I was on a panel on “AI” yesterday (was in German so I don’t link it i this post, specifics don’t matter too much) and a phrase came up that stuck with me on my way home (riding a bike is just the best thing for thinking). That phrase was

AI systems are just tools and we need to learn how to use them productively.

And – spoiler alert – I do not think that is true for most of the “AI” systems we see sold these days.

When you ask people to define what a “tool” is they might say something like “a tool is an object that enables or enhances your ability to solve a specific problem”. We think of tools as something augmenting our ability to do stuff. Now that isn’t false, but I think it hides or ignores some of the aspects that make a tool an actual tool. Let me give you an example.

I grew up in a rural area in the north of Germany. Which means there really wasn’t a lot to to TBH. This lead to me being able to open a beer bottle with a huge number of objects: Another bottle, a folding ruler, cutlery, a hammer, a piece of wood, etc. But is the piece of wood a tool or is it more of a makeshift kind of thing that I use tool-like?

Because an actual tool is designed for a certain way of solving a set of problems. Tools materialize not just intent but also knowledge and opinion on how to solve a specific problem, ideas about the people using the tools and their abilities as well as a model of the problem itself and the objects related to it. In that regard you can read a tool like a text.

A screwdriver for example assumes many things: For example about the structural integrity of the things you want to connect to each other and whether you are allowed to create an alteration to the object that will never go away (the hole that the screw creates). It also assumes that you have hands to grab the screwdriver and the strength to create the necessary torque.

I think there is a difference between fully formed tools (like a screwdriver or a program or whatever) and objects that get tool-like usage in a specific case. Sometimes these objects are still proto-tools, tools on their way of solidifying, experiments that try o settle on a model and a solution of the problem. Think a screwdriver where the handle is too narrow so you can’t grab it properly. Other objects are “makeshifts”, objects that could sometimes be used for something but that usage is not intended, not obvious. That’s me using a folding ruler to open a beer bottle (or another drink with a similar cap, but I learned it with beer).

Tools are not just “things you can use in a way”, they are objects that have been designed with great intent for a set of specific problems, objects that through their design make their intended usage obvious and clear (specialized tools might require you to have a set of domain knowledge to have that clarity). In a way tools are a way to transfer knowledge: Knowledge about the problem and the solutions are embedded in the tool through the design of it. Sure I could tell you that you can easily tighten a screw my applying the right torque to it, but that leaves you figuring out how to get that done. The tool contains that. Tools also often explicitly exclude other solutions. They are opinionated (more or less of course).

In the Python community there is a saying: “There should be one – and preferably only one – obvious way to do it.” This is what I mean. The better the tool, the clearer it’s guiding you towards a best practice solution. Which leads me to thinking about “AI”.

When I say “AI” here I am not talking about specialized machine learning models that are intended for a very specific case. Think a visual model that only detects faces in a video feed. I am thinking about “AI” as it is pushed into the market by OpenAI, Anthropic etc.: “AI” is this one solution to everything (eventually).

And here the tool idea falls apart: ChatGPT isn’t designed for anything. Or as Stephen Farrugia argues in this video: AI is presented as a Swiss army knife, “as something tech loves to compare its products to, is something that might be useful in some situations.“

This is not a tool. This is not a well-designed artifact that tries to communicate you clear solutions to your actual problems and how to implement them. It’s a playground, a junk shop where you might eventually find something interesting. It’s way less a way to solve problems than a way to keep busy feeling like you are working on a problem while doing something else.

Again, there are neural networks and models that clearly fit into my definition of a tool. But here we are at the distinction of machine learning an “AI” again: Machine learning is written in Python, AI is written in LinkedIn posts and Powerpoint presentations.

Tool making is a social activity. Tools often do not emerge fully formed but go through iterations withing a community, take their final shape through the use by a community of practitioners and their feedback. All tools we use today are deeply social, historical objects that have embedded the knowledge and experiences of hundreds or thousands of people in order to create “progress”, to formalize certain solutions so we can spend our brain capacity on figuring out the next thing or to just create something beautiful or fun. Our predecessors have suffered through proto-tools and all the hurt that comes from using them so we wouldn’t have to. And this social, temporal context is all part of a tool.

And the big “AI” systems that supposedly are “just tools” now do not have any of that. They are a new thing but for most problems they hope that you find ways of using them. They do in a way take away hundreds of years of social learning and experience and leave you alone in front of an empty prompt field.

So no, I do not think that the “AI” systems that big tech wants us to use (and rent from them) are tools. They are makeshifts at best.