Skip to main content

Processing

Language Models​

Language models are advanced AI systems capable of understanding, interpreting, and generating human language. These models are trained on enormous datasets of text and learn patterns, word combinations, sentence structures, and even the subtleties of different languages and language uses. The core of many modern language models is the transformer architecture, which uses self-attention mechanisms to determine which parts of the text are important in a given context.

When processing language, these models use statistical methods to predict the most likely next word or the most likely next sentence in a text. They can understand context over long pieces of text and thus generate not only grammatically correct but also coherent and relevant content.

When using a language model for, for example, a chatbot or a text generator, the model is given certain prompts or initial data, and based on that input, the model generates text that logically follows from the given context. The goal of these models is to produce text that appears as human-like as possible, both in terms of content and style.

Text-to-Image Models​

Text-to-image models are AI systems capable of generating visual representations, such as photos, illustrations, or other types of imagery, from written textual descriptions. These models use advanced neural networks, more specifically generative adversarial networks (GANs) or variations such as diffusion models.

The process starts with a user-entered text description. The model evaluates this text and tries to understand its meaning and context. Then the model generates images that correspond to the textual description, using what it has learned during training, which involves huge datasets of text-image pairs.

During training, the model learns associations between textual descriptions and visual features. For example, if the model repeatedly sees the phrase 'a yellow sun above a blue sea' along with images illustrating this scenario, it learns to recognize and reproduce these elements in future image creations.

The results are often surprisingly accurate and detailed images that match the input text description. These models are becoming increasingly sophisticated and can depict complex scenarios with multiple objects and abstract concepts. They are used in a wide range of applications, including artistic creations, game design, virtual reality, and more.

AI-School Model Access​

It is important to realize that AI-School provides access to various AI models included in the central model catalog. These models are offered by external providers via an API. An API, or Application Programming Interface, is a set of rules and definitions that allow software programs to communicate with each other. It functions as a kind of 'language' understood by programs to exchange information and invoke each other's functions. AI-School itself does not have language models, text-to-image models, or text-to-speech models.

We are not responsible for the results of the different models. However, we have paid attention to selecting the best and most interesting models for schools.

Processing Procedure According to the DPIA​

The DPIA describes the basic procedure for generating a response as follows:

  • The user enters a prompt on the AI-School portal.
  • The prompt is stored in the AI-School database.
  • AI-School offers the prompt to the chosen AI model via an API connection and a general account without linking to the AI-School account.
  • Account and student administration data are not automatically added to the model.
  • The AI model's response is sent back to AI-School.
  • The response is stored in the AI-School database.
  • The response is presented to the user.

AI-School does not automatically send account and student administration data to the AI model. The model does receive the content of the prompt without linking it to the AI-School account. A user may have included personal data in the prompt; in that case, this data can be processed by the chosen AI model.

Additional Context​

For functions such as chatting with documents, multimodal input, or certain tools, the prompt can be supplemented with user-selected context, such as document fragments, images, or search results. This context is also offered to the model without linking to an identifiable user, but the content may contain personal data if the user has added it themselves.

WhatsApp