»Sapere aude!«

The current hype surrounding Al tools such as ChatGPT can easily obscure that AI already has a long history. Not only in science but also in business, medicine, administration, and, last but not least, in most people's everyday lives, AI technologies have already become firmly established. In this interview, Prof. Dr Clemens Beckstein explains what characterizes generative AI tools such as ChatGPT and why they rely primarily on human intelligence. The computer scientist also reveals what Immanuel Kant can teach us about handling AI.

Interview by Ute Schönfelder

You're a Professor of Practical Computer Science with further specialization in artificial intelligence. How would you define the term »intelligence« for your specialist field?

As a computer scientist, the idea of trying to define »intelligence« would never cross my mind. But I nevertheless believe that the term »artificial intelligence« can be meaningfully used as a synonym for digitalizing areas in society, business, culture, and science that previously operated without computer support and that are so complex that they are assigned the attribute »intelligent«.

According to this »definition«, AI is an ascriptive term: its meaning changes in line with the progress we are making in constructing digital, programmable systems. What was called AI 20 or 30 years ago, such as voice assistants and robots, differs from what we call AI today. In this sense, AI is a moving target.

What type of AI does your work focus on?

For the most part, it is classical, symbolic AI , such as symbolic knowledge representation and processing, intelligent computer support of the scientific research processes, and algorithmic network analysis, but also research on foundational and philosophical aspects of AI. Currently, my team and I are focusing on using AI technologies to digitally model the research process in the humanities. For this purpose, we collaborate with partners from the humanities through a cross-faculty research team called MEPHisto (Digital Models, Explanations and Processes in the Historical Sciences).

How do AI tools like ChatGPT actually work?

There are now very many very different AI systems, and they all work differently. ChatGPT is a specific software tool that associates texts, a so-called generative pre-trained transformer (GPT). Grossly simplified, such a transformer is nothing but a gigantic, algorithmically compressed association table consisting of two columns of text. One column lists all conceivable questions, or »prompts«, while the other lists the most suitable responses for those prompts.

A huge amount of publicly available digital texts is processed to create this table. For a conversation with ChatGPT, this model computes the probability with which any theoretically possible word should be the next or final word uttered by ChatGPT in this conversation. This statistical model is called a large language model (LLM) or a generative pre-trained transformer (GPT).

So, how intelligent is ChatGPT?

Only as intelligent as the humans who develop, fine-tune, and use the model. In and of itself, ChatGPT is not intelligent at all. It does not have a mind or even a memory of its own and cannot learn anything autonomously.

But ChatGPT does possess a great deal of human intelligence: the digitalized, written cultural heritage of humankind, as well as vast quantities of data that people have scraped and filtered from the internet and every other digitally available source to construct and fine-tune the language model of ChatGPT. Human intelligence can also be found in the underlying technology and in the public institutions that work hard to ensure a socially compatible alignment and responsible use of ChatGPT by establishing ethical and political frameworks.

It is only through people that ChatGPT becomes what it is: an exceptionally literate, extremely easy-to-train »stochastic parrot«External link that can speak with polish—but does not really understand what is being said.

Nevertheless, ChatGPT in particular has advanced leaps and bounds in the last year. How would you assess its progress to date?

The reason for the progress achieved recently, especially with respect to generative AIs, is not so much technological breakthroughs but rather scaling phenomena. The progress is mainly due to a huge increase in resources used to train these systems: ever larger and better optimized artificial neural networks and ever more powerful hardware for their simulation, together with a gigantic volume of data from almost all areas of our life, which are primarily collected by the four big players: Microsoft, OpenAI, Meta and Google. Exactly how long this scaling will continue to produce significant improvements in AI systems remains to be seen.

How reliable is the output from ChatGPT and the like?

At the moment, it actually isn’t that reliable. As I’ve outlined, the output from ChatGPT is nothing more than statistically founded statements. Its hit ratio is currently around 80%. That isn’t going to change significantly because, like all machine-trained systems, ChatGPT still only has incomplete knowledge of the »right« output and the »right« behaviour. Everything depends on the data used to train it. Another issue is that its output cannot be explained or validated and often enough contains prejudices and biases.

The problem here is not only that the model lacks the technical ability to distinguish between »right« and »wrong« behaviour but also that these categories are socially negotiated and often ambiguous. Strictly speaking, Microsoft, OpenAI, Meta, and Google currently decide what is »right« and what is »wrong« because they control the data used to train foundation models like ChatGPT and set the standards for fine-tuning which is supposed to eliminate prejudices and biases.

Where do you think regulation is needed in relation to AI?

It’s the same for AI as for dogs: whether big or small, aggressive or toothless, the biggest risk with dogs is at the other end of the lead: the human controlling them. That is to say we as a society are responsible for ensuring that AI is used responsibly and accountably in politics and business, and by each and every individual.

I am, therefore, rather disappointed by the current European proposal on AI regulation because it omits a lot of things. It is not the technology that is being regulated, but the application of AI. It is obviously a good thing that mass biometric surveillance or »social scoring« are to be prohibited. But there has actually been consensus on these issues for a long time. What I think is missing from the current draft law, however, is clear regulation of the foundation models themselves, in particular transparency regarding the data used.

Is there also a need for development on the part of users?

In my opinion, this area has the greatest need for development—at least if we do not want an unconsidered use of AI to unintentionally (but culpably) reverse humanity’s emergence from its self-inflicted immaturity, mindlessly reversing more than 200 years of enlightenment!

This is not about having to acquire or develop new skills. Instead, we simply can remember the advice of the great Enlightenment philosopher Immanuel Kant, who wrote: »Sapere aude! Have the courage to use your own understanding!« This is something we should do consistently, despite ChatGPT and the like!