Airbnb says AI writes 60% of its new code. Nobody has explained what that means.

Brian Chesky told Wall Street on May 7 that nearly 60% of the code Airbnb’s engineers produced in Q1 was written by AI. “This is significantly higher than our peer set and benchmarks,” he added on the earnings call.

He didn’t say how Airbnb counts it. Lines of code? Commits? Characters? Pull requests that touched an AI suggestion? The methodology was not published and nobody on the call asked. That makes the number useful as a signal of corporate enthusiasm and not much else.

The claim arrives in a season of escalating AI-coding announcements from every major tech CEO. Each number is bigger than the last. None come with a reproducible definition. And the single largest independent study of AI-generated code in production found numbers that look nothing like what executives are telling Wall Street.

The one-upmanship problem

Airbnb is now third in what’s become a CEO leaderboard for AI coding claims. Sundar Pichai said 75% at Google Cloud Next in April. Snap claims 65%. Meta has an internal target of 75%+ of committed code. Each number arrives on an earnings call or a keynote stage, where the incentive is to impress investors, not to publish reproducible methodology.

The largest independent measurement tells a different story. A study covering 4.2 million developers between November 2025 and February 2026 found AI-authored production code at roughly 27%. The gap between that number and the CEO claims isn’t necessarily dishonesty. It’s probably definitional. If you count every line where an AI made the initial suggestion, even if a developer rewrote half of it, you get higher numbers than if you count only code that shipped as-is.

One analysis of Google’s 75% claim noted a conspicuous absence: “One number is conspicuously missing: the rejection rate. How much AI-generated code gets rewritten before it ships?” GitHub Copilot’s public acceptance rate hovers around 30%. That means 70% of suggestions get thrown away. Whether the discarded suggestions count toward the headline number matters a lot, and none of these companies clarify it.

What Chesky actually said about engineering

The more interesting quotes from the call weren’t the 60% number. They were the organizational claims.

“AI gives you huge productivity gains,” Chesky said. “Where you might have needed a team of 20 engineers before, an engineer can now spin up agents to do a lot of work under supervision.” Claude Code is the only coding tool he named on the record. Airbnb’s customer support bot, which resolves over 40% of issues without human escalation (up from 33% last quarter), runs on Alibaba’s Qwen model, a separate system that’s attracted a Congressional inquiry.

Chesky also warned that “pure people managers” are an endangered category. “I do not think there is going to be as much of a role for pure people managers,” he said on the Invest Like the Best podcast. He cited Jony Ive at Apple as the model: a leader who stayed hands-on with the craft.

But Airbnb is not cutting headcount. The company grew by roughly 900 employees year-over-year to about 8,200. Q1 revenue hit $2.7 billion (up 18%), net income was $160 million, and 156.2 million nights were booked. Chesky’s framing is a productivity multiplier, not replacement.

The trust gap

Developer sentiment on AI coding tools has split in a way that these CEO numbers don’t capture. Adoption is at 84%, but positive sentiment has dropped to 60% from over 70% two years ago. Only 29% trust AI code accuracy. Just 3% report high trust. And 66% say the recurring frustration is code that’s “almost right, but not quite.”

That’s the core tension. Everyone uses the tools. Barely anyone trusts them.

Gergely Orosz, writing in The Pragmatic Engineer, described what he calls “the grief when AI writes most of the code”: the sense among experienced engineers that the skill they spent years building is being devalued by tools that produce plausible but often subtly wrong output. “It took a lot of effort to get good at coding,” he wrote. “Once you’re pretty good, you have something that’s valuable.” The open question is whether the value migrates to a higher level of abstraction or just erodes.

AI-generated code also produces 1.7x more issues than human-written code, with higher rates of critical defects and security vulnerabilities. Microsoft’s own DELEGATE-52 benchmark found that even frontier models corrupt 25% of document content over long editing workflows. The gap between “AI writes 60% of the code” and “60% of the code is good enough to ship” is the part nobody on an earnings call has time to explain.

The Qwen controversy adds another wrinkle. Airbnb uses Alibaba’s Qwen model for its customer support chatbot, not for coding. But House committees launched a joint investigation in April into Airbnb and Anysphere (Cursor’s parent) over use of Chinese AI models. Rep. John Moolenaar said the models “are trained by China’s censorship regime and introduce hidden vulnerabilities.” Chesky has described Qwen as “fast and cheap.” Whether political pressure forces a switch to a Western model is an open question that could affect the cost math behind the 60% number.

The financial backdrop

Airbnb’s Q1 was fine. Revenue hit $2.7 billion (up 18%), net income was $160 million, and 156.2 million nights were booked. The company grew headcount by about 900 employees year-over-year to roughly 8,200. No layoffs were announced. CFO Ellie Mertz noted the company is “ramping up our use of AI internally and anticipate that expense will ramp over the course of the year.”

For context: Disney’s leaked internal dashboard showed 4,800 product and tech staff burning 3.1 billion Claude tokens in nine workdays. Shopify CEO Tobi Lutke told employees last year they must prove AI can’t do the job before requesting new headcount. Cloudflare cut 1,100 jobs on May 7, explicitly saying AI made the roles obsolete.

The cultural shift is real. But Chesky is the only CEO in this group who’s simultaneously growing headcount and claiming 60%. That’s either confidence or contradiction. The next few quarters will show which. If Airbnb’s engineering team shrinks while the 60% number climbs, the “productivity multiplier” framing was always a precursor to cuts. If headcount stays flat or grows, it’s a genuine bet on humans and AI working together at scale.

What this means for you

If you manage engineers, the pressure to adopt is real and backed by every major CEO on the planet. The 60% number will show up in your next board deck or quarterly review whether you believe the methodology or not.

If you write code, the honest picture is that AI tools speed up the repetitive parts and produce variable-quality output on the hard parts. A Copilot acceptance rate of 30% means the tool is still wrong more often than it’s right. The companies claiming 60-75% are measuring something different from what that sentence implies.

Chesky himself acknowledged limits: “I do not think anyone has figured out AI for travel or e-commerce yet.” He listed four specific chatbot shortcomings for travel: text-heavy interfaces, inability to compare options, lack of interactive controls, and poor multi-user trip planning. If the CEO making the biggest AI coding claims also says the product application isn’t solved, that’s worth holding in the same frame.