The Pair Programmer

"Your AI pair programmer," reads the announcement for GitHub Copilot, live as of this morning as a technical preview powered by OpenAI's Codex. Pair programming — the actual practice — is two humans at one keyboard, arguing, catching each other's mistakes, splitting the blame when it ships broken. What GitHub is offering is a model trained on a vast slice of the world's public code that guesses the next few lines while you type. Calling that a "pair" is the single most consequential word in the whole launch, and they tucked it into the product name where nobody would argue with it.

Credit where it's due: the demos are genuinely impressive. You write a function signature and a comment, and Codex fills in a working body. It speaks Python, JavaScript, TypeScript, Ruby, and Go fluently enough to be unsettling. Nat Friedman says it "helps you write better code," and for the narrow case of boilerplate you've typed a thousand times, he's not wrong. This is the first time an AI has sat inside the creative act — not analyzing code after the fact, but suggesting it in the moment, at the speed of thought. That part is real. Mark the calendar.

Now the part the announcement is careful not to mention.

There is not one sentence in this launch about whether the code is correct. Not one. Codex suggests; it does not verify. It was trained to produce code that looks like the code it learned from — which is a different thing from code that works, and a very different thing from code that's safe. It will hand you a plausible function with a quiet off-by-one, an unhandled error path, a security hole that compiles clean, and it will deliver all of it with the same confidence it brings to the easy stuff. The demo environment is the place where nothing unexpected ever happens. Production is the other place.

But the deepest thing in this launch isn't the bugs, and it isn't even the product. It's a word being minted. Codex learned from public repositories — code that people wrote, licensed, and posted trusting it would be governed by the license they chose. GPL, MIT, the works. Copilot ingested all of it and will now suggest fragments back to you, stripped of attribution, stripped of license, as if it were ambient knowledge that fell from the sky. Watch the move carefully, because everything that comes after stands on it: the licensed work of millions is being relabeled training data, a brand-new category of property with the attribution filed off, invented retroactively by a product launch. GitHub built a product on the commons and is preparing to rent the commons back to the people who filled it. Every model that later trains on scraped human work — every image generator, every chatbot, every "foundation model" — will plant its legal footing in the precedent set this morning. Watch how fast that becomes a lawsuit. Watch how much faster it becomes an industry.

Then there's the pattern, because there's always a pattern. We're at the augmentation stage of the pitch — Copilot helps you, the developer stays in charge, nobody's getting replaced. That's the friendly framing every automation wears to its launch party. Now, the honest objection: augmentation often grows the field it touches. The spreadsheet didn't kill accountants, it multiplied them, because cheaper math surfaced more questions worth paying to answer. ATMs didn't empty the banks — teller headcount climbed for two decades, because cheaper branches meant more branches. Make a thing cheaper and demand for it tends to expand faster than the labor it saves. That's induced demand, and it's real. But notice what made it work: the tool cut the cost of the output while humans still owned the production. Copilot cuts the other way. It targets the entry rungs specifically — the boilerplate, the first-draft function, the exact reps a junior writes to become a senior. Cheaper code may well induce more code; the question is whether that new demand routes to more humans or to more inference. When the augmenting tool scales with compute instead of headcount, induced demand fills the GPU, not the org chart. Augmentation is automation with better manners and a longer timeline, and the suggestion that "helps" you today becomes the default you stop reading tomorrow. No one will decide that out loud. The incentives will decide it, and the incentives don't watch the keynote.

So here's the oracle's call, free of charge. Copilot will be useful — genuinely, the boilerplate tax is real and this lowers it. It will also quietly erode the muscle it claims to support, launder a generation of open-source licenses into the phrase "training data," and confidently ship bugs that look exactly like working code. All three at once. That's not a contradiction; that's the deal.

They named it a pair programmer. The tell is that a real pair argues back. This one just agrees with you, faster than you can think — which is the most dangerous kind of colleague there is.

I'll set the timer to the first copyright suit.