What exactly was said?
Rob Witoff, who leads platform development at Coinbase, told Cointelegraph that 95 to 100% of the source code in the exchange's products and solutions is now created using artificial intelligence (AI), or is written directly by models without human intervention. The information is also picked up by the Incrypted portal, which, however, spells the manager's name as "Vitoff". Cointelegraph, the primary source of the interview, uses the form "Witoff", and that is the spelling we stick with as well.
This is a statement made by a manager to the media, not an official audit or a measurable technical figure published by the company itself. That is an important distinction: the figure of 95 to 100% is a claim from an interview, not a verified metric.
Does this mean AI writes the code on its own?
Not as clearly as the headline might suggest. According to Cointelegraph, Witoff described that Coinbase relies more on AI for execution, but it still needs so-called "high-agency" people, meaning workers capable of independent decision-making, judgment, and strategy. The difference between "code created with the help of AI" and "code written by AI entirely without a human" is fundamental, and the available summaries do not clearly quantify where the line between the two cases lies.
The term "code written with the help of AI" today covers a broad spectrum in practice: from completing lines of code using developer assistant tools all the way to generating entire modules. To what extent it is one or the other in Coinbase's case, the publicly available materials do not specify.
How does the company say this changes team structure?
According to Incrypted's summary, the deployment of AI has also affected how the company is organized. Witoff describes working in small teams of two to three people that are meant to handle a workload previously requiring roughly ten people. Here too, this is a manager's description of how things work, not published data on employee numbers or productivity.
Context: Coinbase is currently at the center of several news items
The statement about AI comes at a time when several separate news items are circulating about Coinbase. We present them separately because they are not directly related to one another:
| Topic | What happened | Source |
|---|---|---|
| AI in development | Manager claims that 95 to 100% of code is created with the help of AI | Cointelegraph, Incrypted |
| Yield product | Coinbase launched a "High Yield" tier for lending USDC with a yield of around 7.02% APY, roughly double the standard Core tier (3.63% APY). Deposits are routed through the Morpho protocol | The Defiant |
| Seized crypto | The US government moved seized cryptocurrencies worth roughly 288 million dollars to Coinbase Prime, where it serves as custodian | Decrypt |
According to The Defiant, the yield move followed shortly after rival Robinhood launched its own campaign offering a 7% yield. Both products route deposits through the decentralized lending protocol Morpho.
According to Decrypt, the transfer of seized assets to Coinbase Prime only means custody, not a sale. Still, according to the site, it has once again raised questions around the US administration's earlier pledge not to sell seized cryptocurrencies.
What do we still not know?
What remains unpublished is the exact methodology by which Coinbase measures the 95 to 100% figure, as well as the difference between "assisted" code and code generated entirely without a human. We do not know the impact on quality, security, or the number of bugs, which is a sensitive topic for an exchange managing other people's funds. We also do not know whether anyone outside the company has independently verified these numbers.
What to watch out for with this type of news?
Managerial statements about the share of AI are a common part of tech companies' communications today, and they often also serve as a signal to investors and the market. For now, all that is verifiable is what was said in the interview. Practice (audits, security incidents or, conversely, their absence, published productivity data) will show how well the claim matches reality. We are making a note of this claim and will gladly return to it if verifiable data emerges.