If AI Is Writing All the Code, Why Does Software Keep Getting Worse?

If AI Is Writing All the Code, Why Does Software Keep Getting Worse?

AI coding tools are better than they have ever been. Engineers using them are shipping faster. The models generating code are improving every quarter. So why does every app on your phone feel like it is getting worse?

An essay recently climbed the front page of Hacker News by asking that exact question. Its author, writing at ptrchm.com, catalogs a week of software failures that feel instantly familiar: a banking app that on average requires three rounds of FaceID before showing a payment confirmation. A car infotainment system that reboots during drives, where tapping the screen to open Google Maps brings up the radio instead. A refrigerator warranty claim form that fails silently on submission, its error visible only if you open the browser developer console 1.

The Hacker News thread drew extensive discussion among developers who recognize the pattern 2. This is not a fringe complaint. It is the unspoken contradiction at the center of the AI coding boom: the tools are getting exponentially better, and the products they produce keep getting worse.

The essay's argument is structural, not technological. Large language models are genuinely effective at finding and fixing bugs when given the chance. Nobody is giving them the chance.

The reason is incentives. Feature velocity looks impressive in quarterly presentations. A quarter spent exclusively fixing bugs does not. The essay imagines a product manager standing before leadership and announcing: no new features this cycle, we will focus entirely on stability. The fact that this sounds absurd in a corporate context tells you everything about why software keeps degrading.

AI coding tools do not create this incentive gap. They expose it. A team measured on shipping speed now ships faster. A team measured on feature count now builds more features. The capability multiplier amplifies whatever the metrics reward, and the metrics reward volume. When AI coding assistants became widely available, the question most teams asked was what else they could ship, not what they could now afford to fix. More code, not better code.

Part of the quality decline is also a story of accumulating complexity. Software was more stable in earlier eras partly because it was simpler, with fewer abstractions and fewer frameworks between the user and the machine. Each new generation of tooling added power at the cost of fragility 1. AI accelerates this trend. When generating code is nearly free, the temptation is to generate more of it: more features, more dependencies, more surface area for bugs, at a pace that testing and review cannot match.

The essay's author is not anti-AI. They describe the current generation of tools as having "given us superpowers" 1. That distinction is the entire story. The failure is not in the models. It is in the management culture that decides where to point them.

The result is a widening gap between what AI coding tools promise and what users actually experience. The promise is software that builds itself, perfectly and at scale. The reality is a banking app that cannot reliably render a login screen, a car that reboots its interface at highway speeds, and a warranty portal that swallows your claim without telling you it failed.

The tools will keep getting better. The incentives probably will not change on their own. If KPIs continue to reward shipping over stability, and AI makes shipping cheaper and faster, then the rate of quality degradation compounds. More code, shipped faster, with less review, pointed at features nobody asked for. The essay's author sees an opening: frustration with broken mainstream software creates space for individual developers to build better alternatives, tools that would have been beyond their reach without AI assistance 1.

That optimism matters less than the diagnosis behind it. The tools are not the problem. The metrics are. Until someone figures out how to make stability worth optimizing, the gap between what AI can build and what users actually tolerate will keep widening. The question was never whether AI can write good code. It is whether anyone deciding what code gets written has any reason to care.