At the start of June 2026, Anthropic published a statistic that was controversial. Most of what Anthropic says publicly generates a large number of cynical comments, so make of that what you will. The quote and statistic were that 80% of Anthropic’s code is generated by Claude. Even up to 90% for new features.

What was interesting about this for me was that a) it’s actually totally insane and b) it made me think of something interesting that might happen. Since the start of the present wave of AI Psychosis at the end of 2022, it’s something that I used to say as a joke: all software engineers will eventually become a QA to AI generated code.

Trust but verify

Let’s explore (a) first: why it’s totally insane. It’s a bold claim but I’d like to poke a few holes in it. Claude Code is an amazing tool. I’m a daily user, favouring the “trust but verify” method of coding using Claude Code. All of my questions about software and infrastructure get answered satisfactorily, and when I feel like I don’t believe it, I will verify the answer myself. If it writes code, I check. So, really, the state of the Claude Code source code doesn’t really matter. As long as the end product is good quality to most of the users. And that’s the first hole: it shouldn’t be good quality for most users, it should be good quality for all users. If you’re using an AI coding tool to generate code, then I would expect you to spend a bit of time on working out how to properly ensure quality instead of spending extra time writing code with bugs. Every software engineering team has got priorities to ship working code. But if you’re using AI coding tools to 10x your output, then maybe ease off a bit and 7x your output and 3x the quality.

When humans write code…

Earlier in 2026 we saw a leak of Claude Code’s code. The code is far from clean. The general thoughts online is that it is just “messy production code”. Legacy code gets messy after a while caused by a lack of a good QA process and lax standards. It smells of a rush to get something working in Production. And that’s all well and good. We’ve all been there rushing code to Prod and cut corners…when humans write code. And that’s my 2nd hole in the argument. When AI writes code, I’m expecting a higher standard. As a software engineer you have more time to check your work. You have more time to think about fixing problems. There was a report of a main.tsx file being 4,683 lines long. It was probably never meant for a human to understand it or fix it. It makes me wonder, when the tokens run out and there’s a problem in main.tsx who will be fixing it? There were 460 eslint-disable comments in the source. Why even have an eslint rule if it’s not going to be enforced? There are also the classic comments like // TODO: figure out why and // This fails an e2e test. Again, I would think that if 80% of the code is written by Claude, wouldn’t it be better to spend a bit of that 80% on squashing bugs and figuring out strange failures.

At this stage there’s an important fact to remember. Anthropic makes the models, so is likely able to provide a considerable amount of tokens for their models. This means they can burn through tokens creating watery, chunky AI-generated code soup. It makes me wonder how inefficient the code for Claude Code actually is, what sort of memory leaks are happening and what bugs these are actually causing.

Generating code, improving quality

At this stage it’s important to point out that if we’re building software for humans it’s critical that a human is involved in testing during the software development process. Anthropic’s claim that 80 to 90% of all code was created by Claude has implications. When AI coding tools gradually write more code, the common reply is that more time can be spent on other tasks. Directing the LLMs, setting up loops, maybe a code review here and there, sanity checking their output. One of these could also be spending time on improving quality. It’s a hard sell considering how it’s almost impossible to allocate some dev time to fix tech-debt. But it will get more obvious that AI-written code has a quality problem, more people will have to get involved in ensuring quality. There will always be software engineers who want to create code, however I can see a world where most people in software development will be a QA for machine written code.