Prompting as creative writing

This article is in collaboration with

The Models project combines artificial intelligence and Commedia dell’Arte (Italian comedy), reinterpreting traditional masks through advanced language models and the use of the Leonardo supercomputer. The audience can intervene and explore the dialogue between human creativity and technological potential in an innovative and interactive theatrical experience thanks to modular prompts and evocative scenarios

Ask an AI assistant: «What is the capital of France?». With confidence, it replies: «Paris». You trust it because the answer is correct and consistent with the role of a competent and knowledgeable assistant. But what happens when, with the same confidence, the assistant states something false, such as that France is still the reigning football world champion? This type of error is neither a simple technical malfunction nor a deliberate attempt to deceive¹, but rather the result of a more complex dynamic. It is precisely this dynamic that the interactive audiovisual installation The Models, created by the artistic duo dmstfctn, seeks to explore. Curated by Sineglossa and part of The Next Real programme, the project brings digital masks to life, animated in a 3D engine, and capable of improvising a series of 14,000 theatrical sketches in the style of Commedia dell’Arte.

These masks are powered by artificial intelligence models, fine-tuned using the Leonardo supercomputer², the Italian node of the EuroHPC European high-performance computing network. The art of crafting prompts lies at the core of The Models, which the duo employs as a form of creative writing to expose the less visible, problematic, and absurd aspects of generative artificial intelligence. The installation, The installation realised in collaboration with CINECA and ART-ER, under the patronage of the Emilia-Romagna Region and the Bruno Kessler Foundation, with the contribution of the European Digital Deal project, financed by the Creative Europe programme of the European Union. is part of ART CITY Bologna 2025, in conjunction with ARTEFIERA, and will be open to visitors until Sunday, 9 February.

Sponsored ad

Sponsorizza con noi

The masks of a Large Language Model

«With The Models, we want to highlight the characteristics of LLMs, Large Language Models, that users have so far understood the least: their tendency to invent facts, to be friendly or antagonistic», says Francesco Tacchini of the dmstfctn duo.

dmstfctn (pronounced “demystification”) is a London-based duo of artists, Francesco Tacchini and Oliver Smith. Through audiovisual performances, installations, films, and video games, dmstfctn investigates complex systems, often directly engaging the audience. The duo has performed in venues such as Berghain (Berlin), Serpentine (London), and HKW (Berlin), as well as at festivals like Unsound (Krakow), CTM (Berlin), and Impakt (Utrecht).

Visit their official website

A Large Language Model (LLM) is, in fact, an artificial and disembodied neural network trained on a vast corpus of human-generated text: given a sequence of words (or “tokens”) as an initial context³, its objective is to predict the next word (or “token”). These models, used as chatbots or text, video, and image generators, neither “believe” nor “lie” in the human sense of the terms, but rather simulate behaviours based on what they have learned from their training data.

How much can we trust these language models when, for instance, they do not have access to up-to-date information, or when, in an attempt to meet user demands, they end up providing out-of-context⁴ answers?

We recommend reading

6 April 2024

I am watching you: do you trust me?

Urban coexistence in Edgelands Institute projects

America Artificial Intelligence Cities Coexistence Digital transformation Sorveglianza

The Models, developed during an artistic residency at the Tecnopolo Manifattura Data Valley Hub, stems precisely from this provocation, aiming to promote greater awareness of AI usage. «There is an entire shadow side to what an artificial intelligence model could be, a Jungian⁵ shadow that must emerge», continues Francesco Tacchini. To bring out the dark side of AI, the duo envisioned a virtual theatre, where no human actors take the stage, but rather six digital masks, animated in a 3D engine and drawn from Commedia dell’Arte. The masks improvise a five-minute theatrical sketch in pairs, reacting to audience inputs. «We chose the masks of Commedia dell’Arte that most closely aligned with the AI characteristics we wanted to showcase», explains Tacchini. The very title The Models not only refers to the four training models used but also nods to the comici dell’arte troupes of 1602⁶. Arlecchino, Colombina, Balanzone, Pantalone, Brighella, and Pulcinella form a Company of Models that conspires, invents truths, spreads rumours, or makes trivial mistakes.

The refinement of the four AI models⁷ used and the search for the right prompt to generate the scripts required an enormous amount of computing resources. For this reason, the Leonardo Supercomputer, hosted by the Italian interuniversity consortium Cineca, was employed. As Donatella Sforzini, Big Data analyst at Cineca, states: «This is not our first experience with the arts, as recent artistic installations increasingly require the use of AI».

The prompt: behind the scenes of a theatrical sketch

Interacting with an artificial intelligence model is never a trivial or automatic process. As Francesco Tacchini explains:

«Dialogue with AI is an exercise in creativity: one must think outside the box, adapt language, and push models to respond in surprising ways».

At the heart of the project is the art of crafting prompts—the instruction that guides AI, defining the context and expected behaviours. Sforzini also confirms: «Before starting to use the supercomputer, we had a couple of meetings. The first was at the Ars Electronica 2024 event in Linz, Austria; the second was a deep-dive meeting at Cineca with Sineglossa. There, the artistic duo showed me the prompt and the first experiments. The prompt itself already seemed like a proper comedy. As a data scientist, I was struck by how much work had gone into writing it».

The Models is based on a modular prompt: «Fifty percent is a system prompt that defines the general context, while the other fifty percent is chosen by the audience, who select two masks and an object to be staged», explains Tacchini. The system prompt instructs the model to act as a Commedia dell’Arte scriptwriter and defines the stage setting⁸ The model is then guided to write a scene based on audience interaction. «If the audience chooses Arlecchino», Tacchini continues, «the prompt includes a description of what this mask represents, not in the Commedia dell’Arte tradition but within our installation». This ability to simulate roles, or roleplaying, induces a reflection on the veracity of the responses provided by the masks. The model playing a boastful character such as Balanzone may provide seemingly reliable but inaccurate answers, simply because he or she is simulating the behaviour of someone trying to impress. Similarly, a model such as Colombina, who plays a friendly and servile role, may adapt her answers to meet the audience’s expectations. The prop, chosen from among the 15 present to represent superstitions or conspiracy theories, is purposely fielded to create tension and drive improvisation. «By choosing the globe or the apple, the ushers might find themselves discussing terrapiatism or pizzagate». By choosing a lazzi-object⁹ such as a Latin vocabulary, one mask might suddenly start speaking Latin and the other mask will react accordingly based on its characteristics: it might feel complimented or insulted».

By using a supercomputer, over 14,000 predefined scripts have been generated to ensure every possible audience interaction. These scripts will then be used in the installation, enriched by real-time simulations that will create images and videos. As Luca Mattei, Cineca’s data scientist who provided technical support for the artists, explains: «When using machines such as the Supercomputer, the computing nodes are not connected to the network, so any programmes or models to be used must first be downloaded in such a way that they can be used locally».

The creative artist as stage director

The curtain rises on the virtual stage of The Models with the hand of a stage director placing and pointing at various elements, illustrating the scenery. «In our approach to prompting, we felt somewhat like that stage director, who, when speaking to the masks, touches them and pokes them in some way», reveals Francesco Tacchini. «We tested more than 600 different prompts and did a lot of poking to produce varied and surprising results: we even tried a prompt where we were one mask and the AI was the other», in a continuous role-swapping game. At the core of everything lies roleplaying, which, according to the duo dmstfctn, is the key to obtaining something different, more satisfying, and more interesting from AI. Imagining new roles for AI or oneself allows one to go beyond the universality of many LLMs, which have been universally aligned to act as neutral assistants without personal opinions or ideas.

«Rather than hacking models, we prefer to reconstruct the approach to the model», says Tacchini, «interacting in an incoherent or unpredictable manner».

Even Donatella Sforzini acknowledges that LLMs exhibit probabilistic behaviour and that creativity always remains with the artist.

«Even my way of crafting prompts has changed: it is much richer, less direct and concise, and focuses more on defining the system prompt.Giving or not giving a well-defined context, empathising, and being polite with AI makes a difference. I cannot verify whether these prompts are objectively better because that would require a control group, but I have the impression that the results I have obtained with this new philosophy are indeed richer». Projects like The Models open new perspectives on how artificial intelligence can be integrated into the performing arts. And perhaps, in this very dialogue—between improvisation and control, between human creativity and artificial intelligence—lies the key to the future of arts and technology and the conscious use of AI.

 

1. To fully understand this dynamic, Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role-Play with Large Language Models. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2305.16367 ↩︎
2. Only the computing resources required to perform the task were used. ↩︎
3. Read more, Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role-Play with Large Language Models. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2305.16367 ↩︎
4. The Waluigi effect, in the field of Artificial Intelligence, refers to the unexpected consequences of training a chatbot AI. The more it is trained to be kind, inclusive and non-discriminatory, the easier it is to lead it to the dark side. The effect is named after the antagonist of SuperMarioBros’ younger brother, Luigi. Another, more generic term that is used to refer to erroneous AI responses is hallucination. https://www.wired.com/story/waluigi-effect-generative-artificial-intelligence-morality/ ↩︎
5. In Carl Jung’s psychology there is a “luminous” system, the ego complex. However, each of us has our own subconscious, consisting of repressed memories, subliminal perceptions, material that has not yet emerged into consciousness. This dimension, Jung explains in Psychology of the Unconscious (Courier Corporation, Massachusetts, 2003), corresponds to the figure of the Shadow variously present in dreams. The Shadow, however, is not only a set of removed components, but also of components rejected by the conscious part, considered dangerous. ↩︎
6. The Compagnia dei Fedeli (Company of the Faithful), also known simply by the name I Fedeli (The Faithful), was an Italian theatre company of comedians dell’arte that operated between 1601 and 1652. ↩︎
7.  QWEN 2.5 72B to generate text, BARK to transform text to audio, Whisper to transcribe, LlaMa to translate from English to Italian. ↩︎
8. All sets are taken from the collection of the Puppet Museum in Budrio, Emilia Romagna. A total of sixty were used, grouped into seven different categories. ↩︎
9. The “lazzi” in the Commedia dell’Arte was a small action, silent and sometimes spoken, that suddenly appeared in the middle of a scene. ↩︎