How (not) to learn about AI with metaphors: And how to use ChatGPT as a metaphor generation assistant

The geneticist Steve Jones once remarked about metaphors and evolution that they go together like statues and bird droppings. This applies pretty much to any complex subject and AI is no exception. Metaphors are great at giving insights but they also let us get away with an illusion of understanding. This is a warning on some examples about AI.

TLDR;

Generated by Claude.ai with edits by me:

• Metaphors give an illusion of understanding AI unless accompanied by actual knowledge of how AI works.
• Common AI metaphors like "intelligence", "learning", "neural networks", "reasoning" mislead if taken too far.
• To learn from metaphors, actively seek where they break down. Make tables mapping source to target domain.
• Use multiple metaphors together to get a richer understanding. Single metaphors mislead.
• Ask ChatGPT to analyze and generate metaphors but don't take first suggestion. Drill down for more.
• Metaphors generated by ChatGPT are of mixed quality. But it will generate more suggestions than a person ever could in under a minute.
• Claude is a good complement to ChatGPT for generating metaphors. Others less useful.

How metaphors work: Mappings between domains of knowledge

What is metaphor? How does it work. Everyone knows it’s a sort of comparison, but what are we comparing and how? Metaphors get its work done through creating mappings between two domains of knowledge. For example, argument and war. We can say, they brought the big guns to the argument and everyone knows what it means because we can project from the domain of war onto the domain of argument.

Traditionally, people say that the source domain is more concrete or better known but that is not really true. How much more do we know about war than arguments? Is it really all that more concrete? Instead, the metaphor gives us a different way of thinking about the target domain through those mappings. Often, that leads to new ideas or better understanding.

But this requires that we have something in the target domain to map to the source. If we know literally nothing about the domain we’re projecting to, we cannot improve our thinking about it. We still only know about the source domain and have just learned some words to use about the target domain.

Take computer virus, if you know nothing about how computers work, you might think that simply placing two computers next to each other in the same room might spread the virus. But in fact, very little that is true about viruses is true about computer viruses. The person who invented the metaphor almost certainly knew much more about computer viruses than real ones.

Metaphors of AI

But most people who resort to metaphors about AI, know very little or almost nothing about the domain of machine learning or Artificial Intelligence. And any metaphors they learn will just give them an illusion of understanding unless they are also accompanied by some knowledge about how AI works that comes from outside the metaphor.

For example, many people have taken to repeating the metaphor of generative AI being just a fancy autocomplete. “All it does is predicts the next word”, they say. This also happens to be almost literally true, so it looks like the person has learned something about AI. But if your only experience with autocomplete is on your iPhone’s keyboard, you have learned nothing about AI. In fact, you are actively primed to make wrong inferences because ChatGPT is more unlike your iPhone’s autocomplete than it is like it.

The same goes for many of the metaphors baked into how we talk about AI. Here are some that we have forgotten are metaphors:

1. Intelligence in AI is not the same as intelligence as humans - we may actually know more about how machine intelligence works than how human one does. Many thousands of hours have been wasted on exploring this mapping of a metaphor that appeared in a grant proposal in 1956 in an effort to get money.
2. Learning in machine learning is not the same as a child learning - the machine is learning only in the sense that it is not purely calculating statistics on predefined parameters but it is not learning in almost any of the ways we know from our own experience. Plus there are many aspects of our learning we may not even realise are important. For example, humans are learning (as in their brain changing) just by existing, computers only learn when a learning algorithm is being triggered.
3. Neural network is not very much like a human neural network - it is based on a mathematical model of the neuron developed in the early 1940s. It is based on the principles of connections with different strengths that update as part of learning but that is very much it. Many people who just hear the words neural network conclude that it is based on how the brain works and thus expect human properties.
4. Reasoning in reasoning benchmarks. Most modern AI systems are evaluated based on benchmarks that test human reasoning. Therefore, we conclude that they must have somehow learned to reason. But if we want to call it reasoning, we also have to acknowledge that it has very few properties we would expect in a reasoning machine. Most notably deliberation, forward planning or backtracking.
5. Algorithms are another popular metaphor brought up when talking about AI. There are literal algorithms in play but what many people imagine when they hear AI algorithms are procedures for coming up with results that resemble step by step instructions. For instance, grammar rules or decision trees we know form expert systems. But no such algorithms are encoded in AI. In a way, algorithm is just a synonym for ‘reasoning’ or ‘intelligence’ but it brings with it many mappings that actively confuse us about how generative AI works.

How (not) to learn something about AI with metaphors

Two ways of not learning from useful metaphors

I wrote much more about (not) learning from metaphor on my blog about metaphors.

Source domain leaks: Intelligence, learning, neural net, reasoning - those are all good and useful source domains for helping us structure our thoughts about the domain of computers performing intelligent tasks. The problem is that they often leak mappings that are either irrelevant or misleading. For example, we only ever experience intelligence behaviour tied to some sort of intention. So, many people look for intentions in AI system even if we know quite well what their targets are.

So you always have to be extremely careful about what gave rise to the metaphor and at what point we are simply bringing over things from the source domain and desperately casting about for something in the target domain to map them onto. Metaphors are great if you are always actively seeking their breaking points. But they are also incredibly leaky ships, and it is better to let them sink with great frequency than constantly run around and try to plug the holes.

Mappings to nowhere: But what’s even worse, is when we have nothing in the target domain to map the metaphor to. Then we just begin inventing things about AI that have no bearing on reality and start expecting things to be true about them that are just not. We may not even be aware that they are metaphors and just think we are making literal statements.

For example, I constantly run into people who think that machine learning is just giving an “algorithm” a lot of data and it will “learn” the relationships using some sort of “machine learning” magic. They have no awareness of issues like identifying features, overfitting on training data, etc. They have no idea about how much data different approaches require and what the data is. Now, a lot of this is very technical and quite specialised information, but without at least some basic awareness of it, you can make almost no useful inferences about what machine learning can do.

There’s also the opposite variant of this, I’d call mapping from nowhere which is when we make assumptions about the source domain of the metaphor which leads us to running in circles and we end up accidentally creating metaphors of the the thing we’re trying to use the metaphor for.

This the case with the many uses of intelligence as a source domain for technology over the years. We don’t actually know that many things about intelligence, so it is easy to map from assumptions onto computers. For example, computers model our logical thinking. But because we don’t actually know exactly how our logical thinking works, we start assuming that our minds actually think using computer-like algorithms. Not because we have any first hand evidence of this but because we have created these mappings from nowhere and are now accidentally using the target domain to structure the source domain.

Two ways of learning from metaphor

I wrote much more about inferential learning on my blog about metaphor.

There are many ways to classify learning something. But for our purpose here, I’d suggest two levels:

1. Pre-inferential - you can only identify which domain something relates to but not say much meaningful about it that is not just repetition.
2. Inferential - you can make useful inferences about things that were not in your course or textbook - this can be at a very basic level or an expert level.

Metaphors can be great for inferential learning but they are also great at giving you the illusion of inferential learning. They are a trigger but not a shortcut. If you haven’t done the hard work, metaphors are great at keeping you on the surface.

Metaphors are great at helping you find some anchor for the new subject in your comfort zone. But is that’s where you stop, you’ve not learned much that is useful outside of a casual conversation.

Metaphor learning method 1: Mapping analysis

To generate new inferences with metaphors, you have to find out not-metaphorical things about the target domain. I suggest this simple principle:

If all you still cannot say something about AI outside of the metaphor, you have not learned anything through the metaphor.

But even more importantly, you should be able to say where the metaphor breaks and why. Make a simple table with 4 columns:

1. Something about the source domain
2. Something about the target domain
3. How they are similar
4. How they are different

It is useful to make those list independently and see how many of the items match. If you were honest, you will find that most things in the source domain have no equivalent at all in the target domain and those that do are only very imperfect matches.

If you can’t put anything in the target domain that means that the metaphor is not going to help you. If you can’t put at least three things in the source domain column, the metaphor is probably not very useful. You may also find that you will need to learn a bit more about the source domain to see what the mappings can do for you.

Metaphor learning method 2: Metaphor coupling

Metaphors are like guinea pigs, you should never just keep the one.

There’s a law in Switzerland that you can not own a single guinea pig because they pine without company. Metaphors are also like that. But instead of pining, they will leak harmful inferences all over your mind.

The biggest danger a metaphor will pose to you, if you let it crowd out other metaphors. They can be very seductive. and in the afterglow of the revelation you experience when you finally come up with a particular metaphor, you can easily start feeling like that’s the only metaphor you will ever need.

But because all metaphors are partial, you can never just have one. Try to come up with as many as you can. Make a list of any random domain you know something about and try to find some mappings. What you’re after is enriching your space of understanding, so rejecting many domains as sources of mappings is just as useful as finding new ones.

Using ChatGPT to help with metaphors

Large language models are extremely mixed when it comes to metaphors. They are unexpectedly good at generating them but their hit rate is relatively low. There are two problems that are present in about 30-40% of metaphors you ask :

1. The metaphors ChatGPT suggests make no sense, or
2. ChatGPT’s explanation of the metaphor is very partial

ChatGPT as metaphor analyser

Large language models are good at following on from examples. You can give ChatGPT or Claude some of the examples from this post and ask it to give you more.

See what happened when I did this in this chat about metaphors. Notice how I never accepted the first answer. What you can’t see in the shared chat is that I often regenerated the answer to have more possible candidates.

At some point, I had to step in and remind ChatGPT about some facts about LLMs and then tell it how to use them. When I did, it did a passable job. Not one I would be satisfied with but enough to start my thinking process.

ChatGPT’s strength here is the speed with which it can do this, not the quality. It would certainly take me a lot less time to fix the problems than to start from scratch.

ChatGPT as metaphor generator

Many (if not most) of the metaphors ChatGPT will generate will be completely useless. For example, one metaphor it seems to come up with repeatedly for LLMs is the Library of Alexandria. This not very good. But the point of this is that it won’t get tired and eventually, it will come up with a metaphor that you’d have never thought of that you can do something with.

It does not have to cover the whole of the target domain, just some aspect of it. ChatGPT is doing just the grunt work, you have to do the actual thinking after.

See here for a chat where I asked ChatGPT to generate metaphors about LLMs. Notice how I did not accept the first thing ChatGPT suggested. I asked for more and drilled down. Also, I did not just ask for metaphors, I specified what kind of metaphors I wanted and gave some examples.

Claude and others

Claude does not allow chat sharing so I cannot show you a full chat but it is not bad at this. I find it more to the point but slightly less good at following instructions than ChatGPT (using GPT4).

Here’s one of the metaphors it generated for me:

LLMs are like a guest at a cocktail party who can chime in on any topic based on snippets of overheard chatter. The longer the party goes, the more tidbits they accumulate to converse credibly.

As in so many ways, I found Bard and Bing to be also-rans. They are set to generate quite minimalistic answers. Bing keeps changing interface. I’m sure I could make them do more but so far I have found no reason to try. Bing did a search on useful properties of LLMs but it did not improve its answers. When it comes to metaphor, Claude and ChatGPT are the workhorses.

Conclusion

The last word generated by ChatGPT based on the text above:

Metaphors can be a double-edged sword in grasping AI. While they can anchor new ideas, they risk giving a false sense of understanding or misleading entirely. To navigate the world of AI, it's not just about finding the right metaphor but critically analyzing its limits. Tools like ChatGPT and Claude offer a hand, but remember: they're the starting point, not the finish line. Always dive deeper than the surface metaphor.