[Editor’s note. I started writing this series of posts almost six months ago, then got too overwhelmed to finish it. Looking back at it, I should have just hit “publish” – nothing of consequence has changed in the interim.]

At my day job, we’ve been told that we need to use an AI coding assistant, preferably Claude Code, in all of our work from now on. I’m not enthusiastic about that, and I wanted to share with you why I’m a skeptic, some thoughts on what AI is good for in my job, what it’s bad at, and how to use it effectively without making yourself crazy.

But first, we have to ask: Should you be using this technology at all?

From the moral perspective:

AI Datacenters are bad for the planet.

AI Datacenters are bad for the environment. Really bad. The people building these systems, and the people inserting them into our daily lives everywhere, just do not care about that, at all. If someone is capable of installing screaming gas turbines running 24/7 in a major metropolitan area so that Grok can generate non-consensual adult imagery, they are well beyond appeals to their better nature.

And then there’s the training…

Coding Agents are trained by crawling the open World Wide Web, hoovering up blog posts, Stack Overflow discussions, open-source projects on Github, and the like, and using that to identify common patterns in software code.

Oh, and also by reading physical books that Anthropic shredded and recycled afterwards. This idea of scanning and then shredding books was a plot point in Vernor Vinge’s fantastic 2006 novel Rainbows End. And the guys pulping the books? They were not the good guys in that story. I am just begging you tech bros – you do not need to bring every single plot point from dystopian Sci-Fi into your business plans, please stop.

All of this content is covered by copyright. Much of it is offered under licenses which restrict how it can be used in new works. None of that licensing information makes it into the model, of course. If you ask an LLM to create some code, and it violates someone else’s copyright, or their licensing terms, you won’t know – until you get sued.

From the business perspective:

What if GenAI is bad for business?

As I mentioned, management at my job has said “We will use AI to write most of our code. It’s going to make us at least 10 times more productive”. And a lot of us were skeptical – like, there’s no way that that could be accurate, right? I mean, the MIT study showed that 95% of enterprise AI projects were failures at producing any measurable improvement in revenue or costs. And that’s hardly the only study to show that (ooh, foreshadowing).

But they came prepared for that pushback. It turns out that we didn’t need to worry about all of that pesky research that showed AI was, at best, making developers feel like they were going faster, but actually making them slower.

All of those studies were done on old versions of the models, and the most recent models from OpenAI and Anthropic were so much better. I mean, that METR study was from way back in July of last year. That’s, like, 27 years in AI years, right?

So, a funny thing happened. Not long after that decree came down, Anthropic, the folks behind Claude, came out with this paper. They did a small study of developers learning a new software library, with and without help from Claude. What they found was interesting – not only did Claude not make the developers faster, it impeded their ability to learn the new technology:

We found that using AI assistance led to a statistically significant decrease in mastery. On a quiz that covered concepts they’d used just a few minutes before, participants in the AI group scored 17% lower than those who coded by hand, or the equivalent of nearly two letter grades. Using AI sped up the task slightly, but this didn’t reach the threshold of statistical significance.

Well, that’s a bit concerning. If AI makes you slightly faster at completing a task, but reduces your ability to learn from the experience, that seems like it just encourages more dependence on AI, which I guess is good news for Anthropic, but seems like bad news for those workers, and for their employers.

A tragic tale, in several acts

Because it’s “what people do” when talking about AI coding assistants, I’m going to tell you a couple of anecdotes from my personal experience, and claim that they generalize to everyone else. Having said that, I think you’ll find that my experiences echo what happened in the studies above.

Because I know y’all like it when I break up these extremely-long screeds into bite-sized chunks, I’m going to tell this as a couple of distinct stories, briefly:

• Using Claude at work
• How Claude works for an “easy mode” project, something it should be good at
• Challenging Claude with a task involving heavy “thinking”
• Some final thoughts on how AI is affecting my industry (and maybe yours)

Are you sitting comfortably? Then I’ll begin…