Coinbase CEO Brian Armstrong rejects calls for new AI regulator as crypto firms increase AI use
Coinbase CEO Brian Armstrong says the US does not need a separate artificial intelligence regulator: current laws, in his view, already cover the harm AI systems might cause. His argument arrives alongside a startling Coinbase figure. AI now writes or assists with 95% to 100% of its code, up from an estimated 40% in February. That happened in five months. I’ll be honest: the headline number is impressive, but also frustratingly vague. Coinbase has not explained where it draws the line on AI involvement or how much engineers change before code reaches production.
The regulatory debate
The dispute is really about pre-release testing. On July 14, Google DeepMind CEO Demis Hassabis proposed a federally supervised standards body to test and certify models for cybersecurity and biological risks. National security would also fall within its scope. He suggested a public-private organization based on the Financial Industry Regulatory Authority (FINRA), starting voluntarily and potentially becoming mandatory later. OpenAI CEO Sam Altman supported much of the idea. Microsoft CEO Satya Nadella did too, saying the goal should be to “avoid any model that breaks the world.” Dramatic wording? Yes. But the concern underneath it is concrete.
Armstrong’s counterargument
Armstrong thinks the proposal risks making companies seek approval twice: first from state authorities, then from an industry organization. He wants neither a self-regulatory group nor a government agency devoted solely to AI. Courts can already compensate people harmed by AI products, he argues, and consumer protection laws already apply. My take: the duplication concern is not theoretical for crypto companies. They have spent years dealing with overlapping regulators, and Armstrong does not want AI developers entering the same maze.
“Why design regulation around a hypothetical problem,” the crypto chief said. “The existing laws which prevent fraud, award damages when victims are harmed (tort), UDAP Laws (Unfair and Deceptive Acts and Practices) etc provide broad protections if one of the frontier labs issues a model that does harm.”
Armstrong also argues that AI companies have a financial incentive to make safe products because customers will leave dangerous tools. Most free-market arguments stop there. That’s only half right. The real issue is timing: users must recognize the danger before serious damage occurs. Can they always do that? Coinbase’s notification error, discussed below, suggests they cannot. The system can fail first and look convincing while doing it.
Coinbase’s AI use
Coinbase is moving fast. Rob Witoff, the company’s head of platform, recently told Cointelegraph that large language models now write or assist with 95% to 100% of the exchange’s code. Coinbase put the share at 40% in February. Five months, roughly 55 to 60 percentage points. That is not a marginal shift. Still, “AI-generated” can mean anything from completing a line to drafting most of a feature, so I would not treat the percentage as a clean measure of autonomous coding.
The company announced a 14% workforce cut in May. It also reorganized into smaller teams that rely more heavily on AI. Engineers continue to review sensitive work such as cryptography; AI handles nearly all internal prototypes. That split makes sense. Prototypes can be thrown away. Broken cryptographic code can put customer funds at risk.
If the arrangement works, Coinbase could spend less and release software sooner. Other companies may copy it. Counter to the usual market narrative, though, faster development is not automatically bullish for Coinbase shares, which trade under COIN. It could improve margins. Poor code and rushed reviews could also become expensive. One flashy percentage is not a ready-made buy or sell signal. Not even close.
AI adoption across the industry
Coinbase is not alone. Gemini and Crypto.com have reduced staff this year while using more AI. Kraken, Messari and Dune have made similar moves. In some departments, the result is smaller teams and more automation, with fewer jobs left behind. Software updates may arrive sooner. Do they become more accurate? Not necessarily.
Coinbase got an uncomfortable demonstration earlier this month: an AI-written notification announced the result of a FIFA World Cup match before the game had started. The company investigated. It was one notification, not a system-wide failure, and that distinction matters. Even so, the incident showed the full failure chain—AI invents an answer, presents it confidently, then publishes it at speed. In finance, a comparable mistake could move prices or trigger trades before anyone catches it. Human review remains necessary. I would not soften that conclusion.
What this means
Armstrong opposes broad AI regulation when officials have not identified a specific gap in existing law. That position fits the crypto industry’s long resistance to rules it sees as vague or repetitive. Meanwhile, Coinbase is betting on smaller teams equipped with AI to produce more software. The upside is straightforward: earlier releases and lower spending. The downside is equally plain. More mistakes like the premature World Cup result could reach users.
Traders should not assume that savings will produce lower fees or faster service. Nor do they guarantee a higher COIN price. Coinbase could keep the money, direct it elsewhere, or spend it resolving compliance and quality problems. Yes, that complicates the efficiency story. It should. The World Cup error was small, but it showed that automation creates costs of its own.
Claims that Coinbase’s AI push will increase blockchain adoption are much harder to defend. AI models consume large amounts of computing power and data. Neither requirement makes decentralized infrastructure necessary. Some crypto companies will probably combine AI with blockchain, but a pairing is not a use case. My take: there is still no guarantee that the combination becomes common, let alone useful.
Investors should focus on the actual regulatory proposals, especially any plan requiring approval from both government agencies and an industry standards group. Two reviews could raise costs and delay releases at Coinbase and other crypto companies. Is that automatically disastrous? No. The impact would depend on which models the rules cover and how regulators enforce them. The powers given to the standards body would matter too.
Comments from technology and crypto executives can move sentiment temporarily. Proposed laws and agency guidance matter more. For COIN, skip the grand predictions and ask concrete questions. How much money does AI save? How much faster can Coinbase develop software? How often do errors reach customers? Regulation could change every answer. So could the quality of Coinbase’s engineering reviews.
FAQ
Q: What is Brian Armstrong’s main argument against a new AI regulatory body?
A: Armstrong says current fraud, tort and consumer protection laws already cover harm caused by AI. He also warns that another regulator could make companies pass through two approval systems.
Q: How much of Coinbase’s code is now AI-generated?
A: Rob Witoff, Coinbase’s head of platform, said large language models write or assist with 95% to 100% of the company’s code.
Q: Who proposed a federally supervised AI standards body?
A: Google DeepMind CEO Demis Hassabis proposed a body that would test and certify advanced AI models.
Q: What does Armstrong mean by a “dual approval process”?
A: He means companies might need approval from state regulators and a separate industry standards organization before releasing AI systems.
Q: Has Coinbase had problems with AI-generated content?
A: Yes. Coinbase investigated an AI-written notification that announced a FIFA World Cup result before the match started.