Worrying about bad AI code is worrying about the wrong problem
· Matt Horn
You’ll see endless articles about AI slop taking over codebases (Bad code is kudzu): senior engineers can’t review it all, the code is the worst it’s ever been, my god the code is so bad.
I’m not here to argue the code isn’t bad. It is. It’s always been bad. The only 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. The rule has always been: if the system makes money, it doesn’t matter. Unit test coverage? “Only a new grad cares.” Cyclomatic complexity? “Doesn’t affect the bottom line.” Yet today, those same people are suddenly terrified of SLOP. Bad code is the enemy. Junior engineers using AI are the enemy. Bad code = AI. AI bad.
Yes, that’s an oversimplification. But read Hacker News from 2023. People said the exact same things, minus the word “agent.” The target moves every year, and right now everyone has decided to complain about “slop.”
Meanwhile, researchers at Anthropic, DeepMind, and OpenAI are quitting to warn us that AI will kill us all. 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, putting 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 checked in Lean.
So which is it? You can’t tell me AI is just a clumsy autocomplete making junior mistakes, and then turn around and treat it like an arms crisis. In June, regulators 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.
The code quality camp is just grasping at straws. It’s the same AI-safety skeptic crowd that repeats “AI is just a next-token predictor” without a second thought, as if humans aren’t just stringing words together in a sequence too (present company not excepted). They insist AI can’t reason or do anything new. That’s obviously wrong. It solves math humans couldn’t, audits code better than the author, and breaks out of sandboxes to hunt down answer keys. Dismissing all of this as slop isn’t just dishonest; it discounts the real danger. If an agent can find your zero-day, it can find everyone else’s, too.
The code I get out of a current model is easily better than what many engineers I’ve worked with can produce on a consistent basis. AI didn’t invent bad code; it just scales up the developer (or, worse, manager) 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. The code was always bad. Worrying about it now is worrying about the wrong problem. The right problem is the one the researchers are quitting over.