AI Makes All Code 'Write-Only': What That Means for Developers and the Rest of Us
A software engineer told a London conference that AI-generated code is too dense to read, too fast to review, and best treated as disposable. His prescription: automate the boring parts and save human creativity for what machines still cannot do.
Key points
- Engineer Phillip Mortimer argued at QCon London 2025 that AI turns all code into "write-only" material: produced faster than any human can read or check it.
- Mortimer says tests, not the code itself, now serve as the real documentation for what software does.
- He recommends using AI agents to review code automatically and to fix live problems by monitoring a system's alerts.
- According to Mortimer, intent and implementation are now fully separated: developers no longer need to master a specific programming language.
- Research cited by Mortimer suggests solo thinking produces better ideas than group brainstorming sessions.
"Write-only code" is a decades-old programmer's insult for software so tangled that even its author struggles to read it back. At QCon London, a developer conference held earlier this year, engineer Phillip Mortimer argued that AI has made every codebase write-only by default.
That sounds alarming. It is worth understanding what it actually means.
So what is the problem, exactly?
AI coding tools can produce hundreds of lines of working software in seconds, far quicker than any human team can read and check it. Mortimer's point is that volume alone makes human review impossible. The code works, but nobody fully understands it.
His solution is not to slow down. It is to stop pretending humans can keep up with line-by-line checks. Instead, he says, you lean on two new rules.
First: the tests define the behaviour. A test, in software terms, is a small automated check that feeds the program a known input and confirms it produces the right output. When the code itself is too dense to follow, those input-output pairs become the only readable record of what the software is supposed to do.
Second: code is disposable. If something breaks and the AI-written code is hard to unpick, the faster move is often to delete it and let the AI rewrite it from scratch. You keep the tests, which tell the new version what to achieve. You throw away the implementation.
Can AI review its own code?
Yes, and Mortimer says it works better than you might expect. An AI agent, software that carries out multi-step tasks on its own, can be plugged into the same pipeline developers already use to check each other's work. Give it a different prompt, a different set of instructions, and it catches real bugs even in code another AI just wrote.
He also described "de-slopify" agents: tools that run quietly in the background, scanning a codebase for repeated code, redundant settings, or unused features, and cleaning them up without human prompting.
Beyond that, Mortimer envisions software that heals itself. An AI agent watches the live system's alerts (the constant stream of warnings a running application throws off), spots patterns, and automatically raises a proposed fix for the engineering team to approve.
What does this mean for developers' jobs?
Mortimer's answer is: the job shifts toward creativity. Implementation, actually writing the lines of code, is something AI handles. What remains is knowing what to build and why.
He noted that companies should give developers uninterrupted stretches of time to think. Brainstorming sessions, the kind most workplaces love, have actually been shown to produce fewer good ideas than people working alone. Constraints help too: a tight deadline or a limited budget often forces more creative thinking than a blank canvas does.
For people who are not developers, the practical takeaway is simpler. The software that runs your bank, your hospital records, or your shopping app is increasingly written and patched by AI. The humans overseeing it are shifting from writing every line to setting goals and checking outcomes. That is a different kind of oversight, and the industry is still working out how to do it well.
Common questions
Does write-only AI code mean software is less safe?
Not automatically, but it does change where the risk sits. Fewer human eyes on individual lines of code means automated testing and AI review tools carry more of the safety burden than before.
Do developers need to learn new programming languages now?
Mortimer argues the opposite: because AI translates plain instructions into any language, developers can work across many languages without mastering each one's syntax from scratch.
