Herausforderungen und Herangehensweise bei der Einbindung eines Large Language Models in den Newsworkflow - betrachtet am Beispiel der Einführung von ChatGPT

Abstract

With the progressive development of Artificial Intelligence (AI) over the past 80 years, it is now finding its way into the various areas of people's lives. In the automotive, insurance and media environments, AI is already an integral part. With the rapid development of Large Language Models (LLM) over the last five years, we are also seeing the application of language processing becoming more and more profitable. With LLMs, however, it is now possible to have more complex texts written by an AI. These thus form a powerful tool to support journalists in their everyday work. The goal of the master’s thesis is therefore to define requirements for an AI for use in journalism and to propose an approach for the introduction of ChatGPT into the news workflow. The approach consists of a non-technical description of concepts to meet the pre-defined requirements and a technical solution to translate the concepts into a system landscape. The master's thesis therefore aims to show the requirements which have to be fulfilled and how they can be implemented in order for AI to be used efficiently and safely in journalism. Furthermore, it shows the limitations of ChatGPT in relation to the requirements and recommends the use of other AI models in addition to ChatGPT to cover even more use cases. ChatGPT is suitable for the use in the news workflow under certain conditions. It is technically possible to integrate the AI model into the news workflow. The introduction phase and the subsequent productive use are iterative processes. The individual process steps result from the experiences during the project and the feedback from the users. The preconditions under which ChatGPT will be integrated efficiently and safely into the news workflow are initially of a nontechnical nature. They include aspects such as new roles, measures to comply with legal and ethical requirements, feedback and average scenarios. The technical solution carries the concepts. By using additional automation software an intermediary is inserted between ChatGPT and the user. The middleman ensures the security, the transfer of information and the initiation of automatic workflows. With this solution about half of the use cases in the requirements matrix are put into action. In order to implement further use cases, other OpenAI models are suitable for the use in the news workflow. As a supplement to ChatGPT, the models from the InstructGPT bundle, DALL-E and Whisper cover more use cases. These thus also include the processing of images and sound. For the implementation of these models, additions to the existing concepts and technical solutions are needed. Additionally, the new models are capable of user finetuning, a method of giving the AI model in-house data as training material. This allows the AI model to learn and output the style and attitude of the media company. Nevertheless, even the combination of models does not cover all use cases. For example, the tagging of images and videos based on the visual component is missing. At the end of the day, AI is a technology that has undergone rapid development in recent years. It has the ability to enable a high increase in efficiency in processes. However, it still requires strict regulation and supervision by humans.