Four experts share their insights on current issues at the intersection of technology and society. In this edition: Vanessa Evers.

With ChatGPT, Claude and the like, people are creating and achieving all sorts of wonderful things. One person uses them to create practice tests for a grandchild, whilst another analyses news stories from various newspapers. It reminds me of the rise of Google. Whatever search query you entered, Google knew the answer. It was magical.

In 1992, I started studying at the University of Amsterdam and was suddenly given an email address. I had no one to email – until I spotted an email address on the cover of an introductory book on computer science: the author’s.‘Hello, I am a student at the University of Amsterdam and I have found your email address in my textbook.’ I received a prompt reply: ‘You’re the first person to email me from another country!’ That, too, was a magical experience.

Now we are facing a new, seemingly magical, development in artificial intelligence (AI). But what exactly is AI? Although I’ve taken courses in it and have been involved in numerous groups and projects in which the pioneers of AI also participated, I actually have no idea. My take on AI – machine learning, and specifically reinforcement learning – is now considered outdated. Machine learning is when the machine starts to recognise patterns in the data you feed it: lots and lots of images of horses and donkeys, neatly labelled, and after a while the software determines for itself when it’s likely to be a horse. Reinforcement learning mainly takes place in the real world. My drone wants to count elephants; it thinks it’s spotted an elephant and takes a photo. The forest ranger looks at it: ‘rhino – wrong!’ So the drone gets a penalty point. Do this a few times and it’ll figure out when it really is an elephant.

LLMs and the exciting new multimodal LLMs work differently. AI cannot actually ‘see’ a photo or ‘hear’ a voice, but only understands numbers. The Transformer shuffles text, image and audio tokens and looks for patterns – connections between images, sounds and text. If these relate to a yellow, curved piece of fruit, they end up as ‘banana’. Give a multimodal LLM a photo of the contents of your fridge and ask ‘what can I cook?’, and it will match the ingredients with recipes it has memorised from the internet. Magical.

So why are we so worried? That same AI can instantly create a hyper-realistic fake video. Things that used to be possible only for specialists with years of scientific training are now achievable with the right prompts. AI creates agents – systems that can not only chat but also use computers, make payments and take decisions.

If I ask AI to run my online shop in such a way as to maximise my profits, I have no idea what it will do. Perhaps it will hack the competition, or use the profits to manipulate the stock market. It depends on which agents it decides to create, and how good it is at doing so.

AI operates in a mathematical world. It doesn’t ‘think’ about whether a particular outcome is actually good for other shops like mine. By the time we finally realise that something is happening that we don’t want, it may be impossible to switch it off.

So… the magic of yesteryear, but with downsides that we are still far from fully understanding. There now seem to be infinitely more ways to misuse this technology. The ball is in each of our courts.

Professor Vanessa Evers, who specialises in social artificial intelligence, is director of the Centre for Mathematics and Computer Science (CWI) in Amsterdam.