This article was written by a living, breathing piece of meat
Here I go, making the case for AI writing. I don’t especially want to make this case, but I feel I have to.
I get it. The backlash comes for a reason. But people have kind of lost their minds here. If you accept writing software with LLMs, then writing isn’t meaningfully different. Both are intended to be read by humans. Both can be driven by humans, leaving the agent harness as a glorified text editor.
That is, you the human are in control. You can do a good job or a poor job. You’ll produce slop, if you allow it. But instead of judging pieces of writing on their own, we have people now looking at every piece of writing they come across with a microscope hunting for AI tells. Slow down.
Your Release Notes Are Not Shakespeare
The point of most content is to get a message across. It’s not to make some grandiose statement. It’s not to be elegantly worded. It’s to communicate some direct thing to its audience.
In software we spend a lot of time just telling people what work we’ve been doing, release notes being one example. Here AI works just fine, once properly prompted. If you’re able to explain to the agent where the writing is going; that could be a blog, release notes, and tell it to use a tone appropriate for it. Release notes for example should be direct; explain what was released, that’s all. No marketing. No “not this, but that” AI tells, just a direct explanation of what happened.
That is to say, purpose determines tolerance. When the purpose is conveying personal thoughts, AI should be used lightly. When the purpose is to convey information, it can be used extensively. Some more examples where it works fine:
- Documentation and READMEs - Nobody wants a “voice” in a setup guide; they want it to work.
- Status updates, stand-up notes.
- Support and FAQs - Using agents for support can be a better experience in some settings. AI provides instant answers.
- Writing that’s intended for other AI to read - You might scoff at this, but it’s going to become more common than you think.
Conversely, I’d never want AI to write:
- A wedding toast. Be a human for once.
- A eulogy. Ditto.
- A job cover letter. It’s supposed to prove you want the job, and you can’t do that unless you write it yourself.
Everything else probably falls somewhere in the middle. That includes essays like this one. I personally prefer to go unfiltered, unpolished. Articulating my thoughts is something I’ve always struggled with, and even when I write them down myself, they always feel somewhat distant from what I truly believe, but at least it’s my own translation of those feelings.
The Moral Panic Has History
People aren’t ready to admit this, I think, but the moral panic around AI writing is nothing new. I’m old enough to remember when graphing calculators starting getting good. It was said that it would mean that people wouldn’t learn math. However, this did not pan out at all. We still have plenty of mathematicians. And having a tool for doing math is acceptable, as long as you understand how it actually works. We should apply the same standard when it comes to writing.
You can say the same thing about a number of other things like spell-check - people won’t learn how to spell words! You can go back to when photography replaced hand-painted portraits, I would imagine.
The tools change but the panic repeats, and yet the work still gets done.
What This Is Really About
There are a few reasons why this became so widely discussed; some are even legitimate. The general discourse around “AI taking jobs” is in play here. Think translators, illustrators, and customer support reps. When it comes to things like documentation, things that normally have been done by humans with a lot of effort, replacing that labor with AI is understandably hard to swallow. I’m not sure what the answer is here, just like everyone else.
But I also think there’s an element of feeling that you’re not putting in the effort when you use AI to write. I mean, I still see this when writing code with AI. I myself feel guilty submitting a PR when I was able to fix an issue too quickly, even though I put in the effort to fully understand the code. It’s the same standard as with code: I trust AI-assisted code once I understand what it produced, and I hold my writing to that same bar.
Humans are quick to judge and slow to forgive. We could all use a reminder to give ]our fellow humans the benefit of the doubt.