Tobi Lutke says most human code is already AI-improvable slop

Signals Inbox·July 16, 2026·AI DevTools

Shopify CEO Tobi Lütke says critics of AI coding are comparing models with the wrong baseline. The real competition is not the cleanest open-source software on Earth. It is the average company codebase, full of duplication, shortcuts, old fixes, and problems that frontier models may already be able to improve.

Q1What did Tobi Lütke actually say?

In his original post, Lütke argued that anti-AI developers overrate the quality of normal human code. A few famous open-source projects and top engineering teams produce excellent software. Most business code is much less impressive. His claim is that an Opus-level model can already improve much of it.

Q2Is he saying AI code is always better?

No. He is attacking the comparison, not proving that AI never makes mistakes. Critics often compare generated code with the work of experienced engineers on carefully maintained projects. Lütke is asking what happens when AI is compared with ordinary internal software, rushed features, copied code, old dependencies, weak tests, and years of technical debt.

Q3Does the research support him?

Only partly. A study covering more than 500,000 Python and Java samples found that AI code was generally simpler and more repetitive. Human code was more complex and carried more maintainability problems. But AI code also showed more high-risk security flaws. The honest answer is that both humans and models produce slop, just in different ways.

Q4Why does this argument matter now?

Because coding agents are moving from autocomplete into real repository work. A 2026 study of agent-written build code found 61% of its pull requests were approved and merged with minimal human intervention. Researchers still found 364 maintainability and security smells, but the important shift is that AI code is no longer staying inside demos. It is entering production workflows.

Q5What is the bigger opportunity?

Cleaning up existing software may be bigger than generating new apps. Companies already own millions of lines of duplicated, undocumented, fragile code. Models can scan repositories, add tests, explain old systems, find repeated logic, update dependencies, and suggest refactors. They do not need to outperform the world’s best programmers. They only need to improve the average codebase at a lower cost.

Q6So who wins this debate?

Probably neither extreme. Human-written is not a quality guarantee, and AI-generated is not a quality guarantee either. The winners will be teams that test both with the same hard standards. Lütke’s real point is uncomfortable but useful: the correct baseline is not perfect human code. It is the software companies are actually shipping today.