Mind Readings: When to Use Generative AI in Writing

In today's episode, let's talk about when to use generative AI when it comes to writing. What are these tools good at? What are these tools bad at? When should you use them? When should you not?

Okay, so writing is not monolithic. Writing is a blanket term for creating using words. And those creations can be wildly different. They can be have varying purposes, they can have varying lengths, and so on and so forth. There's a lot of variation in writing.

So we're going to talk about two, maybe three different dimensions of writing. First, the first dimension is the purpose of the writing. Is it, and I'll break this into two camps: creative versus commodity. There is a lot of writing that we do on a day-to-day basis that is commodity writing, meaning it's it really wouldn't matter who wrote it, um, because it doesn't have to be a particularly artistic, it doesn't have to be particularly creative. You know, when you're writing a memo to a client, uh updating them on the status of a project, do you need that to be purely surprise-winning?

Do you need to have it capture your voice and and the flavor of your personality? No, you need to say here's the status of the project, right? And that's an example in the commodity space where, yeah, that is something machines should do. Uh for example, uh, this is a uh different domain. Think of the last painting that you saw that you really wanted to buy in a hotel room.

Have you ever wanted to buy the painting in your hotel room? Probably not. Why? That is commodity art, right? It's it's inoffensive, it's generic, it's fine, it's okay, it's suitable, right?

But it's not, you don't look at it and go, wow, that's that just moved me. You're like, here's my bag, throw it on the bed. Oh, oh look, there's a painting over the bed, cool. Right? And it's usually like a flowers, a sailboat or uh uh uh, you know, um some buildings.

It's it's very rarely. I've not seen any paintings, and I've stayed in a lot of hotel rooms where that was essential. I had like I had to take a photo of it. I've never taken a photo of a of a painting in a hotel room that that really moved me that much. So that's commodity content, right?

You make commodity content all the time, status updates, emails, meeting agendas. Commodity content is the domain of AI. Commodity content is the domain of generative AI. It is where generative AI is best. It is where generative AI does a good enough job, or maybe even slightly better than good enough to pretty good.

For example, there are uh a number of applications. There's one I remember in the early days of of Chat GPT, where a contractor who was dyslexic, severely dyslexic, um, had a friend who wrote him an app that translated his shorthand into fully, you know, full grammatically correct uh memos for project updates for clients, right? That is a brilliant use of AI because it's commodity content. It really doesn't matter the tone or the quality of the writing so much as it gets the job done. The other end of the spectrum is the creative writing.

When you are writing poetry, when you are writing long form works, when you're writing a uh great piece of fiction, machines today in their current form are not doing a great job of that. With some exceptions. There are exceptions to the to that statement, but for the most part, um the machines don't create great creative writing because the very nature of how language models work, right? They are probability engines. They choose the highest probabilities, and the highest probabilities are the most generic ones because they're the most common.

Creative writing by default, if you're doing it well, is all relatively low probability words, right? Uh, my friend Ann Hanley started her newsletter this week saying uh, what's up, uh, what's what, Walnut, right? That is not a word sequence that you're gonna find a lot, and that that unusual tone of machines are not going to know to do that. Now, again, there are some ways to get around that with very, very detailed prompts or with fine-tuning models, but for the most part, most people are not going to do that. So that that's sort of one dimension commodity versus creative writing.

The more the writing you're doing is commodity, the better machines are going to be for that task. Now, here is the question. The question is, do you know what kind of writing you're doing? And should it be a commodity or should it be creative? I would argue for something like you know, a status update on a project or a meeting agenda.

Yeah, that's pretty commodity. It's you can art invest a lot of time to make artful and very creative meeting agendas, but it's not really going to change the outcome of the meeting. And you might get some notice for it. You know, for example, Dr. Bronner's gets noticed for the enormous amount of text on their soap bottles, but for the most part, it doesn't really change the experience.

It's not something that you actively seek out because of that writing. Is your corporate blog commodity writing? Should it be? Those are questions. Because people are saying, yeah, we can use machines to do this commodity writing.

Well, that's true. Absolutely true. You can do machines to do that commodity writing, but should it be a commodity? If the purpose of content marketing is to attract new audiences and to compel people to want to work with you, there's a lot of low probability things in there, right? Your tone, your personality are parts of your sales process.

And if you're using if you're creating commodity writing for a creative task, it's going to come out pretty pretty generic, and that's not really good for marketing or sales, right? Generic is generally not good for marketing. Umless you're that one company that makes the the like the white beer can with the word beer on it. And then that's you know an exception to that rule. So the second question is on the task.

Large language models have two fundamental modes comparative and generative. Um the easiest way to think about this is editing versus writing. There are plenty of tasks that you do that are creative, that are writing. Generative models are not as good at writing as they are at comparative tasks, at editing. That is why without sufficiently strong, big, huge prompts, they don't write well compared to a very creative human writer.

That is also why they summarize so well, because they have the original version and they're basically making a comparative version. They compare by they generate these summaries by comparing back to the original. And so summarization, extraction, uh rewriting. These are tasks that are comparative in nature on the model side, which means they're editing tasks. And as a result, these models are very good at it.

QA answering, yeah, that is generative. They're not as good, they're not as good at that. You've all had the experience of uh having these things generate like factually incorrect statements. That's because they're that's not what these models are really good at. They're really good at comparative tasks.

So that's sort of the second aspect is is what kind of writing are you doing? Is it are you asking them all to do? Is it is it comparative or is it generative? And you match that with your commodity versus creative. Having those dimensions gives you the ability to classify when you should be using AI.

For commodity writing tasks, yeah, you should be using generative AI. For creative writing tasks, maybe not. For the commodity tasks, are you doing commodity generation? You can still use AI for that. Are you doing commodity comparative tasks like summarization?

You absolutely should be doing AI with that. For creative writing, should you be doing summarization and extraction? Yes, because that's what again, it's it's good at that and can still preserve tone and things with summarization. But should you be using it for generation and creative writing? Probably not.

You're gonna lose a lot uh in that. So those are a couple of different ways to think about when to use AI, generative AI within the writing process. Take a look at the things that you want to do with AI, take a look at the things that you already do with AI, and ask yourself which of those buckets it fits in. And that will help guide whether it's a good idea or not to to be using artificial intelligence in those contexts. There are other considerations as well and a lot of people will use very specific examples but you you fundamentally want to look at the underlying architecture to to make the judgment about when AI is a good choice or not.

Thanks for tuning in. Talk to you next time if you like this video go ahead and hit that subscribe button.