Welcome! I’ve lost track of what issue number we’re on, so I guess that means I’m an old hand now.

On today’s quest:

— Two ways to specify tone
— Custom GPTs
— Copyright infringement developments

Tip: Get Output in Any Tone You Want

Last week, I talked to a guy who isn’t impressed with the output he’s getting from ChatGPT. It’s bland. It’s boring. And it’s true.

Out of the box, the writing probably won’t be what you want. But it’s easy to fix. I’ll outline a basic way to fix it and a more interesting way.

1. Tell it what you want

First, you can specify the basic tone for responses. These are some general suggestions I got from Claude:

• Formal/professional: Respond in a polite, professional manner using proper grammar and language. Avoid slang, colloquialisms, emoji, and excessive humor.

• Friendly/conversational: Use a casual, conversational tone as if we're having a friendly chat. Feel free to use first-person pronouns, contractions, humor, etc.

• Enthusiastic/excited: Respond in an upbeat, energetic tone conveying enthusiasm and positivity. Use exclamation points, positive language, etc.

You can put these in custom instructions so they apply to every conversation, or you can use different instructions for different prompts.

2. Let it explain to you what you like

Second, here’s where it gets interesting. Let’s say you give it instructions like the short lines above, and you still don’t like what you get. It can be hard to describe exactly what it is you like about the way you write.

But the AI tools are astonishingly good at describing writing.

Enter a piece of writing that is like what you want — an old blog post you loved or 10 of your favorite social media posts — and ask the tool to describe the tone.

For example, I asked Claude to evaluate the tone of a blog post I wrote when I was a journalism professor about websites getting rid of their comment sections. The analysis was actually as long as the article, but here’s a condensed version (bold added):

I have always been earnest. Sigh.

Next, I asked Claude to describe the tone of an AI Sidequest newsletter from last week. I would say I’m trying to make these “more fun” than my professorly blog post, but Claude spews a litany of adjectives and specific features that form the tone (again, this is a condensed version):

As you can see, if I were trying to get Claude or ChatGPT^† to tweak a piece of writing to make it more like my newsletter, giving directions that incorporate all these specifics descriptors is more likely to work than what I would have originally said: “Make it more fun.”

AI still can’t replicate a human

As an aside, the whole analysis fascinates me. In both cases, my writing style is something I’m doing subconsciously. I didn’t consciously set out to have an earnest, observational tone in my blog post, and although I am aiming to be friendly and personal in the newsletter, I haven’t been using first and second person, contractions, and parenthetical asides in a calculating way. It’s just what I do when my brain is saying “be friendly.”

The only part I’m doing somewhat consciously is including personal anecdotes and observations, which is something AI can’t replicate. AI can’t say, “I listened to the new Beatles song for you; here’s what I think.” It can’t tell you, “I’m trying to make these newsletters more fun than my professorly blog post.”

And I’m including those on at least purpose because I want you to know that I’m actually writing this newsletter. Fingers to keys. It’s me!

I am finding myself doing more of the same in my Grammar Girl podcast too. In the last show I recorded, after pronouncing a particularly difficult word, I had a little aside saying, “You can bet I had to look up how to pronounce that one!” My show is scripted, but that wasn’t in the script. I just threw it in as I was recording because in the back of my mind, I now always have a tickle that says “Make sure people know this is you and not AI.”

Tip: Train ChatGPT on Your Own Data

I’ve been playing with “GPTs,” custom tools OpenAI announced last week that let you train an “assistant” with your own data. I’m impressed, but it’s only available to people with a paid ChatGPT subscription, it’s still in beta, and it’s a little more advanced, so I suspect most of you either wouldn’t have access to it or wouldn’t be interested. But it’s definitely the #HotNewThing of the last few days.

I’m sure I’ll have more about it in the future, but if you want to dig in now, I watched this how-to video on YouTube to get started, and Christopher Penn has a good overview in his email newsletter, Almost Timely News, which I’ve subscribed to for more than a decade and always find insightful.

News

I have deadlines for my real job, and my husband has been sick, so I’m keeping news short today.

US judge dismisses parts of AI copyright lawsuit against Meta

howtobe247.com

A judge dismissed the claim that the output of Meta’s AI model violates authors’ copyrights. His thinking seems to be that nothing you generate is similar enough to violate copyright. This is the “Sarah Silverman” case you may have heard of. However, he is allowing the authors to amend another key claim: that the company violated their rights by using their books as training data (and from a pirated source no less).

“Wired” did a good job back in September describing the pirated books databases that were used to train many AI tools and the position of the Authors’ Guild, which feels using these databases is copyright infringement (not surprisingly) and many legal experts who aren’t so sure.

The “Hard Fork” podcast also had a copyright lawyer as a guest, and her opinions were interesting — and not that encouraging for authors. What stuck with me is how expensive it is to bring these claims. Her example was of a legal website that is suing some model that used its training data. If they win, the legal website will get a payout, but none of the authors who wrote the articles in the database will get anything.

What is AI sidequest?

Using AI isn’t my main job, and it probably isn’t yours either. I’m Mignon Fogarty, and Grammar Girl is my main gig, but I haven’t seen a technology this transformative since the development of the internet, and I want to learn about it. I bet you do too.

So here we are! Sidequesting together.

If you like the newsletter, please share it with a friend.

^{*My first post was in December 2022 about ChatGPT recommending books to me that didn’t exist: ChatGPT Sees Dead People.}

^{† I have been using these tools mostly interchangeably, often just depending on what tab I find first, but in the last week or two, I’ve been happier with the output I get from Claude over ChatGPT. Although I know these tools don’t think, the best description I can come up with for why I prefer Claude is that the responses seem more “thoughtful.” However, as far as I can tell, Claude doesn’t have custom instructions yet, so that’s a downside.}

^{Written by a human. Copyright 2023, Mignon Fogarty, Inc.}