How to survive the four horsemen of the AIpocalypse (webinar)

Confront the chaos that generative AI can unleash on your content and discover how to regain control. In this practical session, Torsten Machert (Congree Language Technologies) and Sarah O’Keefe (Scriptorium Publishing) revealed the four biggest threats that undermine quality when you rely on GenAI for content creation.

Sarah O’Keefe: It is way, way cheaper to build out the maturity of your content, to do the terminology work, to do the structure work, to do the metadata work, label everything, give it categories, give it classifications ahead of time than it is to try and remediate the content after the fact, after it’s been processed, after it’s been ingested into the AI and then spit back out. My fear right now is that we’re seeing a lot of, “Ingest everything, spit it back out, then consider how to fix it.”

Resources

• Congree Language Technologies
• The true measure of success for AI initiatives (Scriptorium podcast)
• Want to stay updated on key industry insights? Subscribe to our newsletter!

LinkedIn

• Sarah O’Keefe (Scriptorium)
• Torsten Machert (Congree Language Technologies)
• Scott Abel (The Content Wrangler)

Transcript

Scott Abel: Hello. If you’re here for How to survive the four horsemen of the AIpocalypse, you’re in the right place. My name is Scott Abel, and I’ll be the host of today’s show. All right, without further ado, take it away, Torsten.

Torsten Machert: Yeah. So Scott is referring to four horsemen. It’s just the three of us here, but let’s see what we can do with it. So I need to talk about myself first just to set the scene. I’m going to show you a picture. And this is a picture of my desk.

Sarah O’Keefe: How did you get a picture of my desk?

TM: Is it yours? I’m sorry. 

SA: I’m sure that’s not mine.

TM: Yeah. So some people say you are a messy. I don’t call it a mess. I call it my desk is a high entropy or my desk is like a neural network. So some people might say, “No, you cannot find any information.” Of course I can because I know the systematics here. Of course, there’s some better way to organize information. This is something like this. There’s a systematic behind that. Of course, in our industry, we don’t have libraries anymore. We use something called content management system. By the way, Scott and Sarah, what is the most popular content management system? I know, it’s a very bad question. It’s called file system. File system.

SO: Yeah. I usually say it’s Excel, but… Yeah.

TM: Yeah, Excel. Yeah, another one. That’s good.

SA: Either that or a desk drawer.

SO: A desk drawer. It’s right there on the left side of the screen.

SA: Exactly.

TM: So this is one way how we can organize a lot of information. We have different shelves for different topics. We could have fiction and physics and chemistry and linguistics and music and everything. And inside of those shelves, we have an order based by subcategories or by authors, et cetera. But I think, Sarah, you are the best person to talk about classification and taxonomy and tagging.

SO: Well, yes, and we will get to that in, I think, some detail. But basically the four horsemen, two of them are basically disorder, lack of organization, and not having a classification system and not understanding the structure of your content or not imposing structure on your content. So eventually, I’m going to start talking about lack of structure, lack of labeling, lack of semantics, and then separately taxonomy, metadata, and a classification system. Those are the two buckets that I’m focused on when I start talking about order and disorder and issues with AI. And interestingly, I think those are effectively one level up, not in a judgy way, but a level up above the things that Torsten is focused on today, which have to do more with linguistics with the text itself. So I’m interested in the containers of the text and, I mean, I’m interested in the text itself, but Torsten’s going to focus on the text itself, right?

TM: Yeah. You mentioned something, order or disorder, so we can call it differently. It’s called entropy. This is an example, again, of order and disorder is an explosion. So it’s an uncontained or uncontrollable event, and that’s part of our nature. So nature, everything tends to get to a higher entropy, so a higher degree of disorder. And this is an explosion where chemical or nuclear energy is being transformed into kinetic energy. So can we use that? Yes, we can contain this principle. And one application is this one, a combustion engine. And within the cylinder, we have an explosion, so we explode the fuel, and then the chemical engineer moves the piston down and does converting the chemical energy into kinetic energy. So what do we need to reduce the entropy is to do some work, and sometimes a lot of work.

But entropy is also applicable to information. All right? And this is the definition I have taken from Wikipedia. Or we say here in Germany, a good artist copy and great artists steal. So I have stolen this from Wikipedia. So you can read it later on. This is the definition for entropy, and as you can tell, it’s also applicable to information theory.

Let me show an example of this one. All right? So the question is, does this sentence have a high entropy or low entropy? So for us, those that understand English, you might think, no, that’s reasonable. I understand what this sentence is all about. So by the way, this is a sentence being used to test fonts. So let’s imagine this text was written using a different writing system, so like Cyrillic, Arabic, Hebrew, Japanese, Korean, or using hieroglyphs. I obviously refer to my handwriting. The question is, can we reduce the entropy of this sentence? Yes, we can, but it doesn’t make any sense. So this is the lowest possible entropy of this content, so we order all the characters alphabetically. That doesn’t make any sense because we have to deal with natural language.

I prepared some other samples. So this sentence. It looks good, but it could come up with some variations of this sentence, so like this. And this is something we don’t want. So we don’t want to get information or content from a GenAI with all those variations. There are minor differences because there’s a difference between hitting a button or smashing a button or touching a button or pushing a button. So why is this bad? So we want to create content correctly, obviously, and consistently, and we want to get information out of a GenAI correctly and consistently. And the content we create is nowadays not only being used to be published to customers and users and maintenance and REAP organizations, so we also want to feed this content into a large language model. So that’s why we need to take care of that and try to reduce the variants that we have within a text.

So there’s a difference in the semantics of hitting and smashing and pressing. I’d agree. So what is the problem for the readers, the users? So they might be asking, “So why were they using, or why was the author using different words here? What is the difference?” And even worse, the translators might be thinking, “Oh, the author, assuming there was only one, was using different words, so I need to find different words in my language as well.” And then again, the entropy increases the quality of the translation.

Another example. So the way I was talking about the button, so the button in a user interface of the software. The shirt has a white button. So this is in English. And in German, for instance, we don’t use the same words. We use different words. But here, I know I’m referring to something else. And in the user interface, the sentence basically provides the context and the other things that I can do with the button on my shirt, I can close it or don’t open the button. And we do not open the button in a user interface.

So this defines the context, and the context is key in a generative AI, because otherwise, a generative AI cannot create reasonable context. And there’s one term that you probably have heard, that’s hallucination. So a GenAI makes things up unless we provide additional information to reduce that. So how does a GenAI work? How does it create a neural network? So it analyzes each and every sentence word by word, goes back, and then creates something that’s called tensors. So tensors is a bit more than a vector. And this could look like this. This is a very simplified version of that.

So based on the sentences you have seen, so we can organize them, we can cluster them. Again, all the words along with the now and button, like smash and click and press and touch and hit, belong most likely to a user interface, to a button in a user interface, and all the other words belong more to clothes, to a shirt, et cetera. And it’s very important to understand that GenAI is not based on… Or it’s based on statistics, and more importantly, on probability.

SO: I just have to tell people it’s just math. I don’t know how accurate that is, but it’s just math. It’s math.

TM: No, I did not dare to use the word math because it sounds so scary to many people. Probability sounds better.

SO: Probability and statistics is better.

TM: Yes.

SO: Okay.

SA: Math teachers are cringing right now.

SO: Yeah.

TM: No, I didn’t mention that. I studied aircraft engineering [inaudible 00:17:47] studying math. Okay, I’m not going to say that. So now, so let’s assume we want to check our content. Because when we create content, it’s all about quality, and we want to achieve a couple of goals. Of course, we want to create content correctly, we want to create them consistently, but there are other goals. So improve readability, improve translatability.

So now, the idea could be, so why don’t we use GenAI to check the quality of our content? So we could send a document, whatever format it is, to GenAI and expect the GenAI to check our content. Does it work? Most likely not because there’s some important information missing. GenAI cannot know what our intention is. It cannot know what the quality criteria are. So we need to add some additional information, and that’s what we call style guide and terminology.

So we are here talking about what’s called a controlled language, and the controlled language is a subset of a natural language defining rules for spelling and grammar and style, and it also defines a dictionary, but we do not call it dictionary for a couple of reasons. I’m going to talk about them in a second. We call it terminology. So the GenAI has no clue how we want to create content. It cannot. The GenAI has no clue what our terminology is.

And if you take some maybe car manufacturers, so like here from Germany also, our customers like Mercedes and BMWs, they have the same components, but sometimes they call them differently, because it’s brand-specific, for a couple of reasons. So if you think of a cruise control system and you look into the owner’s guide of Company A and B and C, then you find completely different terms for the same thing. But the GenAI cannot know that. So why can the GenAI not know that? Because it has never seen, so seen in quotes, the information from those manufacturers, because they would never agree to use their information to feed or to train a large language model because the information is classified or they want to protect the IP and for many other reasons. So that’s a very bad idea to do it like this.

So what do we need to do? We need to find something allowing us to measure the quality of content. And I don’t know if you ever seen this formula. It’s the Flesch-Kincaid score. Really interesting. Flesch was an Austrian. So we already made in Germany very bad experience with an Austrian guy, and this is another one. So in the 1950s, he came up with this formula saying, “Oh, you can measure the quality of content. You can measure the readability.” I do not know how he came up with those numbers, but I give you some examples demonstrating why that cannot work.

So I measured the Flesch-Kincaid score using a couple of sentences. So the first of them are those that you’ve already seen. It’s rather high, and the third one is rather high. And then the fourth one, it’s a quote from one of my favorite American actors by the name of Groucho Marx. “Time flies like an arrow, fruit flies like a banana.” High readability. So it’s for a fourth grader. And then the last one, Lorem ipsum, it’s called blind text and it’s used in typography. So when you design the layout, you don’t have the actual content, so then you put something in like that. It looks a bit like Latin, but it’s not. But still, the readability is high. And this is rubbish, or garbage, depending where you live. So we need something else. So we need what I already mentioned, a corporate language, or I use a different word, controlled language. But any corporate language is a controlled language. But what defines how our corporate language, our controlled language looks like, and there are those three categories, at least those three you see here.

So let’s talk about the target audience. So we need to know what a target audience is. So we need to know what other skills, what is the native language, the cultural background, and everything you can see here. So that defines the first category, defining how our final corporate language looks like. 

Also, the information type. So there’s a crossover between information types and the target audiences, obviously, but we create content differently. So the content in a classical manual looks different than in a parts catalog or in marketing, a collateral or brochure. So sales creates content differently. Training creates content differently using different tools. So they use PowerPoint, where we have more phrases rather than complete sentences.

And the last category is the purpose. We have descriptive information, we have instructions, procedures, we create learning material, and we have translations. And when it comes to translations, we have two categories of translations. So human translations, so made by human being, and machine translation. And the outcome or the output of the translations or the quality of the translations is depending, again, on the source language. And nowadays, we want to feed our content into a large language model.

So many of our customers are interested in using GenAI, but they can’t because they’re not allowed. It’s company policy not to talk to a publicly available GenAI. So they have two options, not to use GenAI at all or to create their own large language model. So now, let’s assume they have the money to do that, because it costs a lot of money, so the hardware that’s needed to create a large language model is massive. So for a medium-sized large language model, it takes like 100,000 processing hours on a GPU. And a GPU costs like $10,000, so you can do the math.

So let’s assume we have the money and we have the hardware, but the amount of data that we have in any organization and any company, no matter how big it is, is small, small compared to the amount of data that was used by companies like OpenAI or Google or Meta to create their large language model. So they were using hundreds of thousands of books, millions of articles to train their neural network. So we can’t do that. And due to this small amount, really small amount of information that we have within an organization, so we need to make sure that the information we create is correct, it’s consistent, and it’s limited. So limited in terms of style and the language we use, and also terminology.

So let’s talk about the first ingredient, and that’s the dictionary or the terminology. I would like to give you a short introduction what we mean with terminology. So first of all, is terminology not? It’s not wordless. It’s not a glossary. And I’m going to explain what terminology is. I’m not sure if you know what this is. It’s called-

SO: I hope it’s cricket.

TM: It’s cricket. What’s that?

SO: I don’t like that.

TM: That’s what the name of this [inaudible 00:26:22].

SO: Well-

TM: It’s a cricket.

SO: Yeah, I think it’s a cricket, but it doesn’t look like the crickets-

TM: It’s a cricket.

SO: …we have here because it’s too small.

TM: It’s too small. Yeah, it’s a cricket. So the first lesson we learn, never begin with the word. It reminds me of one of my favorite songs from one of my favorite bands, Scottish musician, Led Zeppelin in Stairway to Heaven. There’s one line. Don’t ask me to sing it. It goes, “Sometimes words have two meanings, and sometimes they have even more.” So that’s why we cannot say, “This is cricket,” and we put it in our word list, because that’s not enough. This is just a word. This is the way how we call a thing. I’m going to explain later what we do with it. I’ll give you some other examples first.

By the way, I wanted to use another game here because in now three or four weeks’ time in Canada, the US, and Mexico, it’s the biggest sport event in the world, and it’s called the Football World Cup. Football. It’s not soccer, it’s football. So we would never call a game football where the players barely use their feet. It would make sense. So it’s football, Football World Cup, Canada, US, and Mexico, starting on June 11. But I was using cricket. It’s more neutral. So what’s that? It’s difficult to tell. It’s in London, it’s a bank. It’s a national bank. It’s a bank.

What’s that? It’s a beautiful lake on a river, but it’s a bank in English. So again, the same word having a completely different meaning. So terminology always starts with a definition, and we call it in terminology management a concept. So concept is something that we have in the real world. So we can see it, we can touch it, we can move it, we can operate it, we can do it, and you find a definition. And then we say, “If this definition applies, we call it like that. And if this definition applies, we call it like that.” So how would that look like in a terminology management system? I’m going to show you in a moment.

So why is this definition so important? So sometimes the definition has the lead. If you just mentioned the word, people, depending where they live, might be shocked or embarrassed. Scott, Sarah, do you have an idea what is this?

SO: Is that currywurst, and what is it doing on a fancy plate?

TM: Nope. It’s from your English cousins.

SO: Oh, bangers?

TM: And they call it… I did not make it up. That’s how they call it. So that’s why the definition is important. They call it spotted dick.

SO: Oh, spotted dick.

SA: Oh.

TM: Spotted dick.

SO: I’ve never seen a spotted dick before.

TM: Yeah. So nothing bad. That’s how they call it. It’s a dessert. So the definition applies whenever we have-

SO: It’s a dessert?

TM: Yes, most of the time. 

SO: Okay, England.

TM: Yeah. No, I’m not going to do this joke. A definition can be a text describing something. Could also be a picture or a drawing. So whenever this applies, we call it like this. So how would that look like in a terminology management system?

So let’s go to the second example I gave. Bank. So the word bank. And this is what I stole again from Merriam-Webster’s dictionary, the definitions for bank in American English. Maybe in British English as well, I did not double check it. So whenever those definitions apply, we call it bank. Then we have the other definition. That was there.

How does it look like in a terminology management system? This is the definition. Again, stolen from Merriam-Webster. And this is the order. If this definition applies, we call it bank. So why in green? Because this is another term in terminology management system. We say this is the preferred term. Or in other words, this is the only term we allow. And the red one is the bad boy. But in terminology management system, it’s not called bad boy. It’s called deprecated term. Don’t use it. It’s forbidden. Don’t use it.

And this is applicable to American English, obviously. So now, the terminology management contains also the translations, and that’s important for the translation process because when it comes to translations, we do not want to send just the source file. We also want to send the terminology to make sure that the translators use exactly the terminology that we have defined in our organization, not in the Oxford Dictionary, not in Merriam-Webster.

So this is an example for German, for instance. So when it comes to translation, so the deprecate term is not necessary, but for the sake of completeness, it makes sense to have that as well. And that means whenever you see the word bank and the context is clear, so we talk about this financial institution, use bank. And when it comes to translation to German, use the German word bank. So this is another one. So the same word, bank, but a different definition. So the English word is the same, and we could also find this deprecated term is different. In the German, it’s a completely different word. And that’s why it’s so important. So homonym, how we call it, in one language does not mean it’s a homonym in the other languages as well. So why is this important? It’s a lot of work maybe, but it’s a requirement to ensure content quality. So what’s that?

SO: A screwdriver?

TM: It’s a screwdriver and?

SO: 50 cents. 50 euro cents.

TM: 50-cent coin.

SO: It’s money, I guess.

TM: 50-euro-cent coin. Yeah. So what’s important? So we begin with the definition, not with the word and not with the function. So why do I use this example here? It’s a screwdriver. So my hobby is photography and videography, and sometimes I’m in the field and I need to mount my camera onto a plate or to mount it onto a tripod. I barely have a screwdriver with me, so I use a 50-cent mount, a 50-cent coin. It does not mean that the 50-cent coin is a screwdriver. So it’s not the function. It’s the sync. So a 50-cent coin is still a 50-cent coin and a screwdriver is still a screwdriver, or a flat blade or a blade, whatever you call it.

SO: I was going to say, flathead screwdriver.

TM: Flathead screwdriver. A lot of variations, but we don’t want to use the… Again, the two words that go together, correctness and consistency. That’s a big idea. So begin with the definition. If you cannot find a definition, it’s not a term. If you do have a definition, give the baby a name. Maybe you find more names, we call those synonyms, and then you say, “This is the good one and all the other ones are forbidden, deprecated, and not the function.” So this example, I think it’s really maybe a bit ridiculous, but I think it helps illustrate what I mean. So the function is here, you have the 50-cent coin, it’s sometimes my screwdriver for my camera only. But it’s not the same thing. It’s completely different things. Good. So let’s say this example is the last one, definition, the word, the deprecated terms, and the different languages.

So what is the relationship now with what we do at Congree with linguistic intelligence and generative AI? So let’s talk about generative AI. It’s creative. It is. So I use it myself. So I mentioned I do photography, and sometimes I do a picture of a lovely place somewhere in this world and I need to get rid of people. Not killing them, obviously, but to get rid of them. I can do this myself manually in Photoshop, but it takes ages. So now with GenAI, I can easily do that. Marking the places, remove the people, and it’s empty. So even the Vatican, the place in front of St. Peter in the Vatican City, it’s empty, even if I take a photo at noon.

It’s versatile. So we can use it for natural languages. We can use it to create videos. We can use it to create photos. We can use it to analyze huge sets of data in research. We can do a bit of forecast. Or I play, since I was a student, Go. And with Go, we learned a lot about this game we never understood for many centuries. So it’s variable, but it’s also predictable. And that’s the nature of GenAI. It’s nothing bad. As I said, it’s based on probability. So it can make things up and it learns from data. So it depends. So the quality or the output of a GenAI is depending on the data that we use to train the model.

So what have we in the product? We call it linguistic intelligence. And the linguistic intelligence is based on rules. And it’s reliable, so we can check the same content over and over again and we will always get the same result. So I guess that you guys have played with prompts and GenAI. And when you run the same prompt again and again, you always get another result. This is something that we don’t want. So we want to be… Or it’s important to be predictable, reliable, deterministic, as opposed to probabilistic. And it learns from rules. So if we want to fine-tune that, if you want to enhance features, so we need to add some rules. So now, the idea is to combine both worlds because the linguistic intelligence can do things that GenAI cannot do. So we know the style guides, the corporate languages our customers have defined. So we know the terminology. The GenAI has no clue about it, I mentioned it.

But the GenAI can create things. That’s something we cannot do. Let me give you a simple example. So let’s assume we have defined. We don’t want passive voice. So now, an author was using passive voice, and depending on the language, it takes a while to make this sentence written in passive voice a sentence written active voice. I think it’s a bit easier and faster in English than in German. If you want to learn more about German, I always recommend read Mark Twain’s The Awful German Language. So we don’t want that because we don’t have time, so we can use GenAI as a tool to accelerate our content review and content creation process, or the other way around.

So what we do is we send the sentence to the GenAI with a prompt dedicated or attached to this specific rule and say, “Could you change the sentence for me and make it a sentence in active voice?” “Oh yes, I can. Here you go.” And then so this issue is gone. And there are a couple of other examples where we benefit from the GenAI and have the GenAI to do things that cannot do by its own. So that’s from my side. So Sarah, now it’s your turn. So I was talking about two of the horsemen, terminology and language.

SO: Yeah, so it’s interesting because obviously those two are critical because the text itself matters. Additionally, and I wanted to circle back to something you said earlier about science and entropy. So if we haven’t already run everybody off by talking about entropy, I wanted to talk about chemistry just briefly. If you think of your content, your text, as being the chemical ingredients that you need in order to do certain kinds of things, you mix A and B together and you get C. So your outcome in terms of when you mix A and B together to get some sort of an output, language is effectively the chemicals, the raw materials that you’re putting in. And if those are not pure, if they have impurities in them, if they are inconsistent, if they don’t have good terminology and all the rest of it, then the results you’re going to get can tend towards chaos or you won’t get the output that you’re looking for.

The two things that I wanted to talk about today are semantics and metadata taxonomy because… I know that’s three things, but they’re two interrelated things. Because what you run into is that that is like the beaker that you’re using to do this chemical reaction. And if your beaker is not of high quality and clean and tempered, it will do things like explode when you put it over a Bunsen burner. Or again, there’s gunk in the beaker, to use a scientific term, and so the results are not what you wanted. So we have this issue of the language itself, the text, the sentences, the phrases going into an AI being like the raw materials, and then we have the actual process of mixing them and giving that mixing process some constraints and some guardrails. And that is what semantics and taxonomy are about.

So when we talk about semantics, what we’re talking about is tagging things. Very often we talk about structured content, but let’s dial this back a little and think for a second about a Microsoft Word file. If you have set up your Microsoft Word file with some styles and you have tagged everything consistently as a Heading 1 or a body text or a note or this, that, and the other thing, that’s semantic. You’re providing some labels to that content, which are then, if they’re expressed as formatting or if you’re using structured content XML or something like that, then explicitly those are available downstream to the AI to latch onto and use in their processing in building out those vector relationships in determining what things are related and how closely they’re related.

Because, for example, if you have a heading that’s for tasks, then it would have an ability to say, “Oh, this was part of a task, which is a procedure, which is instructions.” So it gets used in the bucket with instructions, as opposed to something that is labeled as side note or pullout quote or reference material. So there’s a really interesting concept, I think, that we have these fundamental building blocks that are perhaps atomic, and then we put those all together and we’re trying to impose some structure and some guardrails on what we’re doing.

And then switching gears entirely, if you think about universal design principles, when you build a house or a road or any sort of infrastructure, it is much, much easier to build universal accessibility into that thing ahead of time. When you build a road, you can put in curb cuts that wheelchairs can pass through instead of a curb that would cause problems. Interestingly, other people also use curb cuts. If people have limited mobility or if they have a stroller or a bicycle, it might be more convenient to use a curb cut. If you build a house, it is very much cheaper to put railings in the bathrooms ahead of time as you’re building the house than it is to put them in later.

And that is what we’re facing with AI ingestion with LLMs, is that it is way, way cheaper to build out the maturity of your content, to do the terminology work, to do the structure work, to do the metadata work, label everything, give it categories, give it classifications. It’s much easier and much cheaper to do that ahead of time than it is to try and remediate the house or the content after the fact, after it’s been processed, after it’s been ingested into the AI and then spit back out. And my fear right now is that what we’re seeing is a lot of ingest everything, spit it back out, and then consider how to fix it. This is called human in the loop. Your best bet is to have the human in the loop ahead of time in the planning phase, not in the after. And right now, everything we’re doing is… Not everything. 99% of what’s being done is after the fact. It’s, “Oh, that didn’t work. It hallucinated. Let me go find some guardrails. Let me fix that.”

Many, not all, but many of the things that cause problems, as Torsten is describing here, can be addressed, note use of passive voice, can be addressed by doing the work ahead of time before your content goes into the LLM either that you’re building or that is being fed or that is consuming your content, whether you like it or not. And while I agree that organizations would prefer to protect their IP and keep it out of the public-facing large language models, there’s not a lot of evidence that people are doing that successfully. Actually, keeping content out is very, very difficult. And even if we successfully kept it out of the initial training phase, there’s still a pretty good chance that somebody like me owns a car and something’s not working, and so I just upload a manual into Claude or ChatGPT. And now, it’s in there. Was it permitted? Probably not. Did I do it anyway? Yeah, probably so. And so here we are.

So big picture for me, we’ve got the terminology and the linguistic constructions, and then we’ve got the, above that, the semantic labeling, the tagging that you’re doing, and the metadata, the classification systems that say this is a procedure or this belongs to… Torsten, you mentioned this. Just because it’s a car manual, you still have the question of which model, and within the model, which trim line. There’s, I think, very few things more frustrating than a manual that says, “If you have the AB trim line, then you don’t have this feature.” Or, “This feature applies only to the CD trim line.” Okay, you know what? It’s 2026. I bought a car. You know what car I bought. Why am I not getting a manual that is exactly tailored to what I just did? That is not rocket science. It’s hard to do in a print workflow and it’s hard to do in a mass-produced PDF or mass-produced print workflow, but it is just not that hard, and furthermore, I’d say does not require AI in the least.

So the problems that I see and the, I worry about bad, terrible things happening and disaster happening as people generate synthetic content, content that is not directly authored by the official owners of the product or the content, is that if the inputs are not good enough, the outputs are not going to be very good. And we can try to remediate it on the backend, but it’s more expensive and probably less effective. So I think that looking at these questions of how do we configure this and how do we do it on the front end, how do we do it ahead of time before this stuff goes in, I think is a question that more people should be asking.

SA: And on that note, I’m going to jump in. So this is Scott just giving [inaudible 00:48:23]-

SO: Hey, Scott.

SA: … that we have 10 minutes left and we also have a couple things that you should know about. First, our original poll shows that 30% of our audience are moderately mature in their level of AI maturity, so they say, 25% are highly mature, 5% are extremely mature, so they could teach other people, 20% are kind of sort of limited, and 20% not mature at all. So about almost half, 40-some percent of the people slightly did not mature in their AI maturity level. There’s that. And we have some questions from the audience that we still want to get to, so I want to encourage you to cover the four horsemen topics as quickly as you can and we’ll get to the questions.

TM: No, Scott, you can go ahead and ask the questions.

SA: All right. And one of the viewers asked if we intentionally were leaving this slide on the screen to teach them something, and if not, I’m going to remove it.

TM: No, I was just too lazy to remove it. No intention.

SO: Less honesty, more…

TM: I’m German, we are straight.

SA: Do we need to cover-

SA: …anything on the horsemen that didn’t get covered? I’m just curious to make sure that you said everything you wanted to before we start taking questions.

TM: Yeah, you can take questions. We are through.

SA: Okay, great. The first question is, at what point does AI-assisted content creation become AI-generated content debt for an organization?

TM: That’s a good question.

SO: Two weeks ago.

SA: Yeah, go ahead. Would you like to go first, Sarah?

SO: Well, I mean, the answer is it depends. It depends on how… And actually, let me back up and say that we’ve been talking a lot about creating content for ingestion into the LLMs, but really when we talk about AI the topic, we need to be talking about AI as a productivity tool on the backend authoring and AI as a delivery mechanism for the consumer. And we’ve been largely saying, “Here are some things you need to do to succeed with that second one.”

But when you talk about AI on the backend as an authoring support system, which is what this person is asking about, the answer is it’s a tool, so use it properly. So used properly, it’s not going to create content debt. Now, if your actual question is, “Hey, my organization has slammed AI into the organization, fired all the tech writers, is generating total garbage and that is then being dumped into an AI,” then yes, you have a content debt problem, absolutely. And you’re not alone.

So I think the question is, how do you use AI in a content creation workflow to support what you’re doing and as a productivity tool that makes your life better, or makes the content you’re generating better, not worse? So AI is really, really good at patterns. AI is good at things like identifying outliers or… I mean, it’s not a spell checker, but if you think of it as a glorified spell checker, it will find things that are not quite right and highlighting for you. That’s one thing you can do. You can, of course, go way beyond that into, “Give me my initial outline, give me a draft,” and then I’m going to look at that.

And I think the answer is that at the point when you start thinking about it more as an author or a co-author as opposed to a supporting tool, that is probably the point where it is going to cut over from useful to problematic. And, oh, sorry, one more thing. If your sources are good enough, you can automate a lot of things. So ultimately the question is, how good is your source material? And that’s just pushing the question of when is the authoring being done by the humans farther back in the work stream.

SA: And to be clear, you could automate a lot of bad things too. So you’re not trying to say that. You’re trying to say that if you want to get the automation deliverable, correct, or as clean and debris-free as possible, you want to clean up front in order for that process to filter down through the results.

SO: Right. And use of AI does not necessarily equal increased content debt. Use of AI could equal increased content debt. And so to a certain extent, the premise of the question is, how much AI is too much AI? And I agree that there is a line, and it’s going to be situational for every organization. Because actually, how good is your data? How good are your inputs? That’s the question.

SA: Great. And we have another question. So this dovetails very nicely to that one. Which of the four issues that you called the horsemen tends to cause the most downstream damage in enterprise environments and why?

And before you answer that question, I tell the audience members that our second poll is open. Which of the four horsemen worries you the most? And your choices are no suitable metadata concepts, no classification, no company-specific language, or no company-specific terminology. And there are people already taking it, so thank you for doing that.

All right, let’s go back to that question again. So which one caused the most downstream damage, do you think, of the four horsemen?

SO: What do you think, Torsten?

TM: If you just talk about the source language, then it’s probably your topics, the classification and metadata. So when it comes to translation and we want to publish and retrieve information in other languages, it’s language and terminology. But they all go hand in hand, so it’s difficult to tell what is more important because they go together.

SA: I was going to say that, “It depends,” answer slips in there. The minute I start trying to parse them, I realize that I could tag it correctly, but then have the content be gibberish, and that would be [inaudible 00:54:34].

TM: Exactly.

SO: I think at the end of the day, if the content is wrong, then you’re doomed.

TM: So in addition to the other question and answer from Sarah, I think, so the biggest misconception is, and you mentioned it, Sarah, so they see GenAI like a magic wand or a magic bullet. We can get rid of all the authors and GenAI does all the magic for us. So it’s like Mark Twain said, “He who has a hammer sees everything as a nail.” But it’s only one tool in our production or content creation chain, not more. It’s a great tool, but it’s not a magic bullet. Not at all.

And that has also changed in the last three years on the conferences. And to me, there was a visible change last year at LavaCon in Atlanta. Before that, there were some people sitting in their dusty attic or dark basement playing around with some prongs and they were talking about that, but it was not enterprise grade, no production grade. And now, we understand all the opportunities we have, but also the limitation and constraints, and we know what we need to do to make it work for us.

SO: Yeah. I think, I’m looking at this poll and it looks as though it’s split really between terminology and metadata. And I mean, that seems right. I think terminology, at the end of the day, these issues with… I mean, bank. There’s also credit union and self-help, and I’m not a financial expert, but there’s all these other terms and they mean specific things. And if you don’t use them properly, then what chance do you have when the content is remixed and reprocessed and translated?

TM: Yes, exactly.

SA: Yeah. And it all goes back to this. I did this talk years ago where I blame my fifth grade teacher, Mrs. White, who was my Language Arts teacher, for all of the problems that we have today because she taught us about synonyms in a chapter in our Language Arts book called Introduction to the Thesaurus. And I found this book on eBay. I went back and found it a few years ago to see if I was crazy and misremembering. I was not. She told me things like, “Don’t ever use the same word more than twice in the same two or three or four paragraphs.” That’s not even a rule. That’s so flimsy. And now, we’re saying, “Use the specific word and use it religiously over and over again so we don’t introduce the slippery, almost but not quite, equivalence that might be in the thesaurus.” Is that also something that maybe we struggle with as human writers that the machines need to be able to help us correct if we find all that?

SO: We are not in techcomm. We are not supposed to be writing fiction.

SA: I agree.

TM: Oh, thank you, Sarah. How did you know what I was about to say?

SO: The rule is for writing for entertainment, writing for interest. So she’s not wrong. It’s just that in this problem set, the example I always use is car seat. Car seat, baby seat, baby bucket, safety seat. Which one is it? Booster seat. Those are all car seats. Infant seat, rear-facing safety thing. Those are all terms for some sort of a car seat and they have specific meaning, and you should pick one or pick the one that you are referring to, not use them all interchangeability because it sounds fun.

SA: Right, exactly. We could go on forever with this, but unfortunately we’re out of time. So audience members, thank you very much. Before we close out, I will tell you, in case you can’t see the results that we can see or the way that we see them, here are the four horsemen results. Let’s see, half of you on today’s show pick no company-specific terminology as the one that was most worrisome to you, followed by 30%, or one in three of you or so, no suitable metadata, and then 10% each for classification and company language. And then we didn’t leave any room for a fifth category, so I’m sure there’s an other bucket that some people would love to throw in there too. But we’ve got our work cut out for us, so thank you very much. And Torsten, for people who want to learn more about Congree, how do they reach out to you to do that?

TM: [email protected]. T-M-A-C-H-E-R-T at congree.com.

SA: All right.

TM: Or you can meet… Even better, Scott. So everybody can meet the three of us in Pittsburgh in October at LavaCon. Not Pittsburgh, sorry.

SO: Charlotte.

TM: Charlotte, sorry.

SO: Pittsburgh was last month.

TM: Yeah, I know. I know. Thank you.

SA: I didn’t know that city was in Pennsylvania. All right, well, without further ado, I will correct that very quick and say, here’s where we’re going to be this October. We hope that you’ll join us the 25th through the 28th in Charlotte, North Carolina-

TM: Yes.

SA: …for the Content Strategy and Technical Communication Management Conference. You can save some money off registration if you use our discount code, TCW, at checkout. As always, we’d like to thank our sponsor, Congree. And thank you, audience members, for being so encouraging and also tolerant of our antics. Please give us a rating on the quality of the information you’ve learned today with our one-through-five-star rating system, and know that you’ve been watching How to Survive the Four Horsemen of the AIpocalypse with Torsten Machert and Sarah O’Keefe. Thanks for joining us. As always, be safe, be well, keep doing great work, and we’ll see you on another webinar in the near future. Thank you. Thank you, panelists.

SO: Thanks, Scott.

TM: Thank you, Sarah. Thank you, Scott.

SA: Bye, everybody.

TM: Have a good one.

Want to stay updated on key industry insights? Subscribe to our newsletter!