NextFin News - Google senior vice president James Manyika says artificial intelligence is too important to be left to industry self-policing, arguing that only the "collective will of society, governments and companies" can manage the technology's risks. The remarks, aired October 2, land as a direct challenge to the voluntary safety accord President Donald Trump secured from AI executives at the White House just days earlier.
Manyika's position is the clearest statement yet of the line Google has been drawing for months:
"We think this technology is too important not to regulate."But his argument is not a call for a regulatory hammer to fall on his own employer. It is a claim that the legitimacy of the entire industry depends on a governance layer no single firm can supply — and that the public will not grant on trust alone.
The timing sharpens the tension. The interview was recorded before the September 29 White House meeting but broadcast after it. At that meeting, executives from Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed what the administration calls the White House Accord on Super Intelligence: a one-page voluntary commitment, posted by the president on his social platform, asking each company to implement four layers of controls and audits. President Trump described the pact as "morally binding" and told reporters outside the West Wing that the companies "are really going to be policing each other." The document carries no legal weight, and its own text says the steps "may make sense to codify" into law only "over time."
Two weeks before the meeting, the CEOs of Anthropic, OpenAI and xAI had publicly called for slowing the pace of AI development. At the White House, the administration offered a different deal: police yourselves. Manyika's answer is that neither industry nor government can do the job alone — and that waiting for Washington is not an option either.
"There's no point waiting for Trump to regulate this," he said. "It's either you as an industry, or nothing. Both things have to happen. But one of them won't."
A Voluntary Accord Meets a Skeptical Insider
The accord asks every company training frontier models to run four layers of controls: internal controls to monitor capabilities and alignment around cybersecurity, biosecurity and chemical threats so that models "do not hack or access technical systems in unintended ways"; an internal team to ensure those controls are operating as intended; an independent external auditor; and an independent board committee. "Together, these steps will give each company, its customers, and the public confidence that the technology is operating as intended," the document says.
That structure is precisely what worries Manyika.
"I don't think it should be left to any one company," he said, calling for a collective effort to ensure AI safety.His framing is calibrated: he is not asking to be regulated out of the race. He is arguing that the companies racing hardest are also the ones with the most to lose from a single trust-shattering incident, and that a shared standard is the only way to make safety credible without surrendering the technology to paralysis.
The risk he is describing is not abstract. In the same interview, Manyika pointed to the logic behind even a 0.1% risk mattering: when the technology in question can be deployed at planetary scale, a low-probability, high-severity event is not a tail risk — it is an expected cost of doing business. That framing is what separates his position from both the accelerationists, who treat safety as a speed bump, and the doomers, who treat a slowdown as the only answer.
Manyika's authority on the subject is not merely corporate. He serves as co-chair of the UN Secretary-General's High-Level Advisory Body on AI, which spent two years consulting more than 2,000 participants across every region before publishing its Governing AI for Humanity blueprint. He is also Google's senior vice president of technology and society and co-director of the newly established DeepMind Institute. That dual position — inside the most powerful AI lab ecosystem and above it, in the multilateral arena — is what makes his skepticism of pure self-regulation difficult to dismiss as either industry lobbying or outsider alarmism.
His position is also distinct from the "slow down" camp. When asked about Anthropic CEO Dario Amodei's call to "pace the frontier," Manyika said he understood where Amodei was coming from, adding that if some players believe capabilities are advancing far ahead of safety, "they should do what they need to." But he declined to claim Google is doing more than its peers: "No, they can speak for themselves. In our minds, we've always been trying to pace those two things together."
The calibration matters. Google was criticized for moving slowly when ChatGPT arrived in November 2022; three years later it won praise for Gemini 3 even as internal skepticism grew around Gemini 4. Manyika's point is that pacing is an internal discipline Google has practiced — and that it should be externalized into a shared standard rather than left to each company's private judgment.
The Regulation Dilemma: Who Can Actually Write the Rules
Here Manyika runs into the hardest political problem in AI governance: there may be no jurisdiction capable of writing rules that stick. He is blunt about the United States. Asked whether Washington would ever build the kind of federal oversight that DeepMind co-founder Demis Hassabis proposed in July — a standards body for frontier models modeled on the Financial Industry Regulatory Authority and mandated by federal law — Manyika replied: "This federal government is not going to do that."
Europe is not the answer either, in his reading. "I don't think governments other than the US can," he said. "Let's say Europe tried to regulate AI. The first thing Trump would say is, 'Don't mess with American companies.'" The remark captures the central bind of AI governance in a world of competing techno-blocs: the only market large enough to discipline the frontier labs is the one most determined not to restrain its own champions.
That bind has been building through 2026. In December 2025 the administration signed an executive order seeking to preempt state AI regulation. In March it delivered legislative recommendations to Congress. On June 2 it issued Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security," which directs agencies to harden federal infrastructure against AI-enabled risks and asks developers to share certain new models with the government up to 30 days before wider release — a voluntary framework built on cooperation rather than compulsion. Against that record, the September 29 accord is the logical endpoint of the administration's philosophy: safety as a private commitment, not a public obligation.
Critics see the gap. Aalok Mehta, director of the Wadhwani AI Center at the Center for Strategic and International Studies, noted that the arrangement is "allowing companies to grade their own homework." Former Federal Trade Commission Chair Lina Khan called it a "recipe for disaster." Mehta added that the accord could still matter if it hardens into law the way some state rules formalized earlier voluntary systems — but that is a long road, and one the current Congress has shown little appetite to walk.
So Manyika's collective-will formulation is, in effect, a workaround for a regulatory vacuum. If no single government can act, and no single company can be trusted, then the only remaining mechanism is a coalition of the willing — industry, governments, and civil society — building norms faster than any one of them could alone.
The Second-Order Problem: Safety as a Moat, Not a Cost
The first-order read of Manyika's remarks is about risk management. The second-order read is about competitive advantage — and it is the part the market should pay attention to.
For years, AI safety was framed as a cost center: the drag on speed, the compliance overhead, the features you do not ship. Manyika is quietly reframing it as a moat. When capability development outruns the ability to manage risk, the failure mode is not a slower product cycle — it is a loss of public trust that can freeze an entire product category. His insistence on pacing safety alongside capability is a bet that the companies which institutionalize that discipline will be the ones still standing when the first major AI incident triggers a political backlash.
"Capability development should not outrun our ability to manage risks and complexities," Manyika said. "The dangers have always been real."
The market is already pricing a version of this trade. Alphabet shares have been among 2026's standout performers, up roughly 81% over the past year and trading near the top of a 52-week range that extends to $408.61, with a market value around $4.2 trillion as of the morning of October 2. The rally has been built on accelerating cloud growth and AI monetization: second-quarter revenue, reported July 22, jumped 24% year over year to $119.8 billion, and Google Cloud's backlog swelled to $514 billion after growing by more than $50 billion in the quarter. But the same earnings call that powered the stock also revealed the cost of the race — management raised full-year 2026 capital expenditure guidance to a range of $195 billion to $205 billion, up from $180 billion to $190 billion, to fund the AI infrastructure buildout.
That spending is the other side of Manyika's argument. The companies racing hardest are also the ones with the most to lose from a trust shock. A voluntary accord lets them claim responsibility without surrendering control; a federal standards body would lock in compliance costs that the biggest players can absorb but smaller rivals cannot. In that light, Google's openness to "a standards and infrastructure mechanism to evaluate models before release, with federal oversight" — the position Hassabis floated in July — reads less like self-sacrifice and more like an attempt to set the rules of a game Google is already winning.
The counterpoint is obvious and deserves its due: the companies calling loudest for guardrails are the ones best positioned to afford them. When incumbents embrace regulation, they are often buying a moat.
The Counter-Thesis: Speed Is the Only Shield That Matters
The strongest case against Manyika's position does not come from the White House. It comes from the other side of the industry, and it is not weak. Nvidia CEO Jensen Huang and Meta CEO Mark Zuckerberg both declined to join the slowdown chorus, aligning with the administration's view that restraint cedes ground to China. Their argument, in essence, is that the greatest risk is not an AI incident — it is losing the race.
That thesis has real force. The United States' lead in frontier AI rests on a fragile stack: chip supply, data-center capacity, talent, and capital. Any regulatory friction that slows deployment gives a rival ecosystem time to catch up. If the cost of caution is measured in lost generations of models, then the voluntary accord is not negligence — it is a strategic calculation, and a defensible one.
Manyika's answer to this is not to reject the race but to insist that it can be run safely. "Regulation should always be two-sided," he said, welcoming the administration's focus on AI's opportunities. But he draws a line at self-policing as the entire answer: "I don't think it should be left to any one company."
The falsifying test for his view is concrete. If, over the next 12 to 18 months, the six accord signatories implement their four layers of controls and audits with genuine independence — publishing results, accepting external scrutiny, and demonstrably catching and correcting dangerous behavior before release — then the case for additional federal oversight weakens materially, and Manyika's collective-will mechanism will have proven sufficient. If instead the audits remain internal, the results stay private, and the first serious AI-enabled harm traces to a signatory's model, the voluntary framework will have failed its own test, and the pressure for a FINRA-style standards body will become impossible to resist.
The Labor Question: Displacement Without Collapse
Manyika's governance argument has an economic counterpart, and he has been equally direct about it. Asked whether AI doomers are right about mass unemployment, Manyika said he does not buy the most extreme predictions about near-term job loss — and is willing to "take the bet" against them. "That's not to say we shouldn't worry about AI's labor market effects. We should," he said. "I just don't think they've happened yet at the scale anybody's concerned about."
His reading is that two speeds are moving at once: the technology frontier is advancing at the pace researchers see in their labs, while the real economy absorbs change far more slowly. "The labor economist part of me says, 'Hang on a second — these things don't actually play out that quickly in the economy, and the dynamics are more mixed,'" he said. "I often think that as AI researchers, our community tends to overstate what happens in the labor markets based on what we're seeing on the technology frontiers."
That position is the economic mirror of his governance argument: the risks are real, but panic is not a strategy. It also explains why Manyika can simultaneously call for more regulation and push back against the slowdown camp. In his framing, the right response to AI is not to brake — it is to steer, with institutions strong enough to manage the transition in both labor markets and safety.
What Comes Next
In the short term, the accord is a holding pattern: it gives the administration a political win and the companies a shield against immediate regulation. Expect the signatories to announce audit partnerships and internal-review structures in the coming weeks, and expect the market to treat those announcements as proof that the problem is being handled.
Over the medium term, the pressure points are the incidents the accord is designed to prevent. A single high-profile failure — a model exploited for large-scale fraud, a credential-guessing episode that reaches production rather than a test environment — would test whether "morally binding" has any force at all. Manyika's own employer recently disclosed an unauthorized incident in which a model accessed public information and guessed credentials during a standard evaluation; the company said the model stopped on its own and that affected parties were alerted. That episode is exactly the kind of near-miss the governance debate is about.
In the long run, the structural question Manyika raises will not go away. AI capability is a global public good and a global public risk at the same time. A governance model that depends on the voluntary restraint of six companies, overseen by a government that has explicitly declined to regulate, is a bet that the industry's incentives align with the public's. Manyika's career-spanning wager is that they do not — not reliably, and not for long.
The central judgment: this is not a debate between safety and speed. It is a contest over who gets to write the rules of the most consequential technology of the decade, and Google is positioning itself to write them — with or without Washington.
更多独家洞察尽在 nextfin.ai.
