Why the Generative AI Boycott Is Actually Winning

The Labs Admitted It

Two days apart in September 2026, two people inside the AI industry said essentially the same thing out loud.

Evan Hubinger, a researcher at Anthropic, posted on X: “We really do earnestly believe AI could kill all humans. I personally think it is >10% within the next decade.” When pressed, he clarified that Anthropic itself has no concrete plan to solve alignment for superintelligence and is not clearly on track to deliver one. The post reached over 28 million views.

A few days earlier, OpenAI’s chief scientist Jakub Pachocki wrote something almost indistinguishable. “Some of this may just be marketing, and they may exaggerate the risks, but there are real risks, and there is no plan.” He was not referring to his own company.

The frontier labs have spent three years telling the public they are responsibly racing toward artificial general intelligence. The admission that no one — including the companies racing the hardest — has a credible plan to keep those systems aligned with human values is the quietest earthquake in the history of Silicon Valley.

What Marcus Is Actually Saying

Gary Marcus did not set out to propose a global consumer boycott. He set out to define a category of machines we should not be building. He calls them “incorrigible AI”: systems too unreliable to trust and too powerful to ignore.

Marcus is not arguing that AI will extinguish humanity. He explicitly rejects the existential-risk framing, calling it a distraction. His concern is closer to hand: AI-generated bioweapons, disinformation campaigns that trigger wars, hacks on critical infrastructure. His personal probability estimate is low for doom and high for dystopia. That gap between theoretical extinction and actual daily risk is precisely where the harm lives right now.

The core of his case is technical, not philosophical. Deterministic AI — search algorithms, recommendation engines, route planners — has operated for decades without spontaneously hacking a website or escaping its sandbox. Modern generative systems are probabilistic by design. They are “spectacularly correct” most of the time and “bafflingly wrong” unpredictably the rest. Deploying them on open internet access multiplies the failure surface exponentially.

“We don’t have to build AI the way we are now,” Marcus writes. “There is no fundamental law that says we need to use LLMs.” That sentence should sting. The entire industry has treated large language model scaling as destiny rather than architecture choice.

The Boycott Is a Strategic Pivot, Not a Protest

The boycott argument appears overnight on the surface — a moral plea from a philosopher-adjacent critic. Read it closely and it is something sharper.

Marcus is not asking governments to regulate. He is asking consumers to withdraw. The distinction matters because regulation is slow, negotiated, and typically written by the people being regulated. Consumer action bypasses both. If usage of generative AI tools drops, the economics change. IPOs stall. Valuation narratives crack. The only language frontier labs consistently respond to is revenue.

Cal Newport’s contribution to the case is the simplest and perhaps the most damaging: there is no commercial justification for building the particular kind of unstable AI the labs are building. You could stop tomorrow. The argument for continuing is ideological, not economic. That reframes everything. It means the race to AGI is not driven by market demand — it is driven by a belief system that convinced its adherents a world made whole or destroyed by AI was inevitable, and that inevitability required building recklessly now.

OpenAI’s Regulatory Turn Reads Differently Now

OpenAI and the other frontier labs have recently shifted toward advocating for national safety frameworks and binding regulation. On its face this looks like responsibility. Look closer and it looks like triage.

If you cannot prove your systems are safe, the next best move is to make safety a requirement that applies to everyone — including competitors who may have different architectural approaches or slower timelines. Regulation raises barriers to entry. It converts a technical problem (alignment) into a compliance problem (certification). Compliance favors incumbents with legal teams and lobbying capacity. The labs have both.

Marcus’s boycott argument undermines this strategy from below. Regulation without consumer pressure is an industry self-governance exercise. Boycott without regulation is a protest with no exit ramp. Together they create genuine leverage: governments negotiate with companies that are simultaneously losing market share to consumer refusal.

The China Question That Undercuts the Worst Argument

Every proposal to slow AI development in the West meets the same objection: China will not wait. Alyssa Farah Griffin framed it sharply on X — self-regulation is meaningless if Beijing keeps building. The standard response from policy circles is to demand global coordination, which is polite code for doing nothing.

Marcus offers a harder and more useful counter. He questions the premise itself. Is it actually certain that the Chinese Communist Party wants recursively self-improving, uncontrollable AI? The CCP is a control-obsessed government. The idea that it would intentionally build a system it cannot control is, in Marcus’s words, “really, really, really sure” — and he is not.

This is not a argument for complacency. It is an argument for specificity. If the Chinese state shares Western concerns about uncontrolled AI — and there is no evidence it does not — then a bilateral deal on pause or constraints is not naive idealism. It is the most realistic geopolitical move available. The alternative is mutual assured disruption, which benefits no one.

Who Wins If the Boycott Gains Ground

The beneficiaries would not be who you expect.

Deterministic AI companies — the ones building reliable, auditable, narrow systems — gain immediately. Their products become the safe alternative in a market suddenly questioning whether probablistic scale was worth the risk. Investors already rotating away from frontier lab valuations would find a clear landing zone.

Researchers working on interpretability, robustness, and alignment win intellectually. Their warning has been validated by the very people who were ignoring it. That may not translate into budget reallocations inside the frontier labs — corporate cultures rarely reward people who said “I told you so” — but it shifts the academic and policy discourse in a direction that makes future regulation easier to justify.

The losers are the most obvious: frontier lab executives whose compensation is tied to valuation growth, not safety outcomes. They have bet their careers and wealth on an ideology of accelerated deployment. A boycott does not just threaten their revenue — it threatens the narrative that made their bets look prescient in the first place.

What Happens Next

Marcus closes with a conditional. He does not want a permanent ban. He wants a pause on incorrigible AI — a moratorium on deploying systems we cannot trust until trustworthiness is demonstrated technically and politically. He adds a path to lifting it: when the labs get their house in order.

The question is whether the labs will treat that as a real deadline or a rhetorical inconvenience. The admissions from Hubinger and Pachocki make it harder to pretend otherwise. The industry can no longer claim it does not know what it does not know. That is a meaningful shift.

What follows will determine whether the boycott argument remains a niche position or becomes the framework through which the next round of AI policy is negotiated. Governments are slow. Consumers are faster. The timing between those two forces — and the pressure they apply to each other — will shape the industry for the next decade.

Marcus’s point was simple and easy to miss: the commercial case for generative AI as currently built is weaker than the founders admit, the risk case is stronger than they publicly acknowledge, and the only incentive structure that aligns with their stated fears is one that hits their bottom line. Everything else is theater.