Unexpected Youth Research Winner: AI "Jacob" for Easy Language

Last weekend, the 2026 Jugend forscht competition concluded with the awarding of federal prizes and numerous special prizes. This year, the International University (IU) sponsored 16 special AI prizes at the state level and one special AI prize at the federal level. Heise online had the opportunity to interview Magnus Schlinsog (18) from Brandenburg, a special AI prize winner, shortly before the federal finals. His project was already highlighted by IU after the state prize awards as a prime example of digital inclusion and education for all. Schlinsog trained an artificial intelligence for “Easy Language” and originally had no plans to participate in Jugend forscht.

Promoting AI Competencies Early On

According to Prof. Dr. Kamal Bhattacharya, Vice President for Research & Transfer, IU's initiative and sponsorship of the special AI prize in agreement with Jugend forscht is part of its own approach to AI in university teaching. He explained to heise online: “We specifically train and promote AI competencies for employees and students.” With Jugend forscht, they want to support young people in “understanding AI as a key competence.”

Bhattacharya did not comment on whether IU will also sponsor a special AI prize as part of Jugend forscht next year, but noted that IU remains in discussion with Jugend forscht. A competition specifically designed for artificial intelligence for pupils in Germany is the Bundeswettbewerb KI (BWKI). Bhattacharya elaborated on why IU preferred Jugend forscht: “Jugend forscht, just like the Bundeswettbewerb KI, is a renowned prize for promoting young research talents. We value both highly. At Jugend forscht, participants can submit their projects in seven subject areas, covering a broad spectrum of science – especially in the natural sciences. In many of these areas, AI is now used specifically to support research processes and gain new insights. For us, it is important not to view AI as a specialized topic, but as a future competence that is relevant in many areas: from education, the world of work, STEM subjects, to societal issues.”

Interview with Magnus Schlinsog

You are a participant in this year's “Jugend forscht” competition and have already received a special prize from the International University (IU), which offered an extra AI prize. You trained a language model that is supposed to respond in Easy Language. How did you come up with the idea to do this?

I've been interested in computer science for a long time – I started attending the CoderDojo at the Hasso Plattner Institute at around 9 years old, encouraged by my parents – and I was looking for a good topic for my seminar paper at school to train an AI for a specific application, because the currently known language models provide a good basis for that. But the decisive idea came from my mother, who works at the Paritätischer Gesamtverband. “Why don't you create an AI for Easy Language – people would surely be interested in that.” I looked into it more closely and thought: “Yes, that's perfect!” I'm not just using language models for what they're already (quite) good at, but I can fine-tune them for “Easy Language” through my training. While you can ask common language models to simplify language via prompts, they usually don't do it according to the actual rules for “Easy Language.” There is a difference between “simple” and “easy” language and how these texts affect people who have difficulty understanding. So, my AI “Jacob” is actually supposed to function like ChatGPT as a chatbot, but for people with special needs.

So you found a goal for your AI training that is concrete and not just abstract and also follows specific rules – so you basically knew who you were training an LLM for and with what?

Yes. There are many people who have learning difficulties or are restricted in their language comprehension in other ways and are therefore unable to understand every text equally. German can become very complicated – for example, due to many chained sentence structures. Easy Language tries to provide a highly simplified form of German. There are some rule sets that can serve as a basis, such as those from the “Netzwerk Leichte Sprache.” These specify maximum lengths for words and sentences, a restriction on subordinate clauses, negations, and also the use of the genitive case, for example.

How did you implement that technically?

To anticipate, when I speak of “AI,” I am more precisely referring to so-called language models, but I use both terms synonymously for simplicity. But yes, how did I do it? I chose a model from one of the major companies that provide models that users can further train. For reasons of digital sovereignty, it was important to me to use a European model, even if there are sometimes better ones for my purposes – for example, Chinese models like Qwen. However, they have the problem, among others, that they were not sufficiently trained with German-language sources, and since I was specifically concerned with language specialization, the focus on European models was also much more suitable for my project. Even though I wasn't always completely satisfied with the process, I ultimately opted for the LLM Mistral Small 3.2.

I then implemented the fine-tuning of the model as follows: I created datasets with which I could later train the language model. I then had a larger, more capable cloud AI conduct conversations with itself, so that I could select those conversations that came particularly close to the goal of Easy Language, to use them as models. The smaller language model Mistral Small 3.2 was ultimately trained on these.

So, in between, you evaluated these conversations, discarded some, and then incorporated those into the dataset that better implemented the rules of Easy Language. Were there any datasets or libraries you could adopt from elsewhere, so you didn't have to do it all yourself?

That is actually the innovation in my project, because such a dataset did not exist yet. While there are texts in Easy Language that can be found online, they are not designed like conversations that a chatbot normally has with a user. To solve this major dataset problem, I had my conversations synthetically generated by a large model in the cloud and then developed an algorithm that can assess how good the generated texts in Easy Language already are. Based on the rule sets, I developed the algorithm to ensure that the model's outputs are as compliant with the rule set as possible in the end.

When one thinks of existing texts in Easy Language, party programs or documents from authorities come to mind, which are linguistically strongly located in the political sphere or explain certain legal foundations more easily.

Exactly. And if my AI is to be able to reproduce more than just something like that, it also needs role models or example texts for everyday communication. Otherwise, it cannot be truly helpful. So I also had to consider what the AI should simplify linguistically if it is to be suitable for everyday use. For example, how should a question about the weather be answered? How is a recipe presented to me if I want to bake something? I had it practice such conversations. For this, I used the aforementioned large cloud model – Mistral Medium 3.1 – because it was significantly more capable than just asking small local models. This large model was then allowed to conduct conversations a thousand times, which I filtered using my algorithm. This way, I could ensure that the data quality was sufficient to proceed.

What was particularly difficult about this work?

It is definitely not easy to represent every rule for Easy Language in an algorithm. I can very easily check how long a word is, but it is not so easy to check whether a sentence is in the genitive case. Sure, there are words to watch out for, like “des,” but then there are also many false positives. Negations are also not always clearly recognizable. And some rules were over-implemented meanwhile, such as avoiding many commas that arise from subordinate clauses. The AI even “cheated” a bit – in quotation marks – and deleted actually necessary commas. So my algorithm is also not perfect (yet). It regulates many things correctly, but to make it even more precise, I have now trained a second AI model based on BERT. That is also the current state of my project. I am now using newer models, as the beginning of my project was some time ago. The ones mentioned before are a bit older, but I started the project around the first half of the 11th grade and just took my Abitur exams – time was a bit tight then.

This means that what you submitted to Jugend forscht is not a completed project, but you are continuing to work on it. And you also mentioned that you were actually working on the topic for a seminar paper. So you didn't have “Jugend forscht” in mind at the beginning?

Yes, my STEM teacher found my seminar paper really good, which I did in the STEM seminar course. She told me, “Well, submit it to Jugend forscht!” She also helped me a lot with the application.

But that raises the question for me of how you were supported for your project at all. Some LLMs can be run and trained on local computers, but you need quite good hardware for that. Did you receive or could you use that from the school? Was it all your private endeavor?

Yes, even using Mistral Small was not easy – partly due to some design decisions by Mistral that forced me to modify the code of some large open-source projects in the meantime. But that's another topic. The small model alone definitely has 24 billion parameters: So it's not exactly small, but not huge either. But that also meant I couldn't just train it on my home computer. Maybe run it, but not train it. So I contacted the Stadtjugendring in Potsdam, and they said: “What you're doing in disability support is wonderful.” And then, fortunately, they funded me to rent graphics cards online. I then continued training it on those.

Oh, wow. That's great! Was it difficult to make these contacts, or are there better conditions locally because, with the HPI in Potsdam, computer science enthusiasts already have a good standing?

That actually had nothing to do with it. There is the Children and Youth Budget in Potsdam. Children and young people can go there and say, “Look. I have an idea. I need some money for it.” And the city of Potsdam then gives up to 2000 euros to each project. What is not used from the money must be returned. And it was really easy to get it. You go to their website and send an email. Then you get an invitation to present the project, and in my case, the money came relatively quickly afterward.

That sounds really good! And I think it prepares you very well for the future. You had to calculate necessary capacities and costs, find partners. These are valuable experiences. Your project is also so practical that you have probably actually used it – certainly also for feedback on quality and further development?

Yes, I am in contact with Lebenshilfe in Potsdam, among others, which also has its own testing group for Easy Language. I had people in some facilities try out Jacob there and ask for feedback. I first tried it with a feedback form, to make it more scientific, but that didn't work very well. However, I then obtained feedback through the supervisor. And I also received feedback from the disability commissioner of the city of Potsdam. The AI has problems here and there that I am still trying to fix, but according to the testers, it formulates its answers in Easy Language and is usable for its intended purpose. Most texts are said to be very easy to understand. One criticism was that there are sometimes delays in the responses, as I am not running this professionally and the rented hardware sometimes reaches its limits.

Based on your experiences with the feedback form, it clearly shows how well or poorly some texts or forms work for people with special needs. Even there, the need for Jacob is evident. How do you want to develop it further? You haven't stopped developing it since submitting your seminar paper.

I am actually still working on improvements. I am also trying out different strategies to evaluate the dataset and train the model. For example, a reward system that is more oriented towards user feedback, rather than just minimizing deviations from the rules. However, I must also add a caveat here: I naturally want to avoid the AI hallucinating just to output any answer, or having conversations with users that could ultimately harm them. So, Jacob should not try to have extensive conversations even if he doesn't have a good answer ready, or keep people engaged in critical situations when they need help. Therefore, there are clear instructions to, for example, quickly refer to support services if the user expresses dangerous thoughts – such as calling 112 or the telephone counseling service. Regarding usability, I would like to enable voice input and output. This is very useful for the target group.

So far, Jacob is browser-based. Is an app also planned? It could also make it easier to capture texts via a camera and translate them into Easy Language, similar to what is already known from Google Lens for foreign languages, for example.

Yes, that's the direction I want to go: keep the LLM multimodal. But first, an API was important to me, but that's the intended path.

Does the name Jacob have a special meaning? Is it an acronym?

No – during development, I experienced that users don't like typing “Hello Chatbot” but prefer to use a name, and Jacob came to mind as a simple German name. However, once I introduced this name, it also led to confusion because, for the loading process, I displayed “Jacob is waking up.” Some testers therefore believed that a real person was doing Jacob's job and sometimes slept. This taught me that while “Jacob” is more pleasant for users in conversation, I also need to instruct the chatbot to repeatedly clarify that it is not human. Questions about its well-being – like “How are you?” – should therefore never be answered with “good” or “bad today,” but rather with: “I have no feelings and cannot have them, because I am a chatbot.”

So you had to deal not only with how an AI can be trained for your own purposes and how you can make it accessible to people, but also with the ethical questions that come with such an offer. What are your next plans?

I first considered taking a year off after Abitur and doing Work and Travel, but that doesn't quite fit now. Therefore, I will first start studying Computer Science at the TU Munich – of course, in the direction of Computer Science – and then also plan to go abroad during my studies.

(kbe)