When creating a text, editing options such as ‘copy and paste’, formatting and a spell-checker are now taken for granted. LLMs (Large Language Models) such as OpenAI’s GPT models, Anthropic’s Claude or Google’s Gemini promise significantly more comprehensive support for text creation. Major advances have also been made in the transcription of video and audio recordings such as podcasts, in automatic tagging, and in image generation, for example, using GPT Image in ChatGPT, Google’s Gemini image models, Midjourney or FLUX.

What generative AI methods based on LLMs have in common is an inherent unpredictability of the results. This is particularly problematic because it is also impossible to predict whether the result is correct. The term ‘hallucination’ has become established to describe outputs that are factually incorrect but linguistically convincing.

GPT stands for ‘Generative Pretrained Transformer’. The name describes how it works: the model generates text (generative), has been trained in advance on large volumes of text (pre-trained) and is based on the Transformer architecture, on which most of today’s LLMs are built. The principle is relatively simple: based on countless learned texts and sentence patterns, the model generates the most probable ‘next word’ for a given text. It then uses the given text and the first generated word to generate the next word. And so on.

In this way, an LLM is able to generate text in response to a given question that is highly likely to constitute a suitable answer. Or, when the phrase “Shorten this text” is added to a text, to generate an abridged version of that very text.

An LLM, like everything in the world of computing, is actually deterministic. Therefore, the answer to a specific question would always be the same; even word for word. So why isn’t ChatGPT deterministic? As described, an LLM always generates the ‘next word’ with the highest probability. Since there are many words that are very close to this highest probability, this is exploited to introduce ‘creativity’ into the behaviour of LLMs. To do this, one of the words with the highest probability is chosen at random. At first, this makes little difference, but as this small element of chance is also added to every subsequent word generated, even a sentence of just 10 words, assuming there are ten options to choose from for each word, already results in ‘10¹⁰’, or ten billion, variants. In this way, an LLM can generate countless texts on a given topic.

If an LLM is allowed to operate on the basis of a given text and with little ‘guidance’, the results are generally good. The smaller the basis within the set of learned documents for the text to be generated, the less reliable the result. A particular problem here is that an LLM can only cite a source for a generated text if it was generated on the basis of one or more specific documents. Why is this the case?

An LLM has learnt all its information and its ability to generate text from an incredibly large volume of texts. Generated text is therefore not based on a single, identifiable source. A direct link only arises when the model specifically accesses a defined set of documents; this process is known as Retrieval Augmented Generation (RAG) and is now state of the art. Web search and grounding functions, which provide evidence to support outputs, are now also part of the standard repertoire. Without such methods, the results are based on all the texts on which the model was trained.

You can think of this in a similar way to how it works for us humans. We would not be able to cite the sources for all our statements either. This is not only due to our limited memory, but also because we, too, base a statement on many sources that we have become aware of over the course of our lives. In doing so, identical information reinforces itself, whilst contradictions cancel each other out.

The InterRed ContentAgents are based on state-of-the-art AI technology. A key distinguishing feature is the reliability of the methods employed. The ContentAgents’ AI always adheres to the set boundaries and rules. LLM technology is used exclusively in the form of self-correcting models. And generated texts always include references to the underlying sources.

A self-correcting LLM independently monitors compliance with specified rules - both in terms of content and form. For example, if it is asked to name three key characteristics of self-correcting language models, each in a single keyword, it adheres precisely to this specification: It names exactly three terms, each limited to one word, without additional explanations or digressions. In addition, it indicates the underlying source for each individual keyword - such as a scientific paper, a technical document or a documented recommendation for use - in order to make the origin and traceability of the terms transparent.

We call our AI, which adheres to rules and boundaries and documents the underlying sources of generated content, ‘SmartAI’.

With ContentAgents, AI can be used for any form of computer-aided writing without uncertainty. Thanks to SmartAI, there is no need for the sometimes time-consuming checking of AI results, which otherwise often eats up the time saved.

InterRed SmartAI stands for artificial intelligence you can trust. As part of our content technology, SmartAI works according to clear rules: secure, traceable and under control at all times. Decisions remain reliable and explainable to humans: Reliable AI.

Max Tegmark, physicist at the Massachusetts Institute of Technology (MIT) and co-founder of the Future of Life Institute, is one of the leading voices for AI development. His assessment of AI is very clear:

‘A powerful tool that we can control – that's the key. I like to drive a fast car on the motorway, but only when I have my hands on the wheel.’¹

He and his team have investigated whether AI can be tested by AI. In the form of a chain of AIs, in which a ‘slightly smarter’ AI controls the next one: ‘Our simulations showed that this does not work.’ This is documented in detail in the scientific study ‘Scaling Laws for Scalable Oversight’ (Engels, Baek, Kantamneni & Tegmark, 2025, arXiv:2504.18530). In it, researchers at MIT show that AI systems cannot reliably control each other. This proves that AI alone cannot deliver reliable results.

That is why InterRed SmartAI relies on a reliable control authority: a transparent, traceable supervisory level that ensures that AI always operates within the framework of human specifications. This allows it to remain a powerful but always controlled tool for greater efficiency while maintaining security and trust.

The ContentAgents automatically analyze all content from the InterRed ContentHub as a basis for your work. Any other sources can be added. The analysis is carried out completely autonomously based on the predefined sources.

The InterRed ContentAgents recognize the topics based on the connections in the individual texts. They learn independently, for example, to distinguish topics about a (parking) bank from topics about financial institutions (banks) and to link them with other suitable content. For example, they would assign articles about the money economy to the topic "bank" (as the financial institution), but never to the topic "bank" (as a park bank).

This is achieved through so-called ‘embeddings’. These allow words and documents to be transformed into mathematical objects (vectors). This can then be used, for example, to measure semantic distance or to determine the word that is semantically closest to another.

The InterRed ContentAgents also recognize people and places and provide further information about them. They can also contextually classify and associate synonyms, i.e. associate words such as "sun" or "warmth" with "summer".

Because they learn independently, constantly readjust themselves and automatically pick up and incorporate new topics, they are clearly superior to manual, hierarchical classification systems such as tagging, ontologies, categorizations or keyword directories. In day-to-day practice, it quickly becomes clear that AI delivers significantly better results than manually maintained and updated classification systems, which can involve a great deal of effort.

All processes for creating a print product can be automated with SmartPaper technology. From the selection of suitable content, the choice of suitable images and image sections to the automated layout and shortening of text. A complete edition can be created automatically in just a few minutes. And thanks to the SmartAI, which adheres strictly to specifications, the face of the product is retained and the result does not have to be laboriously checked. Among other things, InterRed SmartPaper uses the artificial intelligence of InterRed ContentAgents and thus the InterRed SmartAI technologies for automation.

Whether in everyday editorial work or at the touchpoint for the target group. The ContentAgents offer a powerful recommender system that can recommend related texts and images. Suggestions for texts to be linked or suitable images can be integrated into a text simply by dragging and dropping. And at the touch of a button, the AI can generate SEO-optimized links based on up-to-date SEO data at any time. The optimal teasers can be displayed on mobile and web. You set the target. Whether reach, reading time, paywall conversion or product purchase, the ContentAgents support the implementation of the respective strategy.

Instead of laboriously selecting texts and constantly adapting them to new circumstances over time, SmartCollections make it very easy to define a group of texts with the help of AI. These can then be used as required. For example, for teasers, a newsletter or part of a print product. As soon as new texts are created that fall within the definition of the SmartCollection, the usage - no matter where - is automatically updated. This feature also makes SmartCollections suitable as a "content radar". You can have the AI automatically inform you about new texts. By e-mail or message in Microsoft Teams, for example.

The integration of ContentAgents into the InterRed FlowEditor makes AI easy to use. Headings, subheadings, bulleted lists, pre-texts, teaser texts, summaries and social posts (Twitter tweets etc.) can be created at the touch of a button while writing in the editor. And, thanks to SmartAI, always reliably according to your own specifications.