So What? Examining AI Writing Styles of Different AI Systems

Well, hey everyone, happy Thursday. Welcome to So What the Marketing Analytics and Insights Live Show. I'm Katie, joined by Chris and John. Howdy, fellas. Hello.

Uh this week we are talking about examining AI writing styles of different AI systems. It's one of the things I said to Chris, because Chris, you've been sort of sharing with me and educating me about like the tells of what each system has. And now I can't unsee it. And so it's like it's once it becomes apparent, it becomes very apparent. So we're actually doing a little bit of the reverse today.

So we're gonna go over a little bit of like what the different systems default writing styles are, but we're also gonna start to almost reverse engineer it and see. I think Chris, what we're doing today is we're going to see which system writes the most like any given human. And I think that human is me. You are correct. That human is you, but it's not the present day Katie.

It is 2022 Katie. It was a little less polished. Um, I mean, here's the thing we all start somewhere, and I was never asked prior to this role to write on a regular basis and give my opinion on anything. So that was sort of a newer thing for me. And I remember when we first started the newsletter, Chris, I didn't have a voice in it at all.

Um, and that wasn't because I didn't deserve one, it was because I wasn't ready. I didn't feel ready to um be doing that much sharing and thinking. And now I'm definitely comfortable with it. So I would imagine that in 2022, prior to any kind of generative AI assisted, you know, tools or rewriting. John, you were editing my blogs for me.

Yeah, yeah. We were doing a lot of the voice, uh, you know, getting the voice tighter and getting things. It's it's funny that it's making me laugh. I'm like 2022. I was probably a better writer then.

I think I've gotten weak and rusty over the years. So um so where should we get started, Chris? Well, let's talk about the writing styles first. And as this is a completely shameless plug. Um Chris and I are hitting buttons at the same time.

Exactly, we're all hitting buttons at the same time. Uh, a lot of the technologies from today's show are in our new AI for writers course, which you can find at TrustInsideSty AI for writers. One of those, it's not a technology, but one of those things is uh in the course is a 16 page glossary of all the different nerds speak. Some of the things that AI does, uh, for example, like anaphora. Anaphora is defined as uh when you repeat the same word or phrase at the start of two more sentences in a row, like we build, we ship, we iterate, right?

That's an example of anaphora. Um, and there's about seven pages of the things that AI tends to overuse because it doesn't have an understanding of style. It understands probability, and it says, hey, these are all the things that great human writers do, and like that amateur, like that that guy first day on the job at an avid video editing station, like, oh, what are they all these transitions do? And you see that like that amateur video with every single transition overused, you're like, no one can watch this man. And star wipe and star wipe and star wipe.

Yep. And then it swirls. Exactly. Well, and the other thing, you know, so it doesn't understand style, and I think this is probably where you're going. It also doesn't understand frequency.

So if it's like, hey, this thing works, let me just keep doing it over and over and over again. And that's something that I find to be incredibly annoying. Even if I'm just looking at like something it's prepared for me to look at internally, not even like you know, a public facing thing, I'm like, I can't, this is this is terrible. So part of understanding AI writing is to know what those tells are and how much they overuse them. And then the second part of it is understanding how you can measure writing, because writing is as much science as it is art.

And when we're working with machines, machines don't understand the art because they don't have emotions, they don't have lived experiences, but they do understand science, they do understand mathematics if we give them the code, literal Python code in this case, to understand it. So for example, adjacent pronoun opener rate. Uh, how many times does the do you have subject pronouns that are adjacent to each other? So if you, if for example, if you're writing fiction, um, and you start every sentence in the in five paragraphs word she, like she did this and she did that, and they're like, uh that's that's a thing you can measure in text to understand, hey, this rate is you know low, high, and typically when you're doing text metrics, you're measuring in frequency per thousand words. So if you have 1500 words, you can measure, okay, how many times does the person use this construction?

And you use these metrics to evaluate a piece of writing to understand what is that style in a quantifiable way. None of these are good or bad, right? These are all just measures that you can apply that. If you create a mathematical fingerprint of someone's writing, you can understand, okay, here's how often this person uses contractions. So John and I probably write very differently.

Katie, uh you and I definitely write very differently, and these numbers are going to be different. Side note, what is up with Claude not using pronouns at all to start a sentence? So, for example, like if I was gonna write the sentence like we did it, it would start with we. Claude would be would just write a sentence, did it, and that would be the whole thing. Like, what is up with that?

It depends on the context of the task, but also it depends on the model. So, even within the Claude family, the models write differently, which is kind of weird to think about. Um, but they all have different training data sets from different periods of time, and they are they are post-trained on different kinds of tasks. So you'll a lot of people have said this. You will probably have noticed that AI models in the last year have uh unguided become worse writers.

They're less creative. They are, you know, less interesting to read. And that's because they're being trained more and more for agentic use for agents. Agents don't need flowery pros. Agents need very punchy do this, do that, do the next thing.

I did it, you know, uh I did this, did that, and so on and so forth. So you're seeing that heritage of make this tool do stuff instead of not make this tool think. Uh that's not what it does. Okay, that makes sense. And you're like on the upside, it that's kind of good news because we don't, despite having an AI for writers course, it's actually not teaching you how to use AI for writing.

We don't want you to use AI to write your stuff. We want you to use AI to be smart about how you're editing, how to make sure that your writing isn't going to get clocked as AI slop, which is still the worst phrase ever. Um so that so what we have, and this is in the course plug. Um there's a Python script called writing fingerprint that will take the mathematics here and give it a text of at least 500 words, more is better. We'll develop a fingerprint.

So let's take a look at the piece of text that I started with. This is from the one of the 2022 Inbox Insights. This is something Katie wrote getting started with a marketing mix model. And this was well before Chat GPT and things like that. Um, and she wrote it, it contains the five Ps exactly as you would expect because that's been around since 2020, late 2020.

And this was a pretty good issue. I mean, it was it was it was informative, it was useful, it tell it helped people understand how should I be thinking about marketing mixed models? Um, and so what we wanted to do is say, okay, well, what does it look like? What does Katie's writing style look like? So let's take a look at what Katie's writing style looks like.

Katie's writing style, make this full size here. This is 1500 words, and it says, uh the closest match to you uh uh in terms of AI models. We'll talk about that in a bit is Kimmy K3, which is a Chinese model. And then you get start getting some of the measures of this the style measures by family. Katie's median sentence like the by the way, one of the things that goes really wrong with writing style analysis, people use averages, averages are terrible to use anytime you are measuring writing, because if you are a creative talented writer, you're gonna have wildly different sentence lengths and things and wildly different usages.

So you want to use uh median at a minimum interquartile ranges are better. John, you look puzzled. No, I'm just thinking about it, you know, it's because it's an interesting thought of how you know, like you said, sentence size and structure are affected by what you're talking about in the story. You know, there's certain points where you're much more descriptive, and then there's other points where things get more active and interesting where you have to get a lot tighter and faster. And like you said, if you if you're averaging things, you're just boiling all that stuff away.

That's kind of a classic uh statistical mess of thinking that the averages are going to guide you when really it's actually bleeding all the good stuff out of your your you know your project. Yep. So if we look at some of the com a lot of these uh terms you can look at in the glossary itself to understand what they mean. One that the new thing is funny is Katie's use of M dashes uh in 2022 writing is zero for 1000 words. That was not a piece of punctuation that she ever used.

Voice contractions, very high. That's a very high number uh number there. Median dependence distances 2.86, which is means that the the a word and the word it modifies are very close together within like two words of each other. AI has a tendency to not do that, by the way. AI has a tendency to have a very large median dis dependency distance where it'll be saying something and then it has a whole bunch of filler before it finally gets to the point.

And then the models that KD writes most like. So we do the filling first. Yes. I want to add something to that, which is interesting because I do use M dashes now. In 20, so I am, you know, I've had no like training on writing.

So just like my typing, I'm totally self-taught. I didn't know what an M-dash was, and I didn't know how to use it correctly until it became such a hot topic. And our good friend Ann Handley um was really sort of talking about it in terms of, and this is why we use it, and this is how you use it. And so I feel like between 2022 and now I've learned a lot about writing. And so I do use them now, but I still use them sparingly because they're not meant to, you know, drive the whole narrative forward.

Um, if anything, I probably use parentheses in the middle of a sentence to add additional bit of extra context more than I do m-dashes because I'm Gen X and that's what we do. We over explain ourselves. We have the asides for everything. Yes. Um so now in order to test this, to test how does how do different models write, we have to have them do as close to an apples to apples comparison as possible, but blinded.

What that means is they can't see each other's work and they can't see Katie's original work. So what we end up doing to do the testing part is uh I say, and at least if I can have the let me pull up the prompt here real quick. The the prompt for this is to is I want you to write a blog post. Ah, here it is. Original prompt.

Let's bring this up. Uh let's make that much bigger. 24. Um, write a 1500 word newsletter article about the topic getting started with marketing mixed models. Cover these points.

Introduction, purpose, people, process, platform, performance, 5P framework for trust insights, and conclusions, seven parts. The article can be plus or minus 50 words. Use the word count, the WCCLI installed on the system to accurately count words. Return the article in markdown format. The tone should be authentic, approachable, and written for a senior marketing manager or marketing director to understand marketing mixed models, why they might consider them and how to get started.

Do not use any web search tools. Really important because we don't want it Googling and accidentally coming across Katie's original work. If we're trying to do a pure writing style test, use no other tools, right? So uh and do not read any of the documents, only use the word count tool, write only from your own knowledge, and then write the output. And so what I did was uh I went through every model that we had access to.

So we have Claude Opus, Claude Sonnet, Microsoft Copilot, Deep Seek Pro, Deep Seek Flash, Gemini Flash, the brand new one, uh, GLM5. And they each wrote their own version of this. So let's go ahead and take a look just for fun at Gemini's, and we'll make this bigger. And Gemini says, if you've spent recent quarters reviewing your marketing dashboards or growing skepticism, you are not alone between Apple's app tracking transparency framework, the steady phase out of third-party cookies, and expanding global privacy laws and so on and so forth. So Gemini, by the way, Gemini has a uh it's very interesting when you start looking at the writing styles of the different models, they you can see they are distinct.

One of the things that annoys me to no end is that people talk about you know AI writing. There is no such thing. There, every model has its own writing style. Well, because every model was trained on a different set of documents for lack of better term, all written by humans. And so it's basically said, okay, based on all of this stuff, what is the best version of writing that I can then say this is what it is, and you're gonna get a different result every time.

Yep. Let's take a look at our our lovely friend Claude that we all know and can recognize. Um you've probably had a version of this conversation. Someone asks what your paid social media spend actually delivered. You pull up the dashboard, it reports the number, then someone asks why that number doesn't match analytics tool or what finance season review.

The room goes quiet, and the meeting moves on. Claude, you can see already see there's triadic rhythm in there, and then there's also buy calling. That gap is not a reporting failure, it is the predictable result of a measurement approach, though. So there's you it's not this, it's that. So you can see um one of the interesting things I think is so funny is as anthropic has made significant inroads in business, particularly at the enterprise level.

More and more content you read online reads exactly like Claude. For me, there's certain words that Claude uses that gives it gives it away. And a lot of it is like the shape of. So like the shape of this is your next one. How things land.

Yeah. And I'm like, why? Why is suddenly everything have to have a shape? I don't get it. Uh-huh.

And so you'll see, you can see even in this text that you know, Claude has a very distinct flavor. Let's take a look at the brand new version of uh chat GPT, GPT 5.6, just to get a sense of how it writes. And even though it's not an outright, it's not this is that, it actually starts out with it, it's not it, this, it's that. Um, it it chat GPT's writing is better than previous versions that of the GPT-5 family, but worse than the GPT-4 family, which a lot of people appreciate for its more creative writing. Again, it's at the whole agentic thing.

But we can see in here, begin with the decision, not the model, right? There's the by calling frequency, even in there. So our next step is for us to take each of these models and fingerprint them. So let's take a look at GPT 5.6. We'll look at its fingerprint.

Here we are. So take a look at this. So it's got it's got shorter sentences, it's got much higher lexical diversity. Um, it's got a little bit higher median distance descent dependency, but it's got a lot of repetition. And it's got a crap ton of punctuation.

So if I put this side by side with Katie's, you can see there's there are some very, very, very different looks between these two. Right? Katie uses a lot of contractions. It uses none. Katie uses first person.

It uses none. Right. What you were saying, Katie. It's like, hey, this thing appears to not use pronouns at all. Um bizarre.

And so we go through and we fingerprint every single one of these models as the first part to understand which model is going to write like Katie. I've already I've already spoiled that, but that's well, that's fine. The second half of this is inside the um the course contents, there is a uh a YAML file, yet another markup language that contains all of the rules that we have about what you should and shouldn't do. You should not you overuse negative parallelism. You can use negative parallelism like once per thousand words.

It's not this, it's that, but like not every paragraph. And so if we go through and we look at these models and we and we say, okay, how does this model do on that checklist? So GPT 5.6 gets a 62. It says, hey, your vocabulary is good, your structure is not great, your shapes, you're way over using a lot of those constructions. Your your voice match is good, your stance is not, and this goes back to that glossary of all the things that it could and it could and could not do.

And it it goes through and it audits now. The audit model for this uses Alibaba Quen because one of the things you can't do is you can't have a model audit itself because it will be biased towards its own writing. So in any test like this, you have to have a referee model that's a that does not do the testing, that does not do the writing, it just is there to test. It's the same concept of uh peer development, is you don't want to have a software engineer testing their own stuff because they're too close to it, and you don't want to have a writer editing their own stuff, they're too close to it. It's the exact same idea.

Like, you know, have have like a neutral third party looking at it versus testing its own. Like if I look at my own writing, I'm like, I don't see the issue. And then someone like John's like, did you miss the five glaring, you know, lack of punctuation and all of your misspellings? The heck, Katie. It's just human blindness, you know.

It's we literally don't see the stuff because we've gone over it so many times, it's completely invisible. And it sounds like the same is true of models. So if we want to sort of like equate it to that, like it's model blindness where they can't see their own stuff clearly. Exactly. And they will give preference to their own language, of course, even in the scoring rubric.

Now, this is also really important. Um just because something gets flagged does not mean it's bad. If we look at Katie on the same human writing evaluation, I don't write like a human. You you get a 71. There's some things you need to rework.

Particularly, so you you um are are fine on the shapes, right? But you don't use a lot of the AI constructions, but there's a lot of other things that it finds, like, hey, the last paragraph adds no new fact. Um sentence paragraph, load bearing claims with a named source, um, isocolon uniform staccato. So just because something is flagged doesn't mean it's bad. Because as we as we started this show, we're kind of we're talking about frequency.

You don't want to overuse one thing, but some use of things is fine. So no one is going to come out perfect on this kind of analysis, nor is a higher score necessarily better than the lower one. But this analysis helps us understand what a human's writing look like. Because again, if we put these things side by side, um you can see in some areas, here's GPT 5.6 and Katie, there's some pretty stark differences in in those color bands as to how they write. There's definitely machine uh influence there.

Thank goodness. Now, from all of that, what we want to do is then say, okay, if we were to convert all these scores and all these numbers, the fingerprint data plus the human writing data and convert them into scores, we could do um similarity matching to say how similar is one thing to another. So let's take a look at the first of those. On similarity, if we look at just styles stylistically, make this a little bit bigger here. For Katie, Katie, the machine that writes the closest to her is um Moonshot AI's Kimmy K3, the K3 model that just came out about two months ago.

Um in second place is uh GLM 5.3 flash, in third place is Deep Seek uh, and then in fourth place is Claude Sonnet 5. Interestingly, Claude and Katie don't write very similarly compared to these other models. Which I could, I mean, I could have told you that. Uh, when I see outputs, I look at that, I'm like, and again, you know, I'm using like summarizations for internal documentation, you know, putting together uh, you know, data and information. And I'm looking at it, I'm like, a lot of times my thing is like, well, you can tell I didn't write this.

The one that really stands out in here, Gemini. When you look at so this also compares models to other models, right? So you can see Gemini's got a big old red in every direction. Why? Gemini doesn't write like any of these.

Like they all write, it writes very differently from everyone. And it is kind of a very strange outlier that it is so far different from everybody else. And again, that may be because the way Google is tuning it and training it, but it is kind of an oddity, which also means if you if you're gonna use just native AI and you want something that's probably going to sound different, Gemini Flash would be the one used because it very clearly does not write like any of the others. That's interesting. The other one that's interesting, the other thing that's is uh somewhat interesting, is when you look at some of the the hot spots in this table, um a lot of models, you know, there's a decent amount of green in the Claude Opus column.

And it's because uh, as Anthropic has famously been very loud about, other AI companies train on Claude's outputs. And so they have uh post-tune, and so a model like Minimax um tends to write like Claude because it's been trained on on what's called model distillation. Which is like this whole thing is very eye-opening. Yes. So the second thing is in terms of the QA violations, and and in this case, we just mean you know, not writing or writing like a machine.

So Katie and Claude Sonnet do the same things, right? So if Katie uses triadic rhythm, sonnet tends to use triadic rhythm and stuff like that. So they have the same quote tells, uh, closest, um, followed by Opus. No, I'm yes, follow by Opus and then followed by GLM 5.3 flash. So those those three models, Katie has a tendency to write similarly to them.

Now, this doesn't, you know, this doesn't mean bad or good, but just means that if we're choosing a model, this would be the second factor. So if we blend all these things together, the model that if Katie was said, I want to use a model that writes as close to me as possible so that I have as little rewriting to do, the model Katie should use is Kimi K3. Which for those watching is not a model I even have the faintest idea how to get started with. So I'm definitely not using it. And if you look at my Claude system, I'm primarily using Opus.

I probably should be using Sonnet more, but I really don't. I use Sonnet for um some schedule tasks. So yeah, that's if you see it, it's not AI, it's me. Is Kimmy is that one of the local models? No, uh, it's uh it's a cloud-based model.

So uh we use it uh through a company called Deep Info, which is a Palo Alto based company. Uh it'd be you would need like $50,000 worth of hardware to run it, you know, effectively on your own. And Katie said I don't have that kind of budget to work with. No. So yeah, I can I can confidently say we won't be using Kimmy K3 anytime soon.

Not locally, not locally. No. Um, it is a very expensive model. It's one of the more expensive open weights models to use. But yeah, to your point, Katie, if you look, Opus and you don't write similarly uh as much as Sonnet.

Sonnet is closer to how you actually write. And you know, it's this a lot of this started, I guess, over a month ago when we had some negative feedback on the uh newsletter. And, you know, the feedback was that, you know, well, this is just AI slop. And I was like, okay, I mean it's me, so maybe I'm AI slop. I don't know.

But it looks like, you know, depending on what you're looking at, there's a version of my writing that could be considered AI. But to your point, AI is trained on human writing. And so we're not saying Katie writes like these models, we're saying these models are the closest to writing like me in my writing style, which depending on the day might change. Yes. I would imagine if you used a different article or a different point in time, these results would be different.

Yes, exactly. So for example, if I took uh this week's uh a newsletter from you know in the last month or two where you have used Claude Opus to refine your original writing, it's gonna be dead on Claude Opus because Claude will restructure the writing, and even if you give it original content, it still restructures it as AI and it comes out sounding more like AI because that's the nature of the probability engine. There's actually a hack for that. Um and I I'll I'll tell it to you now. I was gonna post about on about LinkedIn later.

But don't tell an A actual, you know, edit this document for me. You tell it, audit this document for me and tell me the smallest effective edits that you'll make and make a list of them. And it will do an audit and it will give you short corrections, and then you say implement the smallest edits only. And so it will do much less rewriting of your stuff. So I actually don't use uh Opus or I don't use Claude to edit my writing.

I do all of my writing myself. I mean editing myself and writing myself in Hemingway. And I also don't use their AI feature. I've I've really tried to make, and then you know, nobody cares, but I'm sort of making sure it goes on record. I'm really trying to make a point to do original human reign, because that is not a muscle that I want to lose.

I want to make sure that I, as the human, can still credibly write something. Um, you know, when I use Opus or when I use Sonnet, it's usually to create something internal to review. Um, you know, of like, hey, what was this call about? Or what are the five things that we need to do for this client? So it's more that kind of writing, and I don't really care, you know, if it's doing the triadic rhythm or whatever, but anything publicly facing, um, I'm doing, I might have Claude help me outline the structure so that I get the narrative clear so it goes from it doesn't just sort of wander off in my stream of consciousness, which is where my first drafts usually go, and there's no point.

So I help it, I use it to help me make sure there's a point to the story, and then I write everything in Hemingway, which is just another editing app. Um, and I guess that's a question is like if I do that in Hemingway, is it still gonna have that like AI written stamp on it? So all the AI detection tools, we covered this on the AI detectors episode. Uh, look at perplexity and burstiness. So if it is making recommendations that you take that, adhere to that structure, because to your point, there are they're trained on human writing.

If you make your language higher probability language, then yes, you'll start to verge into that territory because AI is never, despite what the AI detectors love to say, it's never just you know yes or no. It's a it's a spectrum of you know what percentage of words here have a lower perplexity or a higher perplexity. So it's I mean, it's a lot to think about, it's a lot of information. And I think, you know, it there's definitely times where it is appropriate to have AI generate something for you in terms of the, you know, and I and what I mean is like, you know, maybe a landing page or maybe you know, it can take your brand voice and do something like that, versus, you know, if you're doing original thinking and a thought piece and an opinion piece, AI has no place in that because you're the human, you have opinions. Not, I mean, AI is gonna pretend it has opinions, but it really doesn't care.

Yep. So what do you do with this information? Um, first things first, when you build that fingerprint document, that fingerprint document is then something you can give to a machine and say, okay, machine, these are the text metrics that if you're going to be doing any kind of editing, this is what you must adhere to. You give it a deterministic list and the Python code that goes with it. So it when it does a draft, it will write it and then it will load up the data file for your writing and say, okay, now I'm going to analyze the thing I wrote.

Now I'm going to add, I have my gold standard. So if I go into the let's go into the Katie uh report here, go into where's my fingerprints? There they are. Uh let's find Katie's fingerprint. Writing.

Yeah. Fingerprinting. This is what comes out. Right. So it's just a long data set of exactly how Katie writes.

It would then you hand this to a model and you say, okay, this is what you must adhere to. You have to produce outputs that meet these measures, these concrete metrics. And it has a script to that and it has to iterate over and over until it hits those numbers. Uh, I think in the writing toolkit we distribute with a course, uh, it's you have to be within one standard deviation of all the different measures. And what you will see is when you do that, the machines they they do churn a lot more tokens, but they come out sounding a lot more like you because instead of saying write like me, which is unhelpful, it says you must do these things, and here's the measure to tell you whether you succeeded or not.

One of the things that is so difficult about people writing prompts is that if you say, Hey, you know, write in my style, you've basically violated the 5P framework by Trust Insights. You've given it no performance measures, and with no performance measure, it doesn't know how it succeeded or not, it has no idea what done looks like. Well, and I think that that's I mean, and that's something you've been teaching for a long time is that performance measure, because to your point, AI is just gonna keep going. And if you if this is true of requirements building, this is true of any kind of you know vibe coding that you're trying to do, or you know, a report that you're building, you have to tell AI when it's done. Yep.

So let's take a look here at some of Katie's measures. If you were to say, hey, machine, here's how to write like Katie. Let's open up the meta. Um Katie's sentence length, right? The median 7.57 words, and they have the uh interquartile range that demonstrates and and for your sentence like this interesting, your sentence length does not vary much in in the first and third quartiles and stuff.

So you have a fairly steady sentence pace of your own. Um, your burstiness, uh, your media you don't have a ton of burstiness. So, like you don't write a 28-sentence paragraph and then have a one-sentence, you know, uh bullet and stuff. You you're pretty consistent. And so this how long is this?

This 530 line diagnostic is what the tool goes, oh okay. I understand now what it is that I understand how to write like Katie, and I can go back and check its work. I mean, that makes sense to me. I think when I was writing, when I was learning about writing, especially in college, I probably would have had more of what's called that burstiness. And that was what my creative writing teacher told me was really not helpful.

Um, and so I've tried to when I think about writing for business, I try to make it as concise and consistent as possible while still sharing my thoughts and my experience, you know, on all those things. It's hard, it's a tricky balance, and so it is interesting to see my writing in terms of math. Yep. The other thing that you pointed out very clearly, A, time matters. So your writing changes over time.

And B, you register and domain matters. So the way that you write or text to like a friend, right? Like you know, text our friend Brooke, I guarantee is not the same writing register as when you write a LinkedIn post or when you write a newsletter. There'll be commonalities because you're human, but you're going to use very a very different even rhythm, right? So like when you're texting somebody, it's a typically a much shorter rhythm thing.

So your sentence length, if there even are sentences, depending, I guess you know it depends on who you are and your age, um, will be shorter. And so if you were to use that as your benchmark and then try to cross-apply that benchmark to something else, it's not going to go well. Uh this may come as a shock to nobody. Um, anyone who has the pleasure of texting with me, it is going to be in full complete sentences with punctuation and grammar. And it is not meant to be aggressive.

I just, you know, I'm gonna put a period at the end of every sentence because that's what goes there. Well, it's funny because like on uh Apple devices, um, if you double space, it automatically adds punctuation for you. And then your kids are like, Why are you mad at me? Put a player at the end of the sentence. Like, I didn't do that.

The machine did it automatically. I don't like, and this is a whole I don't want to go down a rabbit hole, but like it's a whole nonsense thing. Like I put a period at the end of a sentence, therefore I'm mad at you. Like, no, I'm I put a period at the end of a sentence because a period goes at the end of a sentence. I text the same.

I believe if you looked at my texts, and I'm not going to show you. If you looked at my texts, they are probably closer to how I write on LinkedIn than a lot of people because I just I just write how I write. Yeah, mine are definitely not the same. Um and the domain matters. Um the other thing that is an issue is a lot of companies have like a brand style guide, a writing style guide for the brand.

And it tends to be pretty bland, which means that if you give that to a machine by itself, you're gonna get slop. It's gonna be slop in the general style of your brand, but it's gonna be absolute slop because the machine won't understand the the nuance of how do I write, you know, as acme corporation in a voice that represents acquirations. We are authoritative and warm and professional and friendly. Like that's like literally every every bad writing product, and it will never generate anything that's worth reading. So you have to also incorporate some of your own human writing into a brand style guide, otherwise, you're gonna create slots.

Yeah, that is definitely something the the corporate brand is so tricky, and a lot of people are going to look at to your point, the channel and the context matter. So a lot of people will look at the consumer brands on social media that are, you know, quote unquote unhinged, and say, Oh, that's our corporate brand. It's like, that's not your corporate brand. That's what you are on Facebook, that's what you are on Threads. That's not what you are on a product sheet that lives on your website.

Yeah, exactly. Think about it like shows. So the the Daily Show with Jon Stewart had a very different voice and register than the Daily Show with Trevor Noah. Same corporation, right? Same corporate entity, the Daily Show.

Wildly different language. John, what do you think your writing style would be? Uh do you use a lot of m-dashes or uh do you use a lot of punctuation? Um yeah, for the most part, I you know, I did have the punctuation rules beaten into me. Um so you know, that's how that works.

The big thing with my writing is, you know, I'm auditorially driven. And so I'm really just taking the way I talk and putting it down on paper, right? I'm not actually creating text, I'm just talking at the machine so that uh my writing style is very closely aligned to when people say their voice. I mean, it is literally my voice, that is what it is coming across in the paper. So um, and yeah, and you know, just a sprinkle of New England sarcasm and anger to spice it up as we go.

I think that's also part of the uh important thing that yeah, AI cannot really match that. The AI Boston accent is just not there. It's not there, no. No. So, in terms of next steps, what you should do, uh, do take the AF for Writers course.

You can get it at Trust Inzest.aslash AF Writers. Um, and fingerprint the models following the process we outlined on today's show with the tools that are in the the course. Um, fingerprint the models on a task that you do that you would want writing for. Fingerprint your own writing, apply those rules, and then see which model behaves the most like you, so that if you have access to it, even if you only have like three, maybe you'll have Copilot, you have Claude, and you have Chat GPT. If you have just those three, figure out which one is closest to you in terms of your style, so that you can say, okay, if I'm doing writing, I'm going to use the model that's closest to me to start with so that I there's less work to clean it up than starting with the model.

It's furthest away from how you write, and you're like, oh my god, I gotta rewrite the whole thing. I should have just done it myself. Yeah. Which is generally my process. It is not time effective at all.

Yeah. Um, but yeah, that's that's how the examining the writing styles of the different systems. And as we saw, some of them, like Gemini 3.5 3.8 flash, write wildly different than everyone else. Um experiment, see which ones take instructions better. As more models become agentic, agent focused, you're going to spend more time giving more specific instructions to them and giving them more materials like these fingerprints because they'll know what to do with them and they'll be able to implement them more smartly.

But by their own natures, they will not write as well naively. Um so it will require more effort on your part. Any final parting thoughts? Don't outsource your critical thinking. Don't outsource your writing.

If you can. But if you have to, for whatever reason, take our course and figure out which model is going to be best suited. Because the last thing you want is there's someone to lose their trust in you because they're just assuming everything is written by AI. Exactly. Alright, folks, we will thanks for watching today.

We will see you all on the next one.ai slash TI podcast. At our weekly email newsletter at TrustInsights.ai slash newsletter. Got questions about what you saw in today's episode? Join our free Analytics for Marketers Slack group at TrustInsights.ai slash analytics for marketers. See you next time.