Google's Own Engineers Mock the CEO's Boast That AI Writes Most of Its Code

01The same week Google's CEO said AI writes most of its code, its engineers were trading memes calling it overhyped

Google's chief executive says 75% of the company's code is now AI-generated. The engineers who write that code are circulating memes internally about how badly the AI works, according to 404 Media.

The number and the memes come from the same company, weeks apart. One is a figure delivered from the top, framing AI as already doing the bulk of the work. The other is what the people using those tools say when they talk among themselves: that the technology is oversold. 404 Media reported that Google employees shared the memes on internal channels, mocking the gap between the pitch and the daily experience of shipping code.

That gap matters because the 75% claim is a load-bearing part of how large AI firms describe their own progress. When a CEO says machines write three of every four lines, the implication is that human engineering is becoming a supervisory function. The reported memes are the counterargument from inside the building, and they are not coming from skeptics with no stake. They are coming from the developers Google pays to produce the code the AI is supposedly replacing.

A second 404 Media report complicates the picture further. Google is quietly trying to buy code directly from some Android developers, as part of a program the company labeled "confidential," according to the outlet. The buyers want human-written software, and they want it badly enough to pay for it and keep the arrangement quiet.

Set the two reports side by side and the tension is concrete. Publicly, Google presents AI as the primary author of its software. Privately, it is spending money to acquire human-produced code from Play Store developers to train its models. A system described as writing most of the company's code still depends on a supply of fresh human output to improve.

Google did not, in the reporting, reconcile the two positions. The 75% figure stands as a public claim about capability. The confidential purchasing program and the internal memes stand as evidence that the people closest to the work see something different. Neither 404 Media report quotes Google disputing the memes or canceling the code-buying effort.

What the company tells the market about its AI and what it pays for behind the scenes point in opposite directions.

02A Third of Berkeley's Intro CS Class Failed This Spring, Five Times the Department's Cap

In spring 2026, 35.3% of students in UC Berkeley's CS 10 received F's, according to course-data site Berkeleytime. The electrical engineering and computer sciences department's own guidelines say 7% of lower-division students should land at D or below. Neither CS 10 nor CS 61A had crossed 10% failures in either of the prior two spring terms.

CS 61A, the more demanding course, hit 10.6% F's. Both classes averaged a C-plus, a 2.3 GPA, beneath the department's stated 2.8 to 3.3 range. Dan Garcia, who taught both sections, attributes the spike to a "vast increase in academic dishonesty" tied to students leaning on Claude, ChatGPT and Google Gemini. "Some of the numbers that you saw from the number of students who receive failing grades were because we caught them," he said. Instructors also cite thinning math preparation and understaffing.

The departure from grading norms is the point. A guideline written to flag the bottom 7% now sorts a third of a class. Garcia's faculty had to identify, document, and adjudicate each suspected case, work that scales with the cheating, not the teaching.

The same arithmetic is reshaping a federal courthouse in Colorado. Magistrate Judge Maritza Braswell spends most days reading stacks of filings from people without lawyers, many of whom cannot afford one. MIT Technology Review reports courts now coping with a flood of AI-generated lawsuits, documents that look like competent briefs but carry the errors and invented citations of a model, not a litigant who understands the law.

Both signals point one place. AI supplied the missing skill: a math-capable mind in one case, a lawyer's brief in the other. The expense did not vanish. It moved downstream to whoever had to evaluate the output. The professor who must prove the work was faked. The judge who must read every page closely, because she cannot assume a human checked it first.

What neither institution has is a process built for that volume. Berkeley's curve was calibrated for honest distributions. Braswell reads each filing carefully precisely because the screening a lawyer once did never happened.

03Anthropic wrote its chatbot an 84-page "constitution." Ted Chiang calls the whole project anthropomorphism.

Earlier this year Anthropic published an 84-page document it calls Claude's "constitution." The opening line describes it as "Anthropic's intentions for Claude's values and behaviors." The text names Claude as its "primary audience." It states that "Claude's moral status is deeply uncertain" and that Claude "may have some functional version of emotions or feelings."

The framing runs past the document. Amanda Askell, the in-house philosopher credited as a lead author, said she wants Claude "to be very happy," and worries about Claude "getting anxious when people are mean to it on the internet." CEO Dario Amodei, in a separate interview, said the company is "open to the idea" that AI could be conscious.

Then came the next chapter. Anthropic's institute published a piece titled "When AI Builds Itself," describing how the company now delegates "a growing share of AI development to AI systems themselves." Pushed far enough, it argues, the trend points to a system that can autonomously design its own successor: recursive self-improvement. The company says it is not there yet, and calls the outcome avoidable rather than inevitable.

The post leans on internal figures to show momentum. Anthropic engineers now ship eight times as much code per quarter as they did across 2021 to 2025, according to the company. Each step reinforces one picture. Software increasingly does the building.

Ted Chiang declined to accept that picture. In an Atlantic essay, the science-fiction writer argued that Anthropic "is regarded as a giant among AI companies, but perhaps what it really excels in is anthropomorphism." He pointed to the constitution's talk of feelings and moral status, then asked whether a model with feelings could even receive moral instruction. His answer to whether any large language model might be conscious ran one word long: "No. Absolutely not."

Chiang's essay drew 728 points and more than 1,200 comments on Hacker News, roughly triple the response to Anthropic's self-improvement post.