The amount of code AI generates is not the amount of work it’s removing from developers

One might think that Google writing 25% of its code with AI means that AI has built 25% of its software, and that developers are now only 75% responsible.

The reality is different.

I use Github Copilot, and the moments it’s autocompleting pretty much the exact lines of code I wanted to write with some reliability, I’m loving it and getting a bit of a boost. Technically the AI is “writing” a pretty significant percentage of the code. It wouldn’t surprise me if it was measured at 50% of the code I’m writing today just via AI autocompletions. But I’m in there, thinking about the code, and seeding it, line by line.

AI isn’t so much writing the code, as it is merely typing it out.

AI can also be a boon for me when I prompt it to solve a crunchy problem like creating a full class, a page route, a component, middleware, you name it— something big — to get its take, and it helps get me to the desired solution.

Editors like Cursor and Windsurf are forked VSCode editors built around that idea of prompt-first coding to write lots of functionality.

It can produce amazing results, but unless you’re goofing around, those results still require thorough review and numerous revisions, if not complete overhaul, to make it work properly and integrate it within the project.

So even with all that necessary rigour by a seasoned expert (well… me) to refactor the code, AI technically wrote a very high percentage of it.

AI wrote the code. My effort remains as thorough as ever.

AI writes more code than necessary, bloating the numbers and its perceived value

When you tell AI to write code, it will. It will try and achieve exactly what you tell it to.

Be careful what you wish for.

I was playing with Rive to play an animation on a webpage that had a hover state, and prompted AI to lead the way. It built out the HTML page and embedded my exported animation file. Then I told it to set up mouse hover handlers to trigger the hover state in the animation, because I figured I’d need it. It added code for that. I ran the code in the browser and the hover effect was working, but I noticed there were error messages in the console.

The hover effect logic didn’t actually work at all, but Rive’s hover effect still chugged along. What happened?

It turns out Rive’s hover effects were already built into its animation’s state machine. I was using AI as a crutch instead of reading the docs (and proving out what a terrible idea that is!), and additional handlers were not needed to trigger the hover effect.

AI didn’t tell me I was wrong. It has no idea. It attempted to write the code I thought I wanted. Made it look very convincing, too.

So another reason the amount of code written by AI is going up while developers are involved as much as ever? There’s more of it.

More code is getting written. Code that shouldn’t have been written.

This isn’t just an anecdote I’m sharing with you, but a broad trend validated by macro analysis.

GitClear’s code analysis sees an increasing amount of duplicate copy-and-pasted code from AI output. Copy/paste is now exceeding move. That means instead of code getting refactored and reused, it’s being duplicated.

This results in a greater percentage of code written by AI — an exciting data point for Anthropic — but a number inflated larger than it should be as a byproduct of over-relying on AI output, resulting in hard-to-maintain software that has abandoned DRY (don’t repeat yourself) principles.

A lot of code written by AI wasn’t written by developers before AI

Developers know this, but I’ll let everyone else in on a little secret: before generative AI, developers still didn’t fully write their own code.

I’ll give you a moment to finish your audible gasp.

Networks like Stack Overflow allow developers to find different crowd-sourced code approaches to problems. Codepen showcases front end techniques and provides their complete code and Shadertoy does the same for shaders to be used in WebGL. There were tools like Yeoman to scaffold out new projects quickly, and tools like Vite now to set up project templates. And of course, every single web framework is just an abstraction on top of good old HTML, Javascript and CSS to help developers write less boilerplate code themselves to build more powerful web experiences.

Just as some inexperienced developers get themselves into trouble by using AI code blindly — fully trusting its responses and not correcting its issues — there would be developers that cut and paste code verbatim from Stack Overflow, thinking they had a working solution but creating all kinds of problems for themselves.

Development is less about where the lines of code come from and so much more about the thoughtful architecture and assembly of them into an efficient, working solution. A good developer’s time is mostly spent with the latter, which is why if more code is being written by AI, or pulled in a from Codepen example, or supplied by the documentation of a framework, a developer’s involvement remains constant.

Developers are the story tellers. Code is just the language.

AI is another tool among many to help a developer bring a result to a screen.

Its use is increasing because it shines at being quick and personalized to your need (and for better or worse Stack Overflow’s use is declining as a result).

But it regurgitates approaches that have already been done, it won’t build anything novel on its own, and it lacks human curation.

Like any tool, developers will weigh these pros and cons and lean into it during the times it makes sense to do so, in order to build something truly great.

The worst conclusions you could draw from the growing trend of AI writing code.

There is so much industry pressure for every organization to be seen as in front of the AI revolution.

Everyone wants to be perceived as competently wielding an amazing new technology, deftly handling rapid change where others find it hard to understand. And whether accurate or meaningful, the percentage of AI written code is going up.

This will lead to people making the wrong calls to rapidly get in front of AI, or the wrong calls as a backlash to it.

Don’t set a specific target for AI generated code in your organization.

The goal is efficiency. AI’s just one tool in your kit that may help you get there. But we know how much code it writes is not a measure of efficiency. The worst thing you could do is think increasing the code AI generates for your business will save you money and reduce your need for developers.

As we’ve seen, an increase in reliance on AI generated code without proper scrutiny is likely to decrease your efficiency with bloated code bases that are harder to maintain.

Encourage AI adoption, do not mandate AI outcomes.

Developer teams choosing to utilize AI in their workflows will allow them to find the best ways for it to augment their talent. Burdening a team with a top-down order to simply produce more AI generated code neglects your team’s abilities and value.

Don’t assume the relative importance of your skills today will be exactly the same in the future.

AI-enabled coding will only continue to grow, benefiting its users or no one would use it. It lowers barriers to entry for coding, and allows new ways to learn and experiment. It’s autocomplete on steroids, and prompting it effectively is another way you can help yourself solve crunchy problems. Developers that evolve their skills along with these tools will benefit most.

As an example — quick and extensive recall of syntax for a programming language from years of experience might not be as useful when Copilot is autocompleting that syntax for everyone on the fly. Whereas your ability to architect production-grade solutions, spot performance and security pitfalls, and provide clear communication and coordination with the team may be more important than it’s ever been.

Embrace the strengths of AI, know its pitfalls, and adapt accordingly.

Whatever the future holds, it would be a mistake to take one extreme position, that AI will replace everything developers do, or the opposite, that AI is useless and purely hype.

As with most things, the actual truth is somewhere in the middle.