Code of Conduct for AI and Easy to read texts - a Talk with Frauke Jessen-Narr and Ruben Rhensius

Domingos: Welcome to a new podcast on digital accessibility. Today’s topic is the so-called AI Code of Conduct for easy to read (in German KI-Kodex für Leichte Sprache, easy to read is Leichte Sprache in Germany). We’ll find out exactly what that is in just a moment. I’ve invited two experts who can speak about it much more competently than I can: Frauke Jessen-Narr and Ruben Rhensius. First of all, thank you both for taking the time to join the podcast.

Frauke: With pleasure.

Ruben: Yes, gladly.

About Frauke and Ruben

Domingos: As always, we’ll start with a short introduction round. Frauke, I suggest you begin and tell us a bit about what you generally do and how you’re connected to the topic of easy to read.

Frauke: Sure. I’m Frauke Jessen-Narr. I originally trained as an occupational therapist—that’s how I started my professional career. Nowadays, I’m a communication educator. I work in the specialist service for Augmentative and Alternative Communication and easy to read at Diakonie Stetten. Diakonie Stetten is a large social service provider near Stuttgart, you could say in the greater Stuttgart area. They support children, young people, and adults in the areas of housing, work, school, and leisure. We’re talking about around 8,000 people who are supported by Diakonie Stetten. I myself am just a very small part of that. I’m responsible for augmentative communication—that is, supporting people who cannot communicate through spoken language—and for easy to read. This includes both internal translations and assignments for external clients.

Domingos: Thank you very much. Ruben, would you like to continue?

Ruben: Yes, gladly. I’m Ruben Rhensius and I work at the Diocesan Caritas Association. The diocesan association is, so to speak, the Caritas organization at the state level—although that’s not entirely accurate, as it is actually oriented around the respective diocese. I work there in a project funded by Aktion Mensch. The project is planned for five years and aims to raise awareness and build qualifications around easy to read among people in the Diocese of Limburg. The area stretches from Frankfurt up to the upper Westerwald region, so from Wetzlar across to Wiesbaden. There is a significant need for easy to read in many different fields. My entry into the topic actually came through AI: how we deal with it and how we use it. That’s also how I came across the AI Code of Conduct, because it’s about how we want to work with it responsibly in the future. I think we’ll explore that in more detail as the podcast continues.

What is easy to read

Domingos: Thank you. The topic of easy to read has already come up several times. I think most listeners generally know what it is. Still, for context: what exactly is easy to read?

Frauke: I’ll take that. easy to read is a special form of the German language. Its purpose is to enable or improve participation opportunities for people with learning difficulties. There are various sets of rules that define quite precisely what easy to read is and what needs to be considered so that texts are truly well understood by the target group. For a long time, one issue was that there were—and still are—many different rule sets with varying focuses. They don’t fundamentally contradict each other, and there is consensus on most points. But in the details, there are differences in formulation.

Frauke: For some time now, there has been the DIN SPEC for easy to read. I’d have to check exactly since when, but it has been available for just under a year now.

Ruben: Yes, about a year.

Frauke: Exactly, that would have been my estimate as well. Since then, we have the DIN SPEC for easy to read with recommendations for German easy to read. It aims to consolidate general rules from the existing rule sets and bring them to a common denominator. I was very pleased when it was finally adopted after such a long time—it really took quite a while.

Frauke: If you look at what defines easy to read—key rules or things that are easy to remember—the first thing you often notice is the visual appearance. easy to read texts usually have a large font size, typically at least 14 points. Line spacing is larger than usual, and sentences are short. That’s probably what stands out at first glance, along with illustrations. easy to read texts usually include images that support the text. This makes them easy to recognize, whether on screen or on paper. These are actually very central features that immediately catch the eye.

Frauke: Short sentences, one idea per sentence. No sentences that stretch across multiple lines. In addition, a very clear structure in the text. It’s also important to always choose the shortest and most commonly used word available for a given concept. However, that doesn’t necessarily mean avoiding abbreviations if they are widely known. No one says “Lastkraftwagen” in everyday life—“LKW” is definitely the more common term.

Frauke: That’s also a good indication of what easy to read is about overall: it is based on actual language use. Content is strongly simplified and presented clearly, but it stays close to what people know and use in everyday life. So no elevated language and no technical jargon.

Frauke: Another important point is to avoid negations as much as possible. If negations are used, they are printed in bold in easy to read texts because they are often overlooked. I always find the example of “kein” and “ein” very illustrative. A single letter determines whether I receive no money or whether a certain amount is available to me.

Frauke: Additionally, no subjunctive mood and no genitive case. There are really many rules that relate to words, sentences, the overall text, and its structure. All of this together ultimately determines the quality of a good easy to read text.

Domingos: Another special feature is that easy to read texts are reviewed by people from the target group. The texts are revised if certain parts are not understandable. That’s something that distinguishes easy to read from plain language, for example.

Frauke: Yes, that’s a very important point. I actually had that further down on my notes, but you’re absolutely right—it’s one of the most important rules: that people from the target group review easy to read texts. I find this extremely important for several reasons. On the one hand, through the reactions of the review group—or through discussions about the texts—I can see very clearly whether something is actually understandable or whether my own assumption about its clarity might not hold true.

Frauke: On the other hand, there is also a political reason. easy to read emerged from the self-advocacy movement of people with disabilities. I think it’s incredibly important that this expertise and empowerment are not taken away from that group again. In my view, this is also a political issue. I don’t know if you agree, Ruben, but I think it’s extremely important.

Ruben: Absolutely. And that brings us to a central point for the AI Code of Conduct. AI offers many possibilities to modify texts and structures. This can quickly lead to the assumption: well, maybe we don’t need the target group anymore. But that assumption is simply wrong. For verified easy to read, the involvement of the target group remains essential. That was also one of the key reasons why we said: we need an AI Code of Conduct—to ensure that the target group continues to be included and remains visible.

AI Code of Conduct for easy to read

Domingos: Yes, thanks for the transition. Let’s dive in more generally: Can you describe what the AI Code of Conduct actually does and what it regulates? Maybe you can expand on that a bit.

Ruben: Yes, gladly. I’d actually like to start a bit earlier: how did it come about in the first place? The idea didn’t originally come from Frauke or me. It was Bettina Mikhail who first brought the topic into the discussion. Among translators and reviewers of easy to read, the topic was already present when the first AI tools emerged. The question quickly arose: how do we deal with this?

Ruben: Bettina then published a post and essentially said: we actually need something like a code—a shared guideline for how we want to work with AI. From there, things gradually developed. Around that time, I met Annika Lange-Knieb. Together, at a Barcamp organized by Lebenshilfe in Hesse, we decided to really pursue this path. Many people joined in and said: yes, we want to contribute.

Ruben: Frauke was also one of the people who helped further develop the document with us. The AI Code of Conduct incorporates many insights we gathered over a longer period. What we have today is the result of a lot of groundwork. For example, we interviewed people with learning difficulties and asked them: what is your view on AI? What do you think about AI-generated texts?

Ruben: We also conducted digital Q&A sessions—with translators, researchers, clients who commission easy to read texts, and people who develop or provide AI tools. All these perspectives, needs, and questions have been incorporated into the code. It was a very labor-intensive process that took over a year. But I believe that only because so many different people were involved could such a strong result emerge.

Ruben: That’s the process—how we got there. And from that follows the question: what is the AI Code of Conduct meant to do? It is intended to support people who want to provide easy to read in actually delivering high-quality easy to read. It is explicitly not about saying: “Okay, AI will handle it.” Because the idea that AI can reliably translate complex texts into easy to read on its own unfortunately does not work yet.

Ruben: To use AI meaningfully, you still need proper preparation of texts, careful post-processing, and above all the involvement of the target group. On many of the points Frauke described earlier, AI is still not reliable enough. That’s exactly why the code aims to provide guidance to ensure quality. Frauke, maybe you’d like to add something.

Frauke: I think the impression you described is very accurate. If I simply take an AI tool, input my text, and get an output that at first glance looks like easy to read, that can be very convincing—especially to an untrained eye, sometimes even to a trained one.

Frauke: That’s because word and sentence simplification already works quite well in many tools. You get short words, short sentences, negations are bolded—these formal aspects initially look correct. But what became especially clear to me while working on the AI Code of Conduct is that the real challenges lie deeper.

Frauke: It’s about text structures. Questions like: what is actually important in this text? What needs additional explanation? How should the text be structured meaningfully? Do I perhaps even need to change the order compared to the original text? AI is currently not very good at handling these aspects. That’s why human expertise is still essential.

Frauke: Even though it’s very tempting—there are now many providers and tools, including within ChatGPT, these specialized… what are they called again?

Ruben: GPTs.

Frauke: Right, GPTs. These suggest: just give me your text, and I’ll turn it into easy to read. But they don’t involve the review group. And you’re absolutely right: this is often a knowledge problem. Many people don’t realize how important it is to truly involve the target group.

Domingos: What exactly does the AI Code of Conduct recommend when dealing with such tools? What rules does it set?

Ruben: Ultimately, it’s about using AI tools in an informed and deliberate way. In consultations, the first question is often: which language form is actually appropriate—easy to read or plain language? And who is the target group I want to reach?

Ruben: This fundamental decision cannot be made by AI. It must always be made by the person using the tool. That’s where it starts. Then it’s about checking whether the rules are actually being followed correctly. If you use AI, you should provide it with clear information: about the rules, the target group, and their real-life context.

Ruben: And in the end, it’s about reviewing: is the text actually correct? Are the rules properly applied? That’s where the AI Code of Conduct comes in. It also emphasizes evaluating quality: does the text simply reproduce the original in a simplified way, or is it meaningfully summarized?

Ruben: Because you can’t assume that someone—even with an easy text—will work through 25 pages. That’s why it’s extremely important to decide which content is truly essential and what can be left out. This is also part of awareness-building: it’s not enough to just feed a tool and say, “That’s fine, this is good, verified easy to read.” That simply doesn’t work yet.

Ruben: That said, you can observe that tools are continuously improving. Language models are steadily getting better.

Domingos: Yes, interesting. For context: even before ChatGPT, there were other tools, but recently there has been a noticeable boom in German translation tools for plain and easy to read.

Domingos: Finally, I’d still be interested in your personal opinion: do you see these tools more as support for your work, or do you work entirely as before and not use them at all? Frauke, would you like to start?

AI as Support for Experts?

Frauke: Yes, gladly. I do use AI tools for translations, but only in specific steps. I still do a lot myself and examine every text carefully. I structure the text, decide which information is actually important for the target group, and what additional information is needed.

Frauke: Before that, I always clarify with the client: who is this text intended for? Who should it reach? And how will it be published—on a website or in print? All of this preparatory work is still done by us. Only then do I begin translating.

Frauke: I usually work paragraph by paragraph or sentence by sentence. Then I use AI tools to generate suggestions. At the beginning, I used to input entire texts into AI tools and then revise the translations afterward. But that doesn’t work because there were too many errors.

Frauke: I constantly had to compare the original text with the AI translation and check: is this factually correct? Does it really say what the original says? That’s why I now work in smaller sections. AI tools provide suggestions—everything else I do myself, using my own expertise, feedback from my colleague, and finally the review group.

Frauke: So the translation process has many steps: I translate, the client provides technical feedback, my colleague reviews the rules and implementation, and finally the target group checks the text. AI only handles a small part of this process.

Frauke: I actually find AI tools quite helpful in that regard. Sometimes you get stuck and just can’t move forward—and then it’s really great to be able to quickly input the text. You get a suggestion that you might not adopt one-to-one, but at least it helps untangle that “mental knot” you were dealing with.

Domingos: Ruben, how do you see it?

Ruben: Yes, that just reminded me of a book title from my studies: “The Fear of the Blank Page.” Frauke calls it a “mental knot,” I might call it the “blank page.” Sometimes it simply helps to have something you can start working from.

Ruben: At the moment, AI tools can be used very well as a tool—but they shouldn’t be used on their own. I find it interesting to observe how these tools are developing. Right now, they are still aimed at translators, not at the target group itself. It will be interesting to see how we can later enable the target group to use these tools effectively. Similar to digitalization in general, we’re still on a learning curve here.

Ruben: We also regularly monitor what’s happening in the market. In some cases, we invite providers and closely observe how their tools work. But so far, I haven’t found a tool that I would say can be used one-to-one without adjustments. At this point, they are really just tools that can support the process at different stages. And importantly, the target group is still not involved in the review.

Domingos: Alright, thank you for that assessment.

Frauke: What I really like about the AI Code of Conduct is that it is very open to technology. That was important to me. We could have written: “None of this works—do it yourselves or hire professionals.” But that’s not what it says.

Frauke: Instead, the code says: you can use these tools—but always check the results carefully, learn the rules, and become knowledgeable yourself. Only then can you assess whether the translation—or the way the AI has processed the text—is actually good and helpful for the target group.

Ruben: Exactly, that’s an important point. It’s not about portraying AI as something negative. On the contrary: AI can help ensure that more texts are translated more quickly—and in a more needs-driven way. That means not only because a client says, “I want this text translated,” but also enabling the target group itself to decide: “This is how I want the text, this is how I need it to be understandable for me.”

Domingos: Thank you very much for that perspective. One final question: where is the best place to follow you if someone wants to learn more about the topic?

Ruben: The AI Code of Conduct is available on the website of the Diocesan Caritas Association Limburg: dicv-limburg.de/leichte-sprache. The code is linked there directly. Otherwise, I’m quite active on LinkedIn, and you can also contact me via my website. Coordination around the code also runs through me and the project office.

Domingos: Great, thank you very much for the interview. It was very insightful—I learned quite a few new things myself. I wish you continued success with easy to read and your projects.

Further Information

More Talks with Accessibility Specialists