Pacing the AI frontier: a signal, not a strategy — what executives must do now

CEOs say we must “pace the frontier.” The pledge is a signal, not a strategy.

Some researchers at Anthropic recently raised stark warnings about long‑term risk, and Anthropic CEO Dario Amodei followed with a public proposal titled “We Must Pace the Frontier.” Senior leaders at several major AI firms quickly endorsed the idea: OpenAI’s Sam Altman, Google DeepMind’s Demis Hassabis, and Elon Musk of xAI all signaled support for Amodei’s framework.

That sequence matters. But rhetoric alone won’t change incentives. The proposal is useful as a conversation starter; it is not, by itself, a binding mechanism that will slow frontier AI. I’ll explain why the pledge is a meaningful signal but without defined, enforceable mechanisms it risks becoming PR. Then I’ll lay out what a credible pacing regime would actually require, the scenarios most likely to play out, and practical steps business leaders should take now.

What Amodei proposed, in brief

Independent outside evaluators with “employee‑like access” to assess models and safety practices.
Agreements among democratic states to align on common safety standards.
Global coordination that includes engagement with authoritarian states on security issues.

On social platforms, industry leaders echoed the call. Sam Altman wrote on X:

“I agree with Dario that we need to pace the frontier. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same.”

Demis Hassabis: “Dario’s essay points towards the right path forward.”

Elon Musk: “Dario is right.”

Why the proposal matters, and why it’s incomplete

The three pillars map to familiar safety tools: external audits, harmonized standards, and international cooperation. Those tools have worked in other high‑risk industries, but only when backed by clear authority, transparent governance, and enforceable penalties.

“Employee‑like access” is a crucial phrase that needs a concrete definition. Does it include model weights, training datasets, fine‑tuning logs, red‑teaming transcripts, compute‑use invoices, and contract terms with cloud providers? Each item raises different legal, IP, and national‑security questions. Without specificity, the phrase is a promise without a way to measure it.

Likewise, “shared safety standards” must become measurable tests: adversarial robustness assessments, standardized red‑team suites, capability taxonomies, or compute‑threshold triggers tied to mandatory review. Otherwise, “standards” can be vague commitments that mean little in practice.

Politics and power: the backdrop that will determine outcomes

The U.S. political environment currently favors competitive advantage over precaution. President Donald Trump has framed AI as a geopolitical prize, “whoever wins AI, wins”, and suggested guardrails are possible but not urgent. Vice‑president JD Vance has warned against safety‑first framing, saying AI won’t be won by “hand‑wringing about safety.” House Speaker Mike Johnson said Congress would not lead on regulation.

That posture matters because law shapes enforcement. Industry lobbying has already influenced legislative outcomes, and political leaders signaling that competition outranks caution make binding, precautionary rules harder to pass without a catalyzing event.

Signals vs. substance inside the companies

Public commitments are valuable, they lower coordination costs and can change norms, but they are also inexpensive signaling devices. Two signals to watch that reveal whether the rhetoric will match reality:

Organizational choices: Reporting has suggested cuts or reorganization of long‑horizon safety teams at major firms, including reductions tied to “superalignment” research efforts. If companies shrink independent research capacity while pledging external evaluation, the balance of effort matters.
Political influence and money: AI‑linked political spending is already significant. For example, Alex Bores’s congressional bid attracted roughly $24 million in AI‑linked PAC spending for and against him, a reminder that policy outcomes will be fiercely contested.

Critics worry the proposal could become a vehicle for regulatory capture. As Sarah Myers West of the AI Now Institute put it:

“AI exceptionalism has led to a total failure to uphold our standards for what products are deemed safe for use.”

And Alvaro Bedoya, the Biden‑era head of the Federal Trade Commission, has argued two linked points: antitrust law does not prevent firms from coordinating to reduce safety harms, but it also prevents coordination designed to shut out cheaper, upstart rivals. Any safe‑harbor that enables industry coordination will therefore need tight guardrails to prevent entrenchment.

What a workable pacing regime would actually require

To be more than a press release, a pacing regime needs legal force, transparent processes, and clear metrics. At minimum:

Define the scope of evaluator access. Specify whether access includes model weights, training data, compute logs, red‑teaming outputs, and cloud contracts. Each should come with data‑handling rules, non‑disclosure protections, and forensic safeguards.
Establish accreditation and selection rules. Create an independent accreditation board that is publicly accountable and certifies evaluators by technical competence, conflict‑of‑interest checks, and rotational appointment. Selection should be transparent and open to appeal.
Separate funding from influence. Use a pooled funding model, such as industry contributions into a neutral trust, or public financing for independent labs. Firewalls must prevent donors from vetoing findings. Independent auditors should report to regulators or a public registry.
Build clear, testable benchmarks. Examples include performance on a standardized red‑team suite, a capability taxonomy tied to compute‑use thresholds, and tests for deceptive or emergent instrumental behavior. Benchmarks determine when mandatory pauses or escalations are triggered.
Design antitrust safe harbors with limits. Any exemption allowing safety coordination should include oversight by competition authorities, narrow scope, sunset clauses, and transparency rules to prevent anticompetitive closures.
Pursue narrow, pragmatic international cooperation. Focus first on verifiable, low‑ambiguity areas: incident reporting and forensic cooperation, joint incident response protocols, and agreed minimum disclosure for dual‑use capabilities. Broad treaties will be politically difficult; start with smaller, enforceable steps.

Three likely near‑term scenarios, and what will signal each

Symbolic restraint (most likely), High‑profile pledges, limited audits, and voluntary disclosure. Signal: short, nonpublic reviews and publicity campaigns framed as “independent” but lacking accredited evaluators or published findings.
Industry‑managed regime (plausible), Firms create a shared standards body, potentially secured with limited antitrust relief. Signal: a trade group or industry board publishes standards and controls accreditation; transparency is limited and participation rules favor incumbents.
Hard regulation (possible after a trigger), Government mandates for licensing, mandatory third‑party audits, export controls, or criminal penalties for noncompliance. Signal: bipartisan bills advancing in committee, regulators issuing binding guidance, or a high‑profile safety incident prompting rapid legislative action.

Practical steps for executives and boards, act now

Assume uncertainty and build optionality. Prioritize near‑term actions with clear timelines:

Within 90 days: Require independent audits for any externally deployed model. Publish a one‑page, non‑technical summary of audit findings for customers and regulators.
Within 6 months: Appoint or hire a senior safety officer reporting to the board (not to product), and fund a small, sustained alignment research capability focused on explainability and adversarial testing.
Ongoing: Engage proactively with policymakers and standards bodies; don’t delegate rule‑making to trade groups alone. Advocate for accreditation, public reporting, and verification mechanisms you would accept for your own products.
Scenario planning: Prepare workforce and product plans for three futures, light regulation, industry‑managed standards, and strict government controls, and run quarterly exercises that stress‑test each plan.

Key questions, answered

Will CEOs’ public support for pacing slow development in practice?

Signal value: high. Binding effect: low, medium. The pledge reduces coordination costs and can change norms, but without legal authorities, accreditation, and enforcement, words risk outpacing action.
Are independent evaluators sufficient to ensure safety?

Not by themselves. Evaluators need legally defined access, transparent selection, independent funding, and enforcement backstops, otherwise results may be non‑binding or hidden.
Could the proposal become a cover for anti‑competitive coordination?

Yes. Antitrust law permits narrow safety coordination but prohibits arrangements that shut out competitors; any coordination vehicle needs oversight, narrow scope, and sunset clauses to prevent entrenchment.
How realistic is coordination with authoritarian states like China?

Hard and limited. Expect incremental cooperation on narrow, verifiable issues (incident response, forensics) rather than broad, enforceable pacts. National security and economic incentives create significant friction.
What should executives do right now?

Act as if some form of meaningful regulation is coming: require independent audits within 90 days, create board‑level safety oversight, fund internal alignment work, and engage regulators to shape enforceable standards.

If this pledge leads to anything substantive, it will be through the hard work of defining access, building accredited evaluators, funding independent verification, and creating legal backstops that bind behavior. Until then, treat public commitments as early signals, useful, but provisional. Boards and executives who want stability should stop hoping others will do the hard design work and start building the structures that make pacing real.