Thilo Stadelmann explains what artificial intelligence really is

A tool, not a being. ChatGPT sounds like a human, but it isn’t one. In his talk, Thilo Stadelmann explains: how we can understand AI, using simple illustrations of how AI works, where the fear of ‘overpowering machines’ comes from, and why he remains optimistic nonetheless.

In short:

• AI does not understand; it calculates probabilities. It delivers what statistically fits best.
• Stories of machines taking control of us are based on worldviews, not on today’s technology.
• The risks are real, but manageable. The future is open and can be shaped.

Thilo Stadelmann is a professor of Artificial Intelligence and Machine Learning at the ZHAW in Winterthur. In a four-part lecture series attended by around 500 participants from the business sector and local authorities, he put the topic into context.

He begins with two examples. The first comes from Google DeepMind: from a single sentence, such as ‘helicopter flight over a cliff with a waterfall’, a photorealistic 3D world is created in which one can move about freely. The second is a riddle that Stadelmann posed to a chatbot: a doctor, who is explicitly stated to be the boy’s father, says he cannot operate on the boy because the boy is his son. Who is the boy’s doctor? The chatbot replied: the mother. Any child would have known the answer. How can a system be so impressive and so far off the mark at the same time?

AI simulates behaviour; it does not think

Stadelmann explains that AI has little to do with intelligence. It mimics intelligent behaviour, just as a computer can simulate the weather without it actually raining. AI is not a single technology, but rather a toolbox containing a vast array of different methods.

The most successful method in recent years is called machine learning. The principle is this: you show the computer a vast number of examples, such as photos of cats and dogs. It adjusts its internal ‘controls’ until it can reliably distinguish between the two, and ultimately produces a probability. Experts refer to this as a probability density function. Large language models (LLMs) such as ChatGPT are based on the same principle. They have learnt from vast amounts of text which word is likely to come next, and construct their response word by word. When asked what river flows through Asunción, the system adds, step by step: ‘the Río Paraguay’.

This also explains the error in the riddle. The standard version appears thousands of times on the internet, almost always with the answer ‘the mother’. The system has learnt this frequency and follows it, even though the question had been modified. It knows neither truth nor facts, says Stadelmann, only statistics.

This leads to three conclusions:

• Not neutral: The systems have absorbed everything available on the internet, including prejudices.
• Always fallible: Errors – so-called ‘hallucinations’ – are inherent to the system. Yet the AI still sounds convinced even when it is wrong.
• Different errors to those made by humans: Sometimes it fails at simple tasks, other times it succeeds at difficult ones. That is why checks are often carried out in the wrong place.

His analogy: AI is to humans as a DJ is to a musician. With dance music, you can hardly tell the difference, yet the DJ mixes recordings whilst the musician plays an instrument. With a romantic ballad, the difference becomes apparent. Similarly, a chatbot is a ‘person simulator’ but not a person, just as a flight simulator is not an aeroplane. According to Stadelmann, one reason we tend to forget this is that providers have a financial interest in us perceiving AI as something human-like.

Where do these images of all-powerful machines come from?

If technology can be explained in such a matter-of-fact way, why do the media so often depict scenarios like those in the film *The Matrix*? Stadelmann replies: These stories are not based on today’s technology, but on worldviews that are prevalent in Silicon Valley. He cites three ideas:

• Humans and computers are essentially the same thing, namely information processors (rationalism).
• Consequently, a computer will eventually be able to do everything a human can do, and thereafter improve itself (singularitarianism).
• Humans should therefore become less human and more machine-like (transhumanism).

These are articles of faith, not scientific findings. Researchers summarise them under the acronym TESCREAL. Stadelmann considers this intellectual triple leap to be ultimately inhumane, because it denies humans value and dignity.

He contrasts this with another perspective: the biblical view of humanity, which served as the basis for the declaration of universal human rights. Human beings are ‘marvellously made’ (Psalm 139), created in the image of God and loved by God. From this perspective, technology is a tool given to humankind to bring about good. On this basis, the same technology leads to a completely different outlook on the future.

How realistic are these scenarios?

According to Stadelmann, the only technical argument in favour of ‘superintelligence’ (Artificial General Intelligence, AGI) emerging in the near future is the so-called empirical scaling laws of machine learning. This refers to the observation that AI models have improved as more computing power and data have been applied. He illustrates this using an image model: with 32 times more computing power, a blurred image becomes razor-sharp. However, he argues that this is an observation rather than a law of nature, and in his view, the effect has been levelling off for the past two years. The idea that machines can improve themselves to the extent required remains, he says, an assumption that no one has yet observed.

Computer scientist Arvind Narayanan from Princeton University also takes a sober counter-view. He describes AI as ‘ordinary technology’, comparable to electricity or the internet: powerful, but it takes decades to become established. Stadelmann also advises people to listen critically to hype and scaremongering, as providers profit from making grand promises.

He also urges people to remain calm when it comes to jobs. AI automates individual tasks, not entire professions. Years ago, it was claimed that radiologists would become redundant because AI is good at analysing X-ray images. Today, according to Stadelmann, more radiologists are needed than ever before, as their work involves far more than simply looking at images: they discuss findings, talk to patients and consult with colleagues. It is this ‘in-between’ work that makes people valuable.

The assessment of the risks comes from the fourth lecture: Gehen Glaube und KI zusammen? Künstliche Intelligenz und Gemeinde (Einführung 4/4)

A huge thing, but no big deal

There are certainly risks, says Stadelmann: dependence on technology and providers, discrimination caused by biased data, uncertainty in the labour market, and the loss of skills and autonomy. However, he does not see these as insurmountable hurdles if the issues are tackled in a targeted manner. These can be kept under control through technical solutions (such as ‘algorithmic fairness’), new habits and clear rules.

His closing analogy is that of a caterpillar which pupates and emerges as a butterfly. For the caterpillar, this is a huge deal, but it is nothing out of the ordinary, and so far every caterpillar has managed it. So it should be with AI: opportunities and risks, benefits and mistakes go hand in hand and should both be kept in view at the same time.

And the future is wide open. The silliest thing would be to withdraw in discouragement and leave shaping it to those who, above all, want to make money. Stadelmann invites us to get involved, including with a Christian voice: if that is missing, something is missing.

Find out more: Wie können wir KI verstehen? Über Technologie, Weltanschauung, und den Menschen (Einführung 1/4)

Book Thilo Stadelmann as a keynote speaker: +1 (704) 804 1054 oder thilo.stadelmann@premium-speakers.com