In this bonus segment, Daniel Heuman shares his strong stance on AI and copyright, how his company safeguards editorial privacy, and why readability scores aren't everything. Plus, we get his great book recommendations.
In this bonus segment, Daniel Heuman shares his strong stance on AI and copyright, how his company safeguards editorial privacy, and why readability scores aren't everything. Plus, we get his great book recommendations.
Draftsmith → draftsmith.ai
🔗 Share your familect recording in a WhatsApp chat.
🔗 Watch my LinkedIn Learning writing courses.
🔗 Subscribe to the newsletter.
🔗 Take our advertising survey.
🔗 Get the edited transcript.
🔗 Get Grammar Girl books.
🔗 Join Grammarpalooza. Get ad-free and bonus episodes at Apple Podcasts or Subtext. Learn more about the difference.
| HOST: Mignon Fogarty
| VOICEMAIL: 833-214-GIRL (833-214-4475).
| Grammar Girl is part of the Quick and Dirty Tips podcast network.
| Theme music by Catherine Rannus.
| Grammar Girl Social Media: YouTube. TikTok. Facebook.Threads. Instagram. LinkedIn. Mastodon. Bluesky.
Mignon:
Greetings, Grammarpalooizians! I am here with Daniel Heuman. We just talked in the main segment about his, all his thoughts about AI, his new product, DraftSmith, but also just general thoughts about AI, editors, and climate, and how it all works. If you'd missed that, check it out. It's in the main feed.
And now for our bonus segment, I want to talk about a few things: copyright, maybe a little bit more about the product, and then his book recommendations. Daniel, thanks again for being here.
Daniel:
It's such a pleasure. Thank you.
Mignon:
You bet.
So, you actually had an article published, or a letter to the editor published in The Guardian about AI and copyright. There's something legislatively going on in the UK, I think, and you had some opinions. Can you talk about those with us today?
Daniel:
Absolutely. I guess my starting point on this is an article by Benedict Evans, and he calls it a completely new argument we've been having for 500 years. He outlines that every time there's something new a new technology comes along, it creates all these issues of copyright questions. And he poses it in interesting ways. So what happens if you have an AI that can imitate Taylor Swift perfectly? Well, that's got to be wrong. What about a person imitating Taylor Swift? Well, that's okay. Why do we have those two lines?
Well, I think those two are fairly clear where most of us would stand on that. Bizarrely, legislation does not necessarily agree with that, and that's kind of the first battle we're all having, which to me are a bunch of questions that are straightforward where rules have been broken.
Then it does get more interesting, though. What happens if you have an AI that's trained to deliver a mix of styles, and maybe it's been trained on half the pop songs out there, and the other half of the authors not get anything? If you've trained on these and paid people properly, is that fair use? I don't even know at that point because it's a mixture, and it's not imitating any one person. So you can get into some quite complex things.
The actual battle that we're having today is not complex. What AI companies have done is very, very clearly wrong. And Andrew Lloyd Webber wrote to the Guardian about this, and I should caveat immediately that I do not agree with most of Andrew Lloyd Webber's politics. I definitely don't agree with Andrew Lloyd Webber's taste in musicals.
Mignon:
Harsh.
Daniel:
We know we won't even bring up "Cats." I'm not going to go there. It's unnecessary. I've said what I have to say.
But I really deeply agree with what you had to say on copyright. Everything you were saying was obviously right. And when you read something like that, you sort of go, "OK, well, he's right. Now I've taken that in." And this is one of those moments of, "Oh, I think I can say something here." I don't really think of myself as a tech CEO, but I guess I'm a tech CEO. It's a very small company, but that's kind of what we do.
And the UK government is banding all of these tech companies together, and it's false.
Actually, we do not want changes in copyright that support AI companies to the detriment of artists. We need artists, we need creatives, we need writers, we need editors; we need all of these people to be protected. And the idea that copyright law could be weakened is exactly the wrong way to go in that debate that Benedict Evans was describing. It's completely wrong.
So no, I'm very strong opinions on the subject. And because I could write from this position of, I guess it's an interesting position of a tech CEO, but I think a lot would feel the same way. It's not banned all tech together. And the Guardian then published it. So then it was just a quiz question: who reads the letters of this stuff? How many people? A surprising number of people saw the letter.
Mignon:
Yeah. Yeah, so your letter basically just said, like, we don't want changes to the copyright law. We want copyright law strengthened.
Daniel:
We want to clarify that, of course, creatives need to be reimbursed or paid for use of their work, and no AI company should have any sort of shortcut around it. And if you think this is going to help the UK to weaken these laws, absolutely not.
Yeah. And I know you have strong thoughts about privacy, especially because you're creating a product for mainly corporate workers who may be submitting things to the product that their companies would not want to go into the training data. So, are you confident that the things that people are putting into DraftSmith that then get sent to ChatGPT are private?
Daniel:
Private? Yeah, absolutely. So, I mean, our history on this is that with PerfectIt, it's an offline product. And as a result, people have that highest level of confidence that it's protected. And so, people with needs for sensitive documents, they don't want anything going near online, have always been able to come to us. So that's, as you say, that's our audience.
And it doesn't — it's not just big corporations. It's individual editors sign contracts that are very clear about what they can and cannot do with a document. The fact that always ticks the boxes because it does not send any data of any kind, anywhere, for any reason. And with AI, the way it is today, well, yes, okay, you can get some of the models running locally, but you're going to deal with such a slowdown. It's not a good enough performance yet. We may well get there, but today, running locally just isn't going to be the answer.
And so, our compromise for that is that we use the Microsoft Azure OpenAI service.
And that's Microsoft's partnership with OpenAI, but it's got this break in it, which is nothing goes to OpenAI, and of course, you are trusting Microsoft with that data. It is Microsoft Azure, but honestly, if we can't trust Microsoft with data, then we're all — you know, SharePoint has every corporate document in the world at this point, so I kind of think I can't prove that they are doing what they say, but they have the highest level of FedRAMP, SOC 2, ISO this and ISO that, you know, around that product. They've got the highest level of security certifications they could possibly have, so I certainly do believe it, and we followed that, and we even went a little bit further. It's not just that we're using a version that has a break — nothing goes to OpenAI, nothing goes to Microsoft; we just erase it immediately afterwards so we can't see it. No one else can see it, and it doesn't need to be stored. So, you cut down on a ton of risks if you're not using this for training and not storing it. It's just like, okay, technically to use the AI models has to be on a server for an instant, but then it's gone.
Mignon:
Wonderful. And then one of the things that you think that your product does well, or that's interesting that your product does that we didn't actually get to talk about in the main segment, is dealing with readability scores.
So, you know, you talked about how it could rephrase things into plain English, which is something I love because I created a course on plain language. But readability scores are kind of part of that, but kind of something different. So how do you think that AI or DraftSmith does readability scores better?
Daniel:
So it's really interesting, with all these things, you have to start from the big picture. And you covered it just there when you mentioned plain language.
Plain language is about so much more than the words you choose. You can't just run an AI, and fix the language, and think that's going to do any good. It's about how you access a document, how you view a document, how you find information that's relevant. And so what we try and do is present very carefully and make sure when we say something that yes, it can help with this.
I think that there's a plain language ISO standard now, and it's something like simpler English relates to three or four of the 60 parts of the standard. People say, "Well, that's only three or four parts, and this is not all that useful." It's like, no, it's three or four parts that the AI can do for you! That's a little bit less work. Of course, it's not transforming the job, and it's kind of the same in readability.
We've got to be really precise about what we're saying. So the history here is readability scores, and the most famous one is called the Flesch-Kincaid index, but there's lots of different scores, and these use things like syllable count and number of words in a sentence. They use statistical measures to determine how simple, how readable language is. And those are really important because if you are a marketer and you're trying to produce something that your audience will understand, it matters.
More importantly, if you're doing safety documentation and you need to make sure that the people who are reading this documentation truly get it, it's got to be clear.
So a score is kind of crucial to that. But as soon as you have a score that's based on syllable count, you have people critiquing it because guess what? It's not gonna do very good.
You can write garbage text that's all, you know, one-syllable words — it doesn't make that much sense. It gets an amazing score. Just purely manipulate the score — doesn't actually have any meaning in terms of will the audience understand it. It's just a score; it doesn't know.
And AI is so interesting here because you can take out the problem. The problem is that people write for the score. So you get this incredibly qualified language professional, knows their stuff, really gets the audience, and they have to get to a certain score. So they're playing around with words to try and make, "Well, if I turn this and make this less natural, but it will bring it down to a sixth-grade reading level. And if I do this, this might be not good at all, but I've got to get this down to second-grade reading level, so I've got to do it."
Whereas let's just take the scores out; let's give those to the AI, and that's what DraftSmith's prompts are doing. So scores are still a little part of it, but the AI is doing the best it can to hit the level that a score would imply, but it's not going to cheat. What it's going to do is it's going to give you a sentence that makes sense, and then a person can look at it and apply their knowledge and their knowledge of the audience without ever considering a number or a score and just go "Oh, is this good? Will the audience understand it?"
So instead of their brain being on "got to get this to second-grade reading level, got to reduce number of syllables," their brain is on "will the audience understand? Is this good English? Has an important point of nuance been lost by the AI?" which, of course, it could do. And I just think that's so much better workflow. So I'm really, I'm excited about the possibility that can make a real difference there and get away from these silly scores whilst recognizing they have some value, but get away from writing. Having the AI do stuff and people judge it, which is what people are really good at.
Mignon:
Does it give you multiple options for its rewrites? Like, I can imagine in a situation like that, you might want to look at four or five different possibilities that all hit the right reading score and sort of pick the one that works best.
Daniel:
Precisely. That's exactly what you want the person to be thinking. So we've built that in two ways. You can set the screen up to have three suggesters.
So you could have second-grade reading level next to sixth-grade reading level next to eighth-grade reading level and just flick between the two. Because actually, it might be that the one for sixth-grade reading level is better for both, or you don't know until you see the suggestion.
And the other one is just to tell the AI to go try again. Because it produces a different answer every time. If you don't like its answer, you click refresh, or we have a design so you swipe to refresh, and it will come up with something different very, very quickly. So you can go in, make your own edits, you can ask it to do it again, you can see different comparisons. You get the person; in each one of those, the person is thinking the right way; they're not thinking about a score.
Mignon:
Yeah, and I think that's key, like, it uses Track Changes, and also you can go in and make your own changes if you want to. If you think, "Oh, this is close, but actually, it gave me an idea, and it would be better if I just switched these things around."
Daniel:
Which is exactly how a lot of professionals are using it. It's not that the AI is doing the thing; it makes the thought process a little easier, especially if you're tired. People don't talk about that so much, but if you're tired and expect to produce your very best work, having that backup of "here's a suggestion" gives that option and provides a check that's just so useful.
Mignon:
That's so true. Fifth hour is so much harder than editing in the first half hour. Yeah.
And your company put out a little, I don't know, it's not a book, but it's almost like a pamphlet about editors taking breaks and having a good lifestyle. Do you want to talk a little bit about that thing you just put out?
Daniel:
Sure. I mean, it was supposed to be a pamphlet. is we found a topic that resonated so much it became, I don't know, it's 30, 40 [pages] … we had 15 contributors. I was expecting when I had this idea to write to people, and a few would say, "Yeah, okay, I could do that."
But the idea it came out of Draftsmith version 1 — we noticed that people gave the exact feedback that I just said, which is this really helps me when I'm tired. I don't need it so much right at the beginning of the day, but when I'm tired, this is a big help, and someone says that you ignore it. Twelve people say it, you start paying attention, and I started asking why does this matter so much? Why does this mean so much to you? And it's what you said. Editors, writers, translators would say the same thing, which is that of course, you've been doing this all day, and all these professions sell themselves on achieving excellence all the time, and that's so completely not possible.
So when I started asking people "Would you like to contribute to this e-book?" the response I got was emphatic, and 15 contributors came in, helped out, some really big names in the industry. And now we continue to, I've been asked to present this, and now I have a slide which is something like, what is the difference between a pilot flying across the Atlantic and a trapeze artist at a circus, and an editor doing the final review of a proposal? And the answer is, well, the pilot flying across the Atlantic has a copilot, and the trapeze artist has a safety net, and the editor hasn't got any of those things.
And that's the job, and I know I'm exaggerating that, and I'm stylizing, and I'm taking it a little too far with that example, but you also get the idea it's like actually that's really, really difficult, and there is a lot at stake on a lot of these documents, and the standard is so high, and guess what? Everyone has ups and downs during the day. So that's why 15 people took part, they all had different strategies for how to handle it, and some of it is "Okay, get a good night's sleep," and some of it is "Be careful of what you eat and what time you're doing certain tasks. Straight after lunch … probably not the best moment for the deepest cognitive thinking that you're going to do during the day." Some of that's obvious, but no one had really written about that before because no one wants to say, "Oh no I'm not perfect." People are selling in the profession based on always being at their best and it's not possible.
So by us doing it, we're not editors, we're a little bit on the outside and then getting people who are known to say, I hope we start a discussion that others will then be able to run with.
And people can just be a bit more honest about the fact that of course there are ups and downs during the day. And of course there's a lot at stake, but you need a strategy to handle that.
Mignon:
And that's a free download, right? Is that at perfectit.com?
Daniel:
It's at draftsmith.ai And the book is called, I always get this the wrong way around, it is "The Pressure of Perfection." Have I got that the right way around? No, I always get it wrong, it's "Perfection Under Pressure."
Mignon:
Ah, wonderful.
Mignon:
Well, let's wrap up with some fun. Let's hear your book recommendations.
Daniel:
So I think the best thing about my book recommendations is kind of something that you do, which is it's how I choose. And I love your strategy.
You get people on the podcast, people you respect, friends, to give you recommendations.
Got this list I want to go back into your archive. I want to get all those recommendations, or I hope maybe you could summarize them somewhere. That's such a brilliant resource you've created.
Mignon:
I've been thinking about it actually.
Daniel:
So my own trick is I love reading, and I get so little time for it. It's just all I like doing. Holidays, I get a good book, and so I have to, with so little time for it, I have to make each choice count. Every couple of years, I will post on social media something like, "My favorite genre is modern classic. I want to read the books that people are talking about now and will still be talking about in five or ten years." And like you, I have lots of friends who are language professionals — lots of editors — and their recommendations are amazing. You'll see, you know, once every couple of years in my feed, you'll see these incredible lists of, like, "Oh, yeah, all these books!" and that will keep me busy for a long time, so things that come out of there.
I loved Jesmyn Ward's "Sing, Unburied, Sing." It's a magical realism, but set in the U.S. South and sort of dealing with issues of the legacy of the Civil War and Reconstruction, but it suits that genre so well, and she's brilliant.
I love "Lessons in Chemistry." That was another one of those recommendations from friends and colleagues. That is Bonnie Garmus. And the concept there was, it's the 1950s in the U.S., and it's been turned into a TV show, so people may know it, but people gave me those recommendations before it became a TV show. I don't read that much, and I get to sound like that person who knows so much more. "Oh yeah, it's a TV show; I read the book." But no, it's a chemistry professor who loses her job but then is offered a chance to teach cooking on TV, and she turns it into chemistry lessons. I think it's 1950s or 1960s USA.
Mignon:
Oh, how fun.
Daniel:
It's absolutely superb. The TV show is good too, but the book is better.
And third one, I would go with Anthony Doerr. It's "Cloud Cuckoo Land." So there's no way to describe this book without sounding incredibly pretentious, but it is three separate stories, and the stories are kind of interwoven and overlap and have common themes. The theme at the end is just a novel that is so in love with storytelling and the power of storytelling. And as I say that, I know it sounds pretentious, but I promise you it is a great read and really, really good.
Mignon:
Now, is he the same one who wrote "All the Light We Cannot See"?
Daniel:
I think it's the follow-up.
Mignon:
Ah, wonderful. Yeah, oh, those sound great. I can't wait to add them to my to-be-read list, and yeah, it's getting really long because people give me, you know, three books every week, so obviously I'm not reading them all because I don't have that much time either. But I've been picking and choosing from guest recommendations, and every one has been fabulous, so yeah, I'm very excited about this list of great books I'm getting, and I hope the listeners are too.
Daniel:
Yeah, I hope you can find some way to compile it. Maybe an AI could compile that list going back over your archive of all the different things that you've been recommended — the ones that come up the most. So it plays straight into how I want to choose my books.
Mignon:
Yeah, I've been throwing them into a Google Doc as they go. So I am planning on putting them out somehow. I'm not sure how. Maybe as a bonus for people or something.
Daniel:
Oh, I look forward to that.
Mignon:
Yeah, well, Daniel Heuman from Intelligent Editing, your new product is DraftSmith. Your old product is PerfectIt, which people like that too. Thank you so much for being here.
Daniel:
Thank you so much. A real pleasure being here. Fantastic, thank you so much.
Mignon:
You bet.