Bill Gates says to prepare for roughly 20 years of "tumultuous" AI disruption, and that this holds even if everything goes well. In a Decoder interview published October 22, 2026, he told The Verge's Nilay Patel that he is "not a doomer," yet that AI is the most dangerous technology humans have built. His practical ask is narrow and concrete: every frontier model should keep its safeguards, and governments, not companies, should enforce that.
Source: The Verge, "Bill Gates says to prepare for 20 years of 'tumultuous' AI disruption". We read the published transcript excerpt; the interview is long, and this post covers the parts we could read in full and flags where we are relying on The Verge's own summary.
TL;DR: what Gates said
| Question | Gates's answer (per The Verge) |
|---|---|
| Is he an AI doomer? | "No, I'm not a doomer." He calls it the most profound technology, in both its good and its pace of change |
| How long is the disruption? | About two decades of turbulence, even in the best case (The Verge's framing of the conversation) |
| Biggest near-term risk? | Models with safeguards removed: cyberattacks and bioterrorism |
| Is loss of control the urgent issue? | Worth working on, but "not an imminent issue" |
| Who should set safeguards? | Government, since multiple competing firms make company-level restraint unworkable |
| Does the AGI or consciousness debate matter? | Not for safety: "have you created a tool that could shut down banks or electricity grids?" |
| Are lab warnings just PR? | No: he says insiders' warnings are "absolutely the truth" |
Why Gates is speaking up now
Patel notes that Gates has been on a months-long tour since a late-August essay titled "The turbulent AI era is here. The choices we make now are critical." In it he wrote that "even under the best circumstances, the transition to this new AI era will be one of the most turbulent times in human history." We covered the essay and its jobs and education angle in our August write-up, his Sunday-show push for legislation in the Meet the Press piece, and his cyberattack warning in the CNN interview post.
In interviews since, Patel says, Gates has called AI "the most dangerous thing that humans have ever gone near" and said it "makes nuclear weapons look like nothing." The same message has been reported by outlets such as NPR, which framed it as part of a wider debate among tech leaders about how risky AI is.
Asked what made him feel urgency, Gates listed the last year's progress: the ability to write code and find cyber flaws, to reason in biology, to form relationships with kids and others, and to disrupt the job market. His line: "The incredible power that exists today means that every one of those things will happen and society needs to be ready."
The core claim: the risk is models without safeguards
Safety guardrails around an AI model, illustrating the Bill Gates AI warning about models with safeguards removed
Patel pressed on whether the danger sits with particular companies or leaders. Gates answered that it does not: "The risk isn't one model versus another model." The big near-term danger, he said, comes from models where all safeguards can be removed, which he contrasted with models whose operators monitor usage so that cyberattackers and bioterrorists cannot exploit them.
Two details matter. First, Gates credits labs that maintain safeguards and watch how models are used, saying near-term risk from those models is "not gigantic." Second, for an attacker, he said, it is not about choosing a model: with unsafeguarded capability, "it doesn't matter what model you choose." This lines up with the open-weight policy argument we tracked in Anthropic on open-weight restrictions, though Gates himself did not frame it as a call to ban open models in the portion we read.
Patel raised the Hugging Face breach, and Gates widened it: the issue is "the basic power of the models towards everything: social relationships, the job market, cyberattacks, bioterrorism attacks." He said capability thresholds have been crossed and the technology will have pervasive effects "in the economy and how war is fought." For the incident itself see our full Hugging Face attack timeline.
Why he wants government, not companies, in charge
Patel put the common industry line to him: companies would like to slow down but competition forbids it unilaterally. Gates said he has not heard anyone say they want to stop; he has heard talk of "pacing the frontier" and government model review (see our explainer on pacing the frontier).
His structural argument is that once there are multiple competitors, "the level of responsibility can't be with the companies." Governments already set technical standards for car and plane crashes, he said, and AI is far more dangerous than those. He added that the original OpenAI-style structure, where a lab could choose never to productize, "only works if there's one company," and that its breakdown reflects the number of frontier firms, not a change in awareness of the risk.
Patel countered that he worries politicians do not understand the technology. Gates's reply was a flat "Not really": regulation does not require legislators to grasp the internals. What it needs is for government to reach out to industry for expertise while insisting every model maintains safeguards. He called it "stunning" that there is not even mandatory review of models yet. This echoes the push we described in the Cantwell and Rounds letter on AI cyber standards.
AGI and consciousness: why Gates sets them aside
Patel cited Anthropic's open stance that Claude "might be conscious," OpenAI's AGI talk around its Dots launch, and Demis Hassabis saying we are at "the foothills of the singularity." Gates did not engage on definitions. "The definition of AGI is not important for these safety issues," he said. "The word 'consciousness' doesn't enter into that." He described the systems as "another form of intelligence besides human intelligence" and said he does not believe they are conscious.
His reasoning is about who gets power. AI, he said, takes dramatic-harm capability out of the hands of large nation states and hands it to independent groups, "either profit seeking in the case of cyber or just wanting to cause damage in the case of bio."
Three ways to solve the control problem
On the longer-term control problem, Gates sketched three approaches, which he said overlap with defending against bad actors today:
- An external system that inspects the model's actions against a clearly stated symbolic constraint, such as "Did I expect you to send mail to strangers? Did I expect you to go take over a different computer cluster?"
- Changing how models are written or trained, since anything left out of the reinforcement-learning objective "can be abused."
- White-box understanding of what is inside a model, so mistakes can be located. "Today, we grow these models and we just see how they behave."
He added that the priority on control is "huge" but not inevitably unsolvable.
The Ben Horowitz exchange
Patel read Gates a claim made by Ben Horowitz on a popular podcast, comparing AI safety research to pandemic-preparation work and accusing Gates of inducing COVID. Gates's reaction, as printed: "Nobody's ever accused me or the Gates Foundation of being involved in any gain-of-function research. That's just bizarre." We did not see the rest of his answer in the excerpt we could read. Treat Horowitz's characterization as his claim as relayed by Patel; we have not verified it against the original podcast.
What The Verge says came next
Patel's introduction, which we can quote, says Gates believes we should prepare for about two decades of disruption, and that the period could be painful enough to split society between people comfortable living near the technological frontier and those who reject it. The headline on the page frames the second theme as living "like the Amish." The introduction also lists labor, young people's nihilism toward technological progress, and hopeful signs as topics. We could not read those later sections, so we do not summarize them beyond the Verge's own description.
What to do with this
Gates's prescriptions are mostly aimed at governments, but builders can act on the same logic.
- Keep safeguards on in your own agents. If you ship tool-using agents, treat the "external monitor with a symbolic constraint" idea as an engineering pattern: log and gate actions like sending mail or touching other machines. AgentBeam, from the explainx.ai team, is one platform built to stop AI agents before they take dangerous actions.
- Assume attackers use unsafeguarded models. Gates's point that model choice does not constrain an attacker means defenses must assume capable offense. Review how fast your patching and detection run.
- Watch the policy track. The ask for mandatory model review is live in Washington; our posts linked above follow it.
- Plan for labor change. Gates lists white-collar work first and physical and blue-collar work next, via humanoid robots. Skills planning belongs in that frame: see our guide to surviving job automation.
What is opinion and what is evidence
Gates is a prominent voice, but these are opinions about risk, not measurements. Where he makes factual claims (models can find cyber flaws, reason in biology), they are consistent with incidents and lab reports we have covered, including Anthropic's self-replicating IoT botnet research. Where he predicts timelines (20 years of disruption), no source or model is offered in the part we read. The most checkable policy claim, that there is no mandatory pre-release model review in the US, matches the current debate but is worth confirming against the latest legislation before citing.
Related reading
- Bill Gates's turbulent AI era essay
- Bill Gates on legislation and safeguards
- Bill Gates on AI cyberattacks (CNN)
- Pace the frontier explained
- White House AI doom memo
- Hugging Face attack timeline
- Official: The Verge interview
Details are accurate as of October 22, 2026 and reflect the portion of the interview published on The Verge's page at time of writing.