The Age of AI Testing

We talk a lot about the explosion of code being written right now. And the numbers are staggering — whatever multiplier you want to use, 10x, 100x, GitHub alone is seeing nearly double the volume of code pushed to repositories. And that's just what makes it to GitHub, which I'd argue is still a fraction of what's actually being generated. It's an industry-level shift, and honestly, I think we're still in the early innings of the adoption curve.

But here's what we're not talking about: we have never generated so much bad code.

The 80/20 Reality

I say this from experience. I've been building a lot — AI pipelines, dashboards, orchestration workflows, internal tools. The creative velocity is real. Something sounds cool, you spin it up, ship it, move on. But when I look back honestly at the output, probably 80% of what I've generated is throwaway. Experiments that didn't pan out. Prototypes that taught me something but aren't going anywhere. And I think that's fine — it's part of the process.

The cost of generation is so low right now, largely subsidized from the AI tooling side, that rapid experimentation is basically free. You can try an idea in an afternoon that would have taken a week two years ago. That's genuinely powerful. The things that do stick — the maintained applications, the production features, the real product work — AI is giving me superpowers there. But the ratio of valuable output to total output is still surprisingly small.

The Graduation Problem

Here's where it gets interesting, and where I think enterprises are really struggling. There's a growing middle layer of AI-generated work that is valuable but not ready for prime time. Someone builds an internal tool. People start using it. It solves a real problem. And then... nothing. There's no process to say, "This crossed the threshold — now it needs proper testing, security review, maintainability standards, and an owner."

What I've seen firsthand is that most organizations don't have any walls around this. They let people generate freely (which is good), and when something becomes genuinely useful, they just... let it keep running in its prototype state. No graduation ceremony. No lifecycle management. No one asking whether this thing that three teams now depend on was ever actually built to be depended on.

The Waste We're Not Measuring

Even setting aside the throwaway experiments, there's a waste problem inside the code that is being kept. When you're generating at speed, you accumulate debt at speed. Dead features. Redundant modules. Untested paths. Code that works but that no one fully understands because it was generated in a burst and never properly reviewed.

We need to start thinking about this the way we think about any other lifecycle:

Pruning. If something was an experiment and it's not going anywhere, delete it. Don't let dead code sit around as organizational noise.

Testing. The code that graduates to "actually valuable" needs real test coverage — not just the happy path that was validated during the demo.

Ownership. Every piece of AI-generated code that enters production needs a human who understands it, owns it, and is accountable for it.

Lifecycle management. Even valuable tools decay. Are we monitoring whether that internal dashboard someone built six months ago still works? Still returns accurate data? Still has users?

The Real Multiplier

AI is giving engineering organizations 100x generative capacity. That part is real. But capacity without governance just means you produce 100x the mess. The organizations that win here won't be the ones that generate the most code — they'll be the ones that build the platforms and processes to separate the signal from the noise, graduate experiments into production-grade systems, and ruthlessly prune everything else.

We're in the age of AI-powered experimentation. That's exciting. But we need to match it with an age of AI-informed quality management, or we're going to drown in our own output.