AI can write for you. But please don't let it.

I’ve spent the last month thinking about writing with AI (and laughing at the images I’ve had AI create for my slides, like the one above). I’ve given 3 presentations on it. One presentation was a brief 5 minutes about how important it is to preserve your unique voice when you are using AI to help you write. The other presentations combined that advocacy with suggestions for strong prompts, exercises for experimenting with the tools, and guidelines for revision. And throughout all of this, I have used AI. 

The caveat for me is that I use it to see what will happen and then I fight with it about what it does. 

For example, I asked it’s opinion on a title for a presentation. Instead of giving me an analysis, it gave me suggestions. I clarified that I wasn’t looking for suggestions. It apologized, and gave me the analysis I was looking for. 

As it produced it’s analysis, it used words like “I” to refer to itself and “you” to refer to me, as if we were two humans having a conversation. I’ve called it out on this before, and it says that “I” and “you” are just placeholders for positions in a conversation: speaker and listener. Which is true. And also, it’s a referent to a self that speaks or writes. And AI isn’t a self. And doesn’t speak. And doesn’t write. An AI is a machine that replicates patterns and generates output in response to inputs.

Even though it looks like it is talking with us, having a conversation, because the mechanism for our interaction is using the framework of a conversation. But it isn’t a conversation! When I’m typing in my thoughts, it isn’t responding with it’s “thoughts”. It doesn’t have any thoughts. It’s responding to the input based on what the data says is the best response–in terms of content and language. 

We’re at a point where AI writing is everywhere and in everything. You can’t open a tool without it asking if you want to use its AI to help you with your work. It’s pushy and it looks so useful–and it can be useful!

Because patterns are useful. And data is useful. And conversations are useful. 

AI is a pattern recognition and regeneration machine that we can interact with by using our own language instead of computer programming language. That makes it easier for us to prompt AI with anything we want to know and get an answer back.

But the key is that we get *an* answer.

This is the first rule of AI writing:

AI will *always* respond to the input. 

And if it can’t find an answer in the data, it may extrapolate from the data what a possible answer could be. Which may be wrong or bad.

I once did an exercise where I asked AI to generate a report and then to provide feedback on how that report could be improved and then to implement that feedback. I went through this loop a few times. The original report was clearly written but incredibly vague and boring–it sounded like every other report that has ever been written, like a template of a report or a satire of a report. The feedback AI gave on the report was generic too: clarify points, make content more concise, tell the reader a story. Those are good suggestions, big picture, but they are too big for a human to typically act on. When it regenerated the report based on its own feedback, the report got weirder and weirder–it still sounded like a report, and it was written clearly and correctly, but the language lost its meaning. I told AI that the writing was getting worse and worse and could it explain to me what it was doing and why it was doing it. The AI told me I was right and tried to explain itself but couldn’t because *it doesn’t know what it’s doing* other than responding to my prompt.

Here’s the second rule of AI writing:

It will always be grammatically correct and sound like it makes sense. But that tricks us into thinking the writing is “good”. 

What makes good writing? A voice. A human perspective. A person with experiences and values and goals trying to express themself to other people with different experiences and different values and different goals.

We have been taught to value clarity and correctness in writing because those are necessary in order for other people to easily understand your ideas. And your ideas were always already in the writing; they were there from the beginning and you were learning how to make them easier for other people to understand. 

Now, AI writing performs clarity and correctness, but it is empty! It has no ideas because it has no experiences or values or goals. And so it looks “good” based on what we’ve been told to value, but it doesn’t *feel* good anymore.

What matters to people is *you*. We do business with people we like and trust. And writing is one of the first ways we learn about who the person is. 

We don’t think about it, but the language that we each use is as unique to us as our fingerprint. Our vocabulary and sentence structure and way of putting ideas into words is a result of our particular background: where we grew up, who our parents were, what we have studied, what we’ve read, what jobs we’ve done, what we’ve become experts in. 

You are the only person who uses language the way you do.

Which is weird because language is a system, a pattern, that we learn when we are babies and don’t think about too much once we’ve mastered it. We are pattern recognizers and regenerators, too. The difference is that we learn patterns through our experience and we recreate them with our own flair. 

People who know you know what you sound like. And if you change what you sound like, they wonder what’s wrong with you. We notice a shift in the pattern that is you, and we notice it through language.

So, if you are doing business with people, you want your communication with them to sound like you, so that they can get to know and like and trust *you*. Sure, you’ll interact with them and they’ll get other information about what you’re like, but language will always be that first bit of information. 

What that means is use AI to write things that are for and about information, things that are intended to inform, not to connect: reports, procedures, legal documents. You’ll still need to provide a lot of instruction in the prompt and do a lot of revision, but the language doesn’t need to be yours because it isn’t affecting how much anyone cares. 

If you use AI to help you with writing that *is* for connection, I would suggest that you write it first and then you ask AI to make sure it is clear and correct, but only to make or recommend *necessary* changes and to avoid making or recommending any unnecessary changes. And then you make the decision about each suggestion and whether it aligns with how you would communicate and corresponds to your knowledge of grammar and clarity.