At the beginning of December, Timnit Gebru, co-head of the ethical AI team at Google, announced on Twitter that she was no longer working for the company. There are different representations about whether she left herself or was released.

So far, the only thing that is clear is that the trigger for the dispute was the planned one A paper was published in which Gebru and colleagues criticized the dangers posed by large AI language models such as GPT3. Google had refused to release the publication – officially because the paper “does not meet the requirements for a publication”. But there is a suspicion that Google wanted to get rid of an unpleasant critic in this way.

After studying the Physics, Wolfgang Stieler switched to journalism. Until 2005 he worked at c’t, in order to then act as editor of the Technology Review. There he oversees a wide range of topics from artificial intelligence and robotics to network policy and questions of future energy supply.

The fact that Google boss Sundar Pichai has meanwhile apologized for the incident in a memo and announced an internal investigation could not appease Gebru. In an interview with the BBC, Gebru accused Google of “institutional racism”.

High-risk technology But the argument about Gebrus’ dismissal holds far more explosive than meets the eye. The ever larger language models, which are now trained with a few hundred gigabytes of data, now show an amazing understanding of language.

In the However, Gebru and her colleagues criticize the use of large language models on a broad front: the excessive energy consumption during training of these models , the impossibility of proper documentation of the training data and the associated extrapolation of prejudices and exclusion, and the exclusive access of private companies to knowledge – all of this paints the picture of a high-risk technology.

In fact, none of this is really new. The paper cites works that have long been published. The fact that pre-trained large language models also deliver “toxic answers” ​​- racist, misogynistic or whatever – when the “prompts”, i.e. the first sentences, are explicitly, measurably harmless Paper, may appear to be a technical gimmick. However, if such models are used across the board for sentiment analysis, to sort job applicants or to make formulation suggestions for office communication, things look different.

Understanding the world What makes the whole discussion so explosive, however, is an often overlooked theoretical detail: the whole The complex of criticism of racism, misogyny etc. theoretically relates strongly to a constructivist view of the world. According to this theory, “words and sentences do not describe things in themselves, but always do so from a certain perspective,” writes the German scholar Andreas Gardt. “As language does not simply passively depict the things of the world, but guides our intellectual access to them, it shapes our image of reality.” In other words: language shapes consciousness and must therefore be used “correctly”. The question of whether the theory is really true is still bitterly debated.

But the joke is: This theory is one hundred percent true for an AI language model. Language shapes their understanding of the world. For AI, words and sentences form the basis for connections and relationships between things. If at some point, as the researchers at Open AI hope, it turns into a generalized artificial intelligence with real intelligence, what will it have learned about the world? Gebru himself does not seem to consider this possibility very likely. However, your criticism of the risks of this type of AI research is fully justified. Such a powerful technology cannot be controlled solely by a small handful of transnational corporations.

(wst)