The Chefbot Thought Experiment

The View from Oregon — 350: Friday 18 July 2025

A few years ago there was a big splash when the Moley cooking robot was introduced. I watched the videos at the time, suitably impressed by the dexterity of the machine. I believe that other cooking robots have been made available since that time. They remain quite expensive, so we’re not yet at a price point at which the mass market will bite (no pun intended). I mention this sophisticated cooking robot as I want to use it at the setting for a thought experiment. Suppose we combine a cooking robot like this with a large language model (LLM) trained on cookbooks and the lore of cooking. Let’s call our cooking robot with an haute cuisine LLM Chefbot. Assume that Chefbot has been trained on every existing cookbook, libraries of information on food generally, the lore of wine and spirits (for appropriate pairings with the meal), and even travel literature focused on food. (Eat Pray Love, anyone?) Further suppose that the LLM is programmed to entertain you and your guests by discoursing widely on cusine and food traditions while cooking you up a meal, sharing anecdotes about the traditions of the meal and literary quotations relevant to the moment. Maybe you can even choose the accent. I suspect many Americans would want to give their foodie LLM a French or an Italian accent. (I once spoke to a Frenchman who worked in the hospitality industry, and he told me that he would sometimes ham up his accent for effect.)

Suppose you and your friends are having your dinner prepared by Chefbot and enjoying all the conversation kept expertly moving along by Chefbot. What exactly are you doing? With what kind of a thing are you interacting? Have you established a relationship with a machine through the time-honored custom of breaking bread together — since even if the machine doesn’t eat, it’s a part of the food preparation, and thus part of the ceremony of shared food? Would the chefbot be a new kind of being? If it is, would human beings then establish a relationship with this new kind of beings, as we have established relationships with horses and cows, dogs and cats? Has the ontology of the world been enriched by something that did not previously exist? Is it a person? Is it a self? Or is it “merely” a machine? Are there many kinds of machines, or are all machines metaphysically indistinguishable?

Chefbot is, minimally, a new kind of machine. But how new? Is there anything unprecedented about it? Is it just another machine, or is it sufficiently different from what came before that we need to acknowledge that the Chefbot is a new form of being? At what point do we recognize a macroevolutionary bifurcation in which a new clade appears — in this case, a new clade of machines, or, you might think, a new clade of intelligence? We are presented with this dilemma with every new machine that is built, but some machines have more impact on human experience and society than others, and those with a greater impact are likely to be invested with a certain aura. Was there anything unprecedented about the smartphone when it was introduced? This was certainly a technology that transformed society to some degree. Individually, all the features of the smartphone had already been in use before the first smartphone was sold, but once the smartphone came into general use, it has been treated as a particular kind of device, as though it deserves its own taxon — the smartphone taxon. What was distinctive about the smartphone (and here I speak in the abstract because I don’t use one myself) was the integration of formerly distinct functions in a single package, and presumably the utility of the operating system, valued for its intuitive funtionality. Sometimes just putting things together in a novel way is felt to be sufficient to constitute a new kind of being, even if all the elements were already present. For this reason we often invoke the idea of something being “more than the sum of its parts,” meaning that the whole introduces something lacking in any one of the parts, or in all the parts taken together but not assembled in the same way. In other words, the whole is an emergent property.

The recent introduce of LLMs already have had a significant social impact, like the introduction of the smartphone. We can only wait to see if they have a transformative influence on society and economics and so will represent a change equal to or greater than the smartphone. Certainly the hype around LLMs has been feverish, and this has led to a lot of precipitate judgments concluding that the world has already been transformed by them. Many people who have interacted with LLMs have thoroughly embarrased themselves by asserting that they “felt” a presence in the machine’s responses, and that something new had been introduced into the world with the advent of chatbots that can pass the Turing test. It doesn’t take much to fool some people, which is why P. T. Barnum famously said that a sucker is born every minute.

The possibility of Chefbot not only poses the problem of the possibility of a new kind of being — is it or is it not? — but also the problem of the new things that Chefbot itself creates. Suppose you’ve had your share of predictable dinner parties with excellent but predictable food. Instead of the ordinary, you ask Chefbot to surprise you. Chefbot asks you (heavily accented in French), “Do you mean zat I should choose at random ze dish I’ll prepare for you and your guests?” And you respond, “No, Chefbot, I want you to create your own recipe and prepare a meal for us that has never before been prepared. I want you to do something completely new.” You’re asking Chefbot to prepare something novel, but, ironically, everything that happens next will be a result of Chefbot’s programming. Whatever Chefbot whips up on a mechanical whim, perhaps drawing on recent conversations that revealed your preferences, and mixing this up with a number of other influences, perhaps from dishes that had elicited your approval in the past, maybe even relying on a randomness generator to add in a few unexpected dashes, will be the result of Chefbot’s programming, and not on creativity.

Chefbot is a machine, and has no creativity. Chefbot may sound conscious and self-aware, but this is an imitation of how a conscious being interacts with other conscious beings — the appearance and not the reality of consciousness. Along with the appearance of consciousness, Chefbot has the appearance of creativity of the sort possessed by (some) consciousness beings. But this creativity is only the result of a complicated alogorithm that mixes together familiar items in unfamiliar ways. However, as we’ve seen with the smartphone, putting together extant features into a single package — be a phone or a meal — can be something sufficiently new that possesses emergent properties. If Chefbot creates a new kind of dish, or even a new kind of cuisine — Chefbot cuisine — is this an emergent property, and, if it is, is this emergent property a marker of a new form of being in the world? If Chefbot is not itself a new form of being, can it create a few form of being?

We could even add a feature to Chefbot that mimics human mastication of food and the sensors of the tongue — the appearance and not the reality of eating — so that Chefbot could “taste” his own creations. If we wanted something more gruesome, we could genetically design a head that was just a nose and a mouth, allowing this head to eat, and connecting the tongue and nerves of the mouth and nose as a computer input, the results of which could be hooked into Chefbot and integrated with Chefbot’s programming. In this way, Chefbot could compare the responses of human beings to the dishes he prepares, building on successes and editing out the failures. But even with all this hardware (or wetware), Chefbot wouldn’t know what anything tastes like, because Chefbot doesn’t know anything, and doesn’t taste anything. Chefbot does not understand food. It doesn’t “understand” anything at all, because it has no understanding. Chefbot doesn’t “know” what foods taste like; it doesn’t even know what a taste is. For Chefbot, “taste” is a word used in certain predictable ways, and if it uses the word in predictable ways, the proper result (however this is defined in Chefbot’s programming) will be obtained. Chefbot doesn’t even know what a “word” is, because a word in natural language must be translated into machine language by a compiler to direct the requisite levels of electrical signals to its electrical components. Chefbot can’t taste. Not only can it not taste, it can’t formulate the concept of taste, because it doesn’t formulate any concepts at all. Because the machine doesn’t think or sense, it can’t even imagine taste by analogy with some sense that it does possess (as we might try to imagine sensing by the pits of a pit viper or the electroreception of a shark), because it doesn’t possess any senses and it doesn’t imagine anything.

You wouldn’t have to have a Chefbot to create chance recipes. We know that cooking can be reduced to a sequence of discrete steps, which is why he have cookbooks, which, as it turns out, are a significant segment of the overall book market. (There are book stores that sell nothing but cookbooks.) The very idea of a recipe is the informal equivalent of an algorithm, and what distinguishes an algorithm is that it can be mechanized. A simple procedure would suffice as a substitute for Chefbot: write on one set of slips of paper the names of ingredients, write on another set of slips of paper quantities, and write on a third set of slips of paper prodecures employed in cooking. Put these three sets of slips of paper into three bowls, scramble the slips together, then remove one slip of paper at a time from each bowl in turn, assembling them into a recipe. If your aleatory method results in a recipe with a cup of salt and a teaspoon of flour, the result probably won’t turn out well. You need an editor simply to scratch off the unworkable combinations.

This is the role that death plays in natural selection — removing the unworkable iterations so that the workable iterations can have a chance in the biosphere — and this is the role that an LLM would play in the culinary creations of Chefbot. However, you could achieve the same through a reasonably short set of rules of thumb for cooking to which any recipe must conform. Certainly the mechanized processes of Chefbot would be quicker, and it might be a lot more fun, but it’s not essentially new, and no new form of being has been created. And this substitute procedure for Chefbot is the reduction of Chefbot to the equivalent of Searle’s Chinese room thought experiment, so my argument here isn’t new either.

This thought experiment won’t remain a thought experiment for long. All the technology to realize what I have described already exists (with the exception of the gruesome example). It’s merely a question of money and motivation as to when Chefbot is built. But the questions I’ve tried to highlight here are fundamental philosophical questions, and they won’t go away any time soon. Even in a future in which there are thousands of Chefbots in thousands of homes, cooking up elaborate dinner parties and entertaining the guests with witty repartee, perhaps surprising all present with a new and unprecedented dish, the philosophical problems will remain. We can even imagine a Chefbot-invented dish becoming a contemporary classic, which ends up in future cookbooks and becomes part of the standard repertoire of cooking robots.