Exquisite corpse is a collaborative drawing or writing game invented by Surrealist artists in the early 20th century. The game involves a group of people creating a drawing or piece of writing together, but without seeing each other’s contributions.

The game typically involves folding a piece of paper into sections, with each person drawing or writing in one section, and then folding the paper to reveal only the last portions of their work before passing the paper on to the next person to continue the process. At the completion of the process, the final composition is revealed for all to see. The end result is a collaborative work that is often unexpected and surreal, as each person’s contribution is influenced by the preceding one, but is not fully aware of what has come before.

The name ‘Exquisite Corpse’ comes from the French phrase ‘cadavre exquis’, which was the original name given to the game by the Surrealists. It is said that the name was inspired by a phrase that was generated randomly using the game: ‘Le cadavre exquis boira le vin nouveau’ (The exquisite corpse will drink the new wine).

While a lot of focus on generative AI, especially large language models (LLMs) such as ChatGPT, GPT-4, and LLaMA, is about its accuracy and avoiding AI “hallucinations,” the artistic or creative aspects of such should not be overlooked. Creative collaboration between human creators with generative models is an open area for exploration. The very fact that an AI like ChatGPT can “make stuff up” is as fascinating as it can be perplexing. The ability to imagine, to “make stuff up,” is, after all, not also one of the hallmarks of human intelligence?

The practice of exquisite corpse was one of the ways the early surrealists attempted to get “out of their heads” to discover new ideas for their work. They courted randomness and wanted to allow their own unconscious minds free play in their art (fascinated, as they were, with Freud’s work on the role of the unconscious in the human psyche). For all their surprising capabilities, at bottom a large language model does one thing really well: given some prompt, it autoregressively predicts the next sequence of tokens (or words) completing that text. Derogatorily, but yet also accurately, such models have been called “stochastic parrots.” I suspect that, were they to be alive to day, the surrealists would have found generative LLMs then fascinating — if only to collaborate with a non-human largely stochastic process (and intelligence).

Playing With ChatGPT

In that vein, I’ve thought it may be fun to try making poetic exquisite corpses, with ChatGPT as my collaborator. Here we can give AI also free reign to “make stuff up.” Will its contributions be deemed “creative”? To experiment with that possibility, I wrote an application which manages and virtually folds a virtual page, if you will, and solicits the AI’s contributions to the process (via OpenAI’s chat completion API).

For example, I began with a few lines, more or less as a kind of stream of consciousness or free association (as best I could at the moment). I then clicked on the “Fold” button on the app. This virtually folded the page, and prompted the AI with only the last line of my contribution (“of what to be foretold”), prefaced by the system message with a set of instructions (a form of ‘zero-shot learning’ AI terminology, it that it is not given examples):

We are playing a surrealist writing game, called "exquisite corpse."
We understand the game, so you are not to discuss the topic.
Your role is as follows:

You are given the last line of a poem, in free verse.
You don't know what came before it.
You are then asked to continue the poem, for at least five lines.
Your poem can be incomplete. Be imaginative!
Return only the lines of poetry you compose, nothing else.

The AI returns some lines. However, the page is again virtually folded (by the application), so I am shown only ChatGPT’s last line. (“Even impossible things are a”.)

As this surely is woefully incomplete, I add a few lines, fold the page again. The AI returns a further contribution to the collaboration, but again I only see the last line: “But our hearts beat with a rhythm of its own.” Feeling as there may be now a complete thought, I click on the “Reveal Poem” to display the complete composition.

One of the interesting aspects is that, in the final result, it may actually be hard to remember where mine and the AI’s contributions begin and end. It becomes one, seamless whole. Great poetry? That I will not venture to judge. But it does feel to be an interesting collaboration between myself and the artificial intelligence. The results are surprising. And fun. (Need we always take the “AI Revolution” always so seriously, after all?)

In another example, I have kept track of my and its contributions to another exquisite corpse:

How it works

We are actually relying on a couple of particular characteristics of ChatGPT (or other large language models). One, of course, is it’s capacity to hallucinate or “make stuff up.” According to a Wikipedia article on the topic:

“OpenAI‘s ChatGPT, released in beta-version to the public in December 2022, is based on the GPT-3.5 family of large language models. Professor Ethan Mollick of Wharton has called ChatGPT an ‘omniscient, eager-to-please intern who sometimes lies to you’. Data scientist Teresa Kubacka has recounted deliberately making up the phrase ‘cycloidal inverted electromagnon’ and testing ChatGPT by asking ChatGPT about the (nonexistent) phenomenon. ChatGPT invented a plausible-sounding answer backed with plausible-looking citations that compelled her to double-check whether she had accidentally typed in the name of a real phenomenon.”

In other words, given a non-existent topic, the AI gives a non-existent elaboration of it. Given the context, it seeks to complete the text. Lacking context, it will naturally fill in the void by doing what it does best: autoregressively predicting one token after another. One prompt engineering approach to control this behavior is to instruct the model that if it does not know something, simply reply that it does not know the answer. However, we are instructing it to, in lieu of the fact it does not know the poem thus far except the last line, simply fabricate a response. While “By 2023, analysts considered frequent hallucination to be a major problem in LLM technology,” we are counting on it! Because in this application, we want it to be a surrealist. And surrealists cultivate hallucinations.

Does this mean we will encourage it to cultivate bad behavior? No, because of another aspect of the model: it is stateless. The model does not retain states from one request to the next. To keep it on task, we have to include the prompt instructing it with each prompt. This seems it’d be a shortcoming in terms of LLMs ability to perform in context or 0 to few shot learning, in that learning does not affect the model’s internal weights. Unlike when pretraining or fine tuning of the model, it does not learn anything new in a long term sense. Similarly, to maintain a continuous conversation, we prompt it with the entire conversation that has come before (within the permitted context window).

But here, we are again counting on that! We do not want it to recall what has come before in the exquisite-corpse, just as if a new participant is handed the folded paper: they only see the one line of text or a few lines of the drawing without context. We are asking the AI to respond without immediate context. We are not encouraging it to develop bad behavior. But given its rich exposure to language, we simply want it to write.

The application and source code can be found on my GitHub repository.

This introductory portion of this article was also written in collaboration with ChatGPT, namely its provision of the exquisite corpse practice’s art historical context (any inaccuracies are, of course, its own).