Three years ago, I wrote an article asking whether AI should be writing our code for us. My answer back then was a cautious no: AI could help, but coding is problem-solving, and I did not want to hand that part over. I ended the piece by saying I would keep my coding skills sharp and my AI in check.
So here we are in 2026. Has anything changed?
The numbers have changed
First, the obvious: almost everyone is using these tools now. In Stack Overflow's 2026 survey, 65.5% of the 12,255 developers who answered the coding-agent question said they used Claude Code, and 58.7% said they used GitHub Copilot. JetBrains' survey of 15,509 professional developers, run from May to July 2026, found Claude Code in use at work by 39% of developers, up from 18% back in January. Copilot went the other way, from 29% down to 21%, even as Microsoft announced 50 million Copilot users in July.
The leaderboard flipped in less than a year. The tool that was the default answer in 2023 is not the default answer anymore. Even the companies keep changing hands: Cursor, the editor a lot of developers switched to, was acquired by SpaceX in August 2026.
The question has not changed
But the question I asked in 2023 is still the same question: is the AI a helpful pair, or a crutch?
The research keeps giving us the same uncomfortable answer: it depends on who is holding it. A study by METR, run on 16 experienced open-source developers, found that they were 19% slower when working with AI, while believing they were 20% faster. They felt productive. The clock disagreed.
And a trial run by Anthropic researchers in February 2026 found something that should worry anyone still learning the craft. Fifty-two mostly junior engineers were asked to learn a programming library they had never used. Half of them got an AI assistant on top of search and documentation. The other half got search and documentation only. Both groups finished in about the same time. But on the comprehension test afterwards, the AI-assisted group scored 17% lower. They got the job done. They understood less of what they had done.
That is the crutch, right there. The code works, the understanding does not come along with it.
A power drill, not a foreman
Photo by Annie Spratt on Unsplash
An AI coding assistant is like a power drill. It drills holes faster than your arm ever could, and nobody sensible builds a house with a hand drill anymore. But a power drill does not decide where the holes go. It does not know whether the wall is load-bearing. In the hands of someone who understands the house, the work flies. In the hands of someone who does not, the holes are faster, neater, and in exactly the wrong place.
That is the deal I keep coming back to. Let the AI write the first draft, never the last word. Let it do the boring parts, the boilerplate, the code you could write in your sleep. But the understanding has to stay yours. The moment you stop reading what it wrote, you have stopped being the developer and become the person who approves things.
Three years ago I worried about losing my problem-solving skills to a shortcut. The tools got dramatically better since then, and my worry did not go away, it just got more specific. The danger was never that AI would write bad code. The danger is that it writes good enough code, fast enough, that we stop asking whether we understand it.
So here is where I land: use the tools. Everyone is. But use them the way you would use a power drill: for the drilling, not the deciding. Read every line before it ships. And every now and then, turn it off and write something from scratch, just to make sure you still can.
That is how you keep the pair helpful and the crutch in the toolbox where it belongs.
What about you? When was the last time you caught the AI writing something you did not fully understand, and shipped it anyway?
Top comments (0)