This is the mechanism behind the concern. More capable AI supports faster AI research, which produces more capable AI. The risk comes from the cycle’s speed and unpredictability, not from any single chatbot feature.
The public should therefore resist two easy conclusions. The statement does not prove an intelligence explosion is imminent. It also cannot be dismissed merely because the exact acceleration rate remains unknown.
A useful test will involve measurable research performance. Independent evaluators need evidence that models can complete long, novel research tasks with limited intervention. Benchmarks alone will not settle the question because laboratories can optimize for known tests.
Real projects matter more. A model that independently identifies valuable research questions, runs experiments, and validates results would support the signatories’ concern. Occasional coding success would not establish the same case.
The Anthropic Google Coalition Faces a Coordination Trap
The primary conflict is not safety against innovation. It is collective restraint against incentives that punish the first participant to slow down.
Anthropic Google researchers can agree that preparation is necessary while their employers continue competing intensely. The same tension applies to OpenAI, Meta, Microsoft, Mistral, and Thinking Machines. Each company wants safer development, but none wants rivals to gain an uncontested lead.
The statement calls this pressure both corporate and national. A laboratory that delays a release can lose customers, investment, talent, and influence over technical standards. A country that restricts domestic laboratories can fear losing strategic ground to foreign developers.
This creates a coordination trap. Every participant benefits if serious risks are managed across the frontier. Yet each participant bears immediate costs when it acts alone.
Voluntary promises help only when companies can observe compliance and trust the measurements. Frontier development occurs behind closed doors, using different architectures, data, computing systems, and internal evaluations. A company cannot easily verify whether a rival has honored a private slowdown.
Governments face the same problem internationally. An agreement would need a definition of covered development, credible monitoring, and consequences for violations. It would also need rules for academic research, open-weight models, military systems, and smaller laboratories.
“Frontier AI” means the most capable class of models at a given time. The category moves as technology improves. A static list of covered systems would become outdated quickly, creating gaps or capturing ordinary products unnecessarily.
Compute monitoring offers one possible signal because advanced training requires significant hardware and energy. However, training size does not map perfectly to capability. Algorithmic improvements can produce stronger results without a proportional increase in computing resources.
Capability evaluations offer another route. Regulators could apply stricter controls when systems cross thresholds in cyber operations, biological assistance, deception, or AI research. Yet evaluations remain vulnerable to incomplete coverage, gaming, and behavior that appears only after deployment.
Washington has already moved toward limited review. A June executive order created a voluntary process for certain US laboratories to share advanced cyber models before release. According to the federal review details, government examination can last up to 30 days.
That order is narrower than the signatories’ request. It focuses on particular national security capabilities and voluntary participation. It does not establish a general authority to pace automated AI research across countries.
President Donald Trump also framed the competing concern clearly. He said he did not want oversight to interfere with the American lead over China and other countries. That position makes an international request politically difficult, even when Washington accepts some domestic review.
The United States wants stronger visibility into frontier systems while preserving national leadership. Developers want predictable rules without exposing valuable intellectual property or losing release speed. International competitors want assurances that “safety” will not become a tool for preserving American advantages.
Those incentives explain why the statement asks for tools rather than immediate restraint. Technical monitoring, shared evaluations, and diplomatic channels must exist before coordinated pacing becomes credible. Otherwise, a call to slow down remains a request for trust among competitors.
The Tradeoff Is Control Versus Concentrated Power
A government mechanism capable of slowing frontier AI can reduce genuine risks while creating new opportunities for political control and regulatory capture.
The signatories describe a world without adequate pacing tools. Their proposed destination remains less defined. The statement does not specify who would trigger a slowdown, which evidence would qualify, or how long restrictions would last.
It also leaves enforcement unresolved. An international agreement needs rules for laboratories that refuse to participate. Controls limited to American companies would not produce frontier-wide pacing, especially if development continued elsewhere.
The absence of operational detail does not invalidate the warning. Short public statements often seek consensus around a problem before negotiating policy. Still, the lack of specifics prevents readers from evaluating proportionality, accountability, or technical feasibility.
One risk is regulatory capture. Large laboratories already possess extensive policy teams, security staff, and government relationships. Complex approval systems can raise barriers for smaller competitors while allowing established companies to shape compliance standards.
A pacing regime might also protect closed laboratories from open models. Governments could restrict downloadable systems more aggressively because their distribution is difficult to reverse. That could reduce public access to research while strengthening a small group of approved providers.
The reverse risk also exists. Exempting open models could leave a major hole if downloadable systems reach dangerous capability thresholds. Any workable system must address capabilities consistently without treating licensing style as a substitute for evidence.
Civil liberties create another concern. Advanced AI research includes code, scientific communication, and publication. Broad restrictions could affect speech, academic inquiry, and international collaboration. Narrow definitions and independent review would therefore matter.
Government discretion also carries political risk. The current review framework has already raised questions about which models qualify and which partners receive early access. Vague authority can favor allies or punish companies involved in disputes.
The concern is not theoretical. Frontier models can affect cybersecurity, defense, intelligence, and economic competitiveness. Those interests create strong incentives for secrecy and executive control, even when an oversight process begins voluntarily.
Demis Hassabis has proposed a more structured alternative. The Google DeepMind CEO wants an industry-funded standards body operating under government supervision. His watchdog proposal would begin with voluntary testing and could later support mandatory requirements.
Hassabis envisions independent experts alongside industry, government, and open-source representatives. Covered systems would face testing before release, regardless of origin or whether their model weights are available. Thresholds would change as capabilities advance.
OpenAI has proposed another institutional route. Its governance blueprint calls for a national framework, a stronger federal AI safety institute, and wider government resilience planning. That approach emphasizes durable institutions rather than a single emergency power.
These proposals provide more detail than the employee statement, but difficult questions remain. Who appoints independent experts? What information becomes public? Can laboratories appeal decisions? What prevents temporary controls from becoming permanent industrial policy?
The most credible system would separate evaluation from political enforcement. It would publish thresholds and procedures before companies cross them. It would also require regular review, independent audits, and clear expiration conditions.
International legitimacy matters just as much. A US-designed process cannot simply declare itself global. Other governments need meaningful participation in standards, evidence review, and enforcement.
The signatories’ core tradeoff is therefore more complicated than “slow down or lose control.” Society must compare technical risk with concentrated institutional power. Both sides require safeguards.
What the Statement Does Not Prove
The signatures show serious concern among informed workers, but they do not establish the probability, timing, or controllability of runaway development.
The 1,293 participants include researchers with direct knowledge of frontier systems. That makes their warning valuable evidence about internal expectations. It does not turn those expectations into independently verified forecasts.
The group is self-selected. Employees who work in alignment, security, evaluations, and governance may be more likely to sign than colleagues focused on products or infrastructure. Readers should not treat the total as representative polling.
The signatory list also combines current workers across many companies. A large combined number can obscure the share represented at each employer. Company size, job function, and seniority matter when interpreting apparent internal support.
The statement’s central premise uses cautious language. It says companies believe they “could be close” to automating AI research. It does not define “close,” quantify expected acceleration, or present evaluation results supporting that timeline.
Some personal comments use much stronger language. One OpenAI employee compared the competition to a deadly race toward an intelligence explosion. That comment illustrates concern, but it is not the statement’s shared factual conclusion.
The difference matters because dramatic forecasts can shape regulation before independent evidence becomes available. Laboratories also benefit commercially when customers, governments, and investors view their systems as approaching extraordinary capability. Safety warnings and marketing incentives can coexist.
This does not mean the concern is insincere. Employees can genuinely fear a technology while companies benefit from describing it as advanced. Policy should therefore rely on reproducible evidence rather than confidence, reputation, or alarming metaphors.
Automated research itself is also a spectrum. AI can improve code completion without choosing research goals. It can solve benchmark tasks without managing a months-long project. It can suggest experiments without reliably detecting flawed evidence.
A slowdown trigger needs to distinguish these stages. Otherwise, ordinary productivity improvements might activate extraordinary controls. Conversely, an overly narrow definition might miss dangerous combinations of individually modest capabilities.
Research evaluation presents another obstacle. Laboratories control access to their strongest models and internal data. Independent evaluators may receive limited testing time, filtered interfaces, or incomplete information about system behavior.
Public benchmarks degrade as evidence because developers train against them. Confidential tests reduce gaming but weaken external scrutiny. Real-world trials provide richer information but can expose people and infrastructure to risk.
The statement also assumes that pacing can buy useful time. That depends on governments and institutions using the interval effectively. A pause without investment in evaluations, security, emergency planning, and international diplomacy merely delays the same problem.
Enforcement can shift development into less visible environments. Restrictions might encourage companies or states to conceal training, divide projects across entities, or misreport capabilities. Monitoring tools must anticipate evasion without creating universal surveillance.
China remains the central geopolitical uncertainty for US policy. American officials routinely frame frontier AI as a strategic competition. Any global mechanism must provide both sides with credible verification and protection against industrial espionage.
Agreement among US laboratory employees does not establish international agreement. It does, however, identify why unilateral company pledges are insufficient. The relevant unit of action is the frontier, not a single employer.
Readers should therefore interpret the letter as a demand for preparedness, not proof that a crisis has begun. Its strongest claim concerns institutional readiness. The world currently lacks an agreed process for slowing automated AI research if evidence later demands it.
Three Signals Will Show Whether Pacing Becomes Policy
The next test is whether a three-paragraph statement produces measurable evaluations, formal institutions, and international participation.
The first signal is a concrete federal standards body. Hassabis wants such an organization operating before the end of 2026. Its design will show whether Washington favors independent technical review or direct executive control.
A credible proposal should define appointments, funding, disclosure, appeals, and conflicts of interest. It should also explain how thresholds change over time. Vague authority would weaken the case that government involvement produces predictable oversight.
Mandatory participation would mark a major shift. The June review process remains voluntary and focused on particular cyber capabilities. Expanding it to automated research would bring the government deeper into laboratory development schedules.
That change would strengthen the statement’s influence, but not necessarily its policy quality. The key question is whether authority comes with public standards and independent accountability. An opaque process would exchange technical uncertainty for political uncertainty.
The second signal is evidence from real AI research tasks. Laboratories and independent evaluators need to show how models perform across long projects, not isolated coding tests. Results should measure autonomy, novelty, reliability, and the amount of human correction required.
Evidence of models completing substantial research cycles would support the case for advance coordination. Continued dependence on close human direction would weaken claims that an uncontrollable acceleration is near. Either result would improve the policy debate.
Reporting failures matters as much as publishing successes. Evaluators should document abandoned experiments, fabricated results, security violations, and interventions. Selective demonstrations cannot establish the operational reliability required for automated research.
The third signal is international engagement, especially from other major AI-producing countries. A US-only framework cannot deliver frontier-wide pacing. It can establish domestic visibility, but competitors would still determine whether restraint remains sustainable.
Early diplomatic progress does not require a comprehensive treaty. Shared evaluation methods, incident reporting, researcher exchanges, and emergency communication channels would be meaningful starts. Agreement on narrow cyber or biological thresholds could test broader coordination.
The European Union’s involvement would add regulatory experience and market influence. Cooperation with the United Kingdom, Japan, South Korea, and Canada could broaden technical capacity. The hardest test remains engagement with China.
International participation would strengthen the signatories’ central judgment that competitive pressure requires collective action. Failure to attract other major developers would expose the limits of a Washington-led pacing strategy.
Developers and enterprise buyers should follow these signals closely. New evaluation rules can change release schedules, model access, audit duties, and procurement requirements. They can also determine which systems remain available for sensitive research or security work.
Knowledge workers should care because faster AI research eventually affects every downstream model. A compressed development cycle can deliver better tools sooner, but it can also reduce the time available to understand failures. Product teams may inherit risks that frontier laboratories have not fully mapped.
Organizations should preserve records of model behavior, approval decisions, and unexpected outcomes. A searchable AI knowledge base can help teams compare incidents across model versions. Internal evidence will matter if external rules begin requiring stronger documentation.
The Anthropic Google coalition has not won a slowdown, and it has not defined one. It has made the absence of a credible pacing mechanism harder to ignore. The people building advanced systems are asking governments to prepare an option their employers cannot exercise alone.
The next three months should reveal whether that request becomes an institution or remains a symbolic letter. Watch for published thresholds, independent research evaluations, and international commitments. If those pieces do not appear, “pacing” will remain a compelling word without an operational meaning.
The question for policymakers is no longer whether every forecast is certain. It is whether uncertainty justifies building reversible safeguards before automated research arrives. For Anthropic, Google, OpenAI, and their rivals, the harder question is whether they will accept those safeguards when competition becomes costly.
