GitHub Copilot's Preview Is a Glimpse of How We'll All Code Soon

I’ve spent the past few weeks with GitHub Copilot’s technical preview installed in VS Code, and it’s the first AI coding tool that’s actually changed how I write code day to day, not just something I poke at once and forget.

Copilot came out of a partnership between GitHub and OpenAI, built on Codex, a descendant of GPT-3 that’s been trained heavily on public source code — think billions of lines pulled from GitHub repos. Instead of the old-school autocomplete that finishes a variable name, Copilot tries to finish your thought. Write a comment describing a function, and it’ll draft a plausible implementation. Start typing a for-loop, and it’ll often guess where you’re headed and hand you the rest.

It’s not magic, and GitHub isn’t pretending it is. In their own benchmark, Copilot got Python function bodies correct on the first try 43% of the time, and 57% of the time if you let it take up to 10 attempts. Read that carefully: more than half the time, even with multiple tries, it still doesn’t nail it. That’s the headline number worth sitting with, because it tells you exactly what kind of tool this is — a fast first draft generator, not an autopilot.

What it’s actually like to use

The good moments are genuinely startling. Boilerplate, common data transformations, regex you’d normally go Google — Copilot handles a lot of that instantly and often correctly. It’s less a pair programmer in the “understands your architecture” sense and more like a very well-read junior dev who’s memorized half of open source and will confidently type fast, sometimes toward the wrong answer.

The bad moments are the ones worth watching for. It’ll suggest code that looks right, compiles, even runs, but does something subtly different than what you asked for. Because the suggestions read fluently, it’s easy to accept them on autopilot exactly when you should be reading closest. That’s the real skill this tool demands: staying skeptical of confident-sounding code.

The licensing question nobody’s settled

The bigger fight brewing around Copilot isn’t about code quality, it’s about where the training data came from. Codex learned from public repositories, many under open-source licenses that require attribution or impose other conditions Copilot’s suggestions don’t carry forward. If the model reproduces a recognizable chunk of someone’s GPL-licensed function verbatim, whose code is that, legally? GitHub hasn’t offered a clean answer yet, and I doubt one arrives before lawyers get involved.

Where this settles matters far beyond one extension. If tools trained on public code become standard issue for developers, the terms under which that training happened set precedent for every model that follows. Worth watching closely, because right now it’s still just a technical preview — but the article for “the year AI writes half our code” is being drafted right now, one autocomplete suggestion at a time.