AI gedachten | thoughts

Around 2018, the first group of children was born that have learned to read and write in a world where ChatGPT was already present. For them, a chatbot is probably something quite normal. A few of them will be amazed: “how is this possible?”: these are the ones that will study computer science or AI. The others just use their chatbots.

Something that already exists when you grow up is taken as a given. Nobody else is surprised or amazed. Many children will ask at one point, “why does the sun rise every day?” but only a few of them are so amazed that they want become astromers.

Growing up with a technology that already exists – you learn how it works because you see how others work with it. This helps you to understand it intuitively – this is also called a ‘mental model’. It does not have to be a technically correct model: for a long time, sunrise has also been understood the wrong way (sun revolving around the earth) but it still helped us to have a basic understanding of the phenomenon.

The technology generations idea indicates that new technology that is adopted through trial and error by society, but younger people who have experienced it while growing up will be able to handle the technology naturally and effectively.

The digital skills of the age group 16-24 are just as high as those of 25-34, and older groups are progressively less skilled. Apparently, around 2023, when Eurostat did the research, the youngest generation was no more skilled than the generation just above it. That latter group is older, but has still the same high level of digital skills; they were born around 1990 and were adolescents when the internet ‘exploded’ with social media. They are usually called ‘digital natives’, as opposed to the older ‘digital immigrants’. Two generations with similar skills probably means that the internet has not changed significantly after the first digital natives group. Note that 2023, when this research came out, also indicated the launch of ChatGPT!

When something new arrives in society, things are different. Everybody has the same questions: where does this come from, how does it work, and what can you do with it? People who know the ansers will explain them to the others. Scientists, or, in earlier times, prophets. Comets that arrived in the heavens above were frequently explained as the end of times.

The way young people develop a mental model resembles how children learn their mother language: by observation, trial and hardly without any explanation. Learning a language at a later age involves learning grammar rules, idioms and a lot of guidance and practicing. Indeed: that is also the difference between data-driven AI systems and knowledge rule-driven AI systems, as described in my previous blog. We may remember the PC courses that our (grand)parents took when in the 80s-90s a lot of private PCs were introduced. Always read the manual first!

Mental models and literacy

It’s good that there is attention to AI literacy lately. The AI Act of the European Commission requires that who works with AI, should have a basic understanding of what it can, and what it cannot do. Literacy is a term that is sometimes understood as just being able to read and write. But that is not enough: when you read, you have to be able to understand what you read, and when you write, you have to be able to build a more or less logical and understandable message.

So, literacy is more than just the outside. AI literacy has not yet been defined very precisely, but for me, having a good mental model of AI is one of the most important aspects. In the usability research field (mainly about apps and webpages), the importance of a good mental model has long been recognized. They have to be based on earlier and comparable experiences, and must help the user to predict what is about to happen. A shopping cart icon (that receives your orders) or a magnifying glass (for search) are examples. Metaphors from the physical world help, too. That is why the ‘save’ icon has been a floppy disk for a long time. When younger users started to comment on an actual physical floppy disk with “hey, cool, a 3D printed icon!”, it was about time to change that icon.

The way a piece of technology is presented can clearly help to support the mental model.

For chatbots, this does not work. They are trained to sound human, convincing and reasonable. The mental model ‘this is a computer that lives in a datacenter somewhere out there’ is not supported – people will develop a personal bond to that thing instead. It also does not help that we use words that make chatbots seem even more human – I have written before about ‘thinking’, but what should we make of ‘personality’ or ‘soul’? There are suspicions that the Pentagon has indicated Anthropic’s chatbot Claude as a ‘supply chain risk’, because it allegedly was equipped with with a ‘soul’, and a constitution, too – which is of course not the American constitution!

No matter how intimidating technology can be, for literacy it is important that we’re not afraid to look under the hood at least once. The drivers of the first cars had to be mechanical engineers in order to keep it running (they needed to have a full-fledged toolbox with them), but that is not needed anymore. Even the most non-technical driver knows that a fuel tank can run empty and needs refilling, that tires can run flat and that the windshield washing fluid must be topped up from time to time.

What should an AI mental model look like?

AI is highly abstract. Everything is hidden behind a screen and you cannot see what’s behind it. Building a mental model for internet was already a challenge (“my google does not work” when there is no wifi) but for AI it’s even harder. This makes a mental model even more important. It helps when it is more or less technically correct, but the most important part is that it helps you to understand what happens.

An example: you sometimes read that LLMs are autocomplete on steroids. Technically that could be defendable, but it has limited contribution to your understanding what an LLM can or cannot do. A car is a bunch of controlled explosions on four wheels, and chess is moving pieces of wood over a board with 64 fields.

A metaphor can be oversimplified.

On the other hand: magical thinking also does not help. As science fiction writer Arther C Clarke famously noted, “every sufficiently advanced technology looks like magic”. Larry Tesler added to that in 1970, “AI is what has not yet been done”. (To which Benedict Evans added, when it has been done, it turns out to be just software.)

That is why I am writing this series of blog posts on AI, each time comparing it with technologies or other familiar phenomena. There are surprisingly many analogies that make AI easier to understand for anyone who wants to become more literate in the topic.

I have made an overview of all analogies that I’ve found so far. You can see that somehow it is useful to draw parallels with cars and everything related to that. Another analogy (which I think I can dive deeper into soon) has to deal with food production. Sometimes analogies from topography are useful, or historical comparisons.

The importance of literacy

Literacy is not only important when you want to work smoothly with new technology. It also helps to avoid trust technology too much, or too less. It helps to understand what is real and what could be fake, like fake news or all techniques used by cybercriminals. And it helps to choose the right type of AI for your job: AI is not one thing, but a fleet.

Learning is done by trial and error. Silicon Valley has translated that to “move fast and break things”. With powerful technology like AI, I’m not sure that is the best of ideas. That gigantic focus on speed can be understood from an economical perspective, but speed does not help to build up a good mental model. I hope that later, we can look back at this period like we can now look at the internet bubble era: when you are in the middle of it, everything is changing super fast and the whole world is confusing, but later on, a few years later, all mental models have settled and we can merrily jump onto another hype.