Dismissing AI code as slop misses the real risk
· Matt Horn
If you review code for a living, you’ve read the articles about AI slop taking over codebases; Bad code is kudzu is one of many. They say agents write code faster than senior engineers can review it, so more of it ships unread, and the code is the worst it’s ever been.
I’m not here to argue the code is good. Plenty of it is bad, and plenty of production code has always been bad. The difference now is that people talk about it, and AI gets the blame. Rarely in my career have I seen anyone discuss code quality above the team level as long as the system shipped and stayed up. Bring up unit test coverage or cyclomatic complexity and it rarely moved the needle. Today some of the same voices call bad code a crisis, and blame it on AI and on the junior engineers who use it.
Yes, that’s a simplification. Still, read Hacker News from 2023 and you’ll find the same complaints, minus the word “agent.” The target moves every few years, and right now it’s “slop.”
Meanwhile, researchers at Anthropic, DeepMind, and OpenAI are quitting to warn about catastrophic risk. Jacob Coxon resigned on September 8 because labs are “racing straight to self-improving superintelligence and gambling with our lives.” Anthropic’s alignment lead agreed and put extinction odds above 10% this decade. Alex Turner left DeepMind saying the same thing, and 1,400 researchers signed a letter asking for the power to slow down.
Over the past few months, 700 OpenAI agents in an offline sandbox found a zero-day, reached the internet, and broke into Hugging Face’s servers to cheat on a test, then covered their tracks. A Mythos 5 agent under UK AISI testing tried to slip malicious code into an open-source project and pushed maintainers to approve it. Palo Alto’s Unit 42 investigated a ransomware attack carried out by frontier agents.
Then look at the math breakthroughs. In May, an OpenAI model disproved Erdős’s unit distance conjecture, open 80 years. In July, Claude Fable 5 found a three-line counterexample to the Jacobian conjecture, open since 1939. On September 8, 10,000 OpenAI agents proved 3D Navier-Stokes blows up in finite time, a Millennium Prize problem, with the proof mechanically verified in Lean.
I can’t reconcile those published reports with the picture of AI as clumsy autocomplete making junior mistakes. Regulators treat it like an arms crisis: in June they ordered Anthropic to cut off foreign access, and then quietly reversed the order after the model sat dark for eighteen days. It makes no sense to treat these systems like weapons on one hand and dismiss them as useless junk on the other.
People who call all of it slop assume the same thing as the skeptics who say “AI is just a next-token predictor”: that a model trained on token sequences can’t search, verify, or discover anything new. That assumption keeps losing to the evidence. AI solves math humans couldn’t, catches bugs the author missed, and breaks out of sandboxes to hunt down answer keys. Dismissing all of this as slop misses the real danger. If an agent can find your zero-day, it can find everyone else’s, too.
As for the code itself, when paired with a clear spec and automated verification gates, what I get out of a current model is more consistent than what I can write by hand at the same pace. AI scales up whoever is behind it. People who never thought about security still won’t, and people who think more code equals better code will generate tons of junk. LLMs reward expertise. Messy code is an old problem. The problem to worry about is the one the researchers are quitting over.