According to a research paper, artificial intelligence is in principle not intelligent at all. This is because AI models are not capable of achieving results outside of their training data. The background.
Researchers from Deepmind, the AI company of Google parent company Alphabet, have published a research paper on the state of development of artificial intelligence. The result: in principle, AI models are not intelligent at all, as they are barely able to achieve results outside of their training data.
Is artificial intelligence intelligent at all?
The research paper has not yet gone through the peer review process – in other words, it has not yet been reviewed by independent scientists. However, Google researchers and authors Steve Yadlowsky, Lyric Doshi and Nilesh Tripuraneni are considered experts in their field.
According to the empirical results, AI models are capable of solving and learning tasks depending on the context. However, the previously trained data must be meaningful enough for this.
“However, when they are confronted with tasks or functions that lie outside the domain of their pre-training data, we show different failure modes of the transformers and a deterioration of their generalization even for simple extrapolation tasks,” the researchers say.
AI can only solve tasks if the training data is correct
In concrete terms, this means that AI models are only as good as the data on which they are based. The studies relate to GPT-2 from ChatGPT developer OpenAI. However, the researchers emphasize that their results generally apply to so-called transformer models.
ChatGPT is one such model – the “T” in GPT stands for Transformer. Such models are able to independently calculate an output from an input. In science, they are therefore most likely to be considered a form of AI that has the potential to be truly intelligent.
This is why transformer models also held the promise that they could go beyond the data they were trained with. However, even despite gigantic data sets, no artificial intelligence has yet developed demonstrable independence. According to the Google researchers, the training data is essential.
This finding is not surprising, as researchers have been lowering expectations of artificial intelligence since the release of ChatGPT. However, some experts still believe that AI is capable of developing “thoughts” that are different from human ones. However, such Artificial General Intelligence (AGI) is still only a hypothetical model.