Computational creativity

• 1. Introduction

  ›

  Computational creativity sits at the crossing of artificial intelligence, cognitive psychology, philosophy, and the arts. The central question it asks is deceptively simple: can a machine be creative? And behind that question lurks a harder one that philosophers and computer scientists have wrestled with for decades. If a machine can only do what it was programmed to do, how can anything it produces ever count as genuinely new?

  This is not an abstract puzzle. It traces back to Ada Lovelace's famous objection to machine intelligence. Modern theorists like Teresa Amabile have echoed her challenge. The field of computational creativity takes that challenge seriously, building systems that generate music, poetry, visual art, jokes, and stories, while simultaneously asking whether any of it constitutes real creativity at all.

  The answers that researchers have found are surprising. Not all computer scientists agree that machines are limited to their programs. Some argue that the boundary between computation and creativity is far more porous than it first appears. Along the way, the field has produced a robot that plays jazz, a computer that writes music in the style of Bach, and a software system that generates chess puzzles by blending features borrowed from paintings and musical scores.

• 2. Defining What Creativity Means

  ›

  AI researchers Newell, Shaw, and Simon proposed that novelty and usefulness together form the cornerstone of a multi-pronged view of creativity. Their framework identifies four criteria. A creative answer must be novel and useful, either for the individual or for society. It must demand that we reject ideas we had previously accepted. It must result from intense motivation and persistence. And it must come from clarifying a problem that was originally vague.

  Margaret Boden focused on the first two of those criteria. For her, creativity should be defined as the ability to come up with ideas or artifacts that are new, surprising, and valuable. She further distinguished between two kinds of creative output. Creativity that is novel only to the person producing it she called P-creativity, for psychological creativity. Creativity that society at large recognizes as novel she called H-creativity, for historical creativity.

  Boden also drew a line between exploratory creativity and transformational creativity. Exploratory creativity involves searching within an established conceptual space. Transformational creativity involves deliberately breaking or transcending that space. She saw the latter as far more radical and far rarer. Her frameworks have guided work in computational creativity at a very general level, providing more of an inspirational touchstone than a technical algorithm. Geraint Wiggins later attempted to formalize her insights into a more mathematical framework.

  Mihaly Csikszentmihalyi took a different angle, arguing that creativity must be considered in a social context. His DIFI framework, standing for Domain-Individual-Field Interaction, holds that an individual produces works whose novelty and value are assessed by the field, meaning other people in society, who then add successful works to the domain of shared culture. That domain then feeds back to influence the next individual who produces something new.

• 3. A History of Creative Machines

  ›

  Combinatorial methods for composing music were explored as far back as the late 1800s, involving figures like Mozart, Bach, Haydn, and Kiernberger. Simple mechanical models were built to explore mathematical problem solving as early as 1934. Then, at the 1956 Dartmouth Conference, creativity, invention, and discovery were listed as explicit goals for artificial intelligence.

  The 1970s and 1980s brought early systems that modeled creativity using symbolic and rule-based approaches. James Meehan's TALE-SPIN, built in 1977, generated narratives by simulating character goals and decision trees. Dehn's AUTHOR, from 1981, approached the problem differently, simulating an author's own process for crafting a story. Harold Cohen's AARON, continuously developed since 1973, produced visual art using a large set of symbolic rules and heuristics for visual composition. Its outputs were good enough to be displayed in reputable galleries. Meanwhile, a system called BACON was said to rediscover natural laws like Boyle's Law and Kepler's Law through hypothesis testing in constrained spaces.

  The 1990s saw more adaptive approaches. Turner's MINSTREL, from 1993, introduced Transform Recall Adapt Methods, known as TRAMs, to simulate creative reuse of prior material. Pérez y Pérez's MEXICA, from 1999, modeled the creative writing process using cycles of engagement and reflection. The JAPE system, from 1994, generated pun-based riddles using Prolog and WordNet, a lexical database developed by George Miller and his team at Princeton. David Cope's EMI, also known as Emmy, was trained in the styles of Bach, Beethoven, and Chopin, generating new pieces through pattern abstraction and recomposition. Its output was convincing enough to persuade human listeners that its music was human-generated.

  In the 2000s, machine learning began influencing creative system design. Researchers Mihalcea and Strapparava trained classifiers to distinguish humorous from non-humorous text. Chess systems like Deep Blue generated quasi-creative gameplay strategies through search algorithms and parallel processing. Dedicated workshops like the IJWCC emerged in the 1990s, growing into the annual International Conference on Computational Creativity, the ICCC, by the early 2000s.

• 4. Neural Networks and Music

  ›

  Peter Todd, in 1989, first trained a neural network to reproduce musical melodies from a training set. He then used a change algorithm to modify the network's input parameters, producing new music in what he described as a highly uncontrolled manner. In 1992, Todd extended that work using the distal teacher approach, an idea developed by researchers including Paul Munro, Paul Werbos, D. Nguyen and Bernard Widrow, Michael I. Jordan, and David Rumelhart.

  In Todd's later work, a human composer would select a set of melodies, position them on a two-dimensional plane using a mouse-based graphic interface, and train a connectionist network to produce those melodies. The network would then generate new interpolated melodies corresponding to intermediate points in that two-dimensional space. During the late 1980s and early 1990s, generative neural systems were also driven by genetic algorithms, with experiments involving recurrent networks that successfully hybridized simple musical melodies and predicted listener expectations.

  The robot Shimon, developed by Gil Weinberg of Georgia Tech, demonstrated jazz improvisation. Virtual improvisation software created through research on stylistic modeling by Gerard Assayag and Shlomo Dubnov, including OMax, SoMax, and PyOracle, generates improvisations in real time by re-injecting variable-length sequences learned on the fly from a live performer.

  Iamus represents a different landmark. Described as the first computer that composes from scratch, it produces final scores that professional interpreters can perform. The London Symphony Orchestra played a piece for full orchestra included in Iamus's debut CD. New Scientist described it as the first major work composed by a computer and performed by a full orchestra. The technology behind Iamus, called Melomics, generates pieces across different styles of music at a comparable level of quality.

• 5. Language, Jokes, and Wordplay

  ›

  Language offers continuous opportunity for creative output, from novel sentences and puns to neologisms, rhymes, sarcasm, irony, similes, metaphors, and jokes. Applied linguist Ronald Carter hypothesized two main creativity types involving words and patterns. Pattern-reforming creativity breaks rules and reshapes language through individual innovation. Pattern-forming creativity works through conformity to rules, creating convergence and symmetry between speakers through repetition.

  Humor is an especially knowledge-hungry process. The most successful joke-generation systems have focused on puns. The JAPE system generates a wide range of puns evaluated as novel and humorous by young children. An improved version called STANDUP was deployed as a tool for enhancing linguistic interaction with children who have communication disabilities. Other humor systems have tackled pronominal reference, explored by Hans Wim Tinholt and Anton Nijholt, and humorous acronyms, addressed by the HAHAcronym system of Oliviero Stock and Carlo Strapparava.

  For neologisms, Tony Veale developed a system called ZeitGeist that harvests new headwords from the web and interprets them relative to word senses in WordNet. ZeitGeist was extended to generate new words of its own, combining elements from an inventory of word parts harvested from WordNet and simultaneously determining likely glosses for those new words. Examples include the generated word gastronaut, with the gloss food traveller, and chrononaut, with the gloss time traveller. The system then uses web search to identify which generated words are both novel and genuinely unused, satisfying the H-creative standard.

  Poetry presents a harder challenge. No general-purpose poem generator adequately combines meaning, phrasing, structure, and rhyme. Pablo Gervás developed a system called ASPERA that uses case-based reasoning to generate poetic formulations of input text. ASPERA retrieves fragments from a case-base of existing poems and combines them using metrical rules into a well-formed poetic structure. In Autumn 2020, The Poetry Review, with ISSN 0032-2156, published Ariel Klein's 50% Human: A poetic interview with AI agents, an early verified instance of LLM-assisted poetry in a major literary magazine.

• 6. Visual Art and Game Design

  ›

  In August 2015, researchers from Tubingen, Germany created a convolutional neural network that can separate and recombine the content and style of arbitrary images. The algorithm, deployed on the website DeepArt, allows users to turn photographs into stylistic imitations of works by artists such as Picasso or Van Gogh in about an hour. In July 2015, Google released DeepDream, an open-source computer vision program that uses a convolutional neural network to find and enhance patterns in images through what the source describes as algorithmic pareidolia, producing a dreamlike psychedelic appearance in deliberately over-processed images.

  Penousal Machado's NEvAr system, short for Neuro-Evolutionary Art, uses a genetic algorithm to derive a mathematical function that generates a colored three-dimensional surface. A human selects the best pictures after each phase of the algorithm, and those preferences guide successive phases into pockets of the search space considered most appealing.

  Simon Colton's Painting Fool began as a system for overpainting digital images in different painting styles, color palettes, and brush types. Later extensions allowed it to generate novel images from its own limited imagination, including cityscapes and forests derived through constraint satisfaction from basic scenarios provided by the user. Artistically, the images the Painting Fool now creates appear on a par with those from AARON, though its extensible mechanisms may allow it to develop into a more sophisticated painter over time.

  In early 2016, a global team of researchers introduced the Digital Synaptic Neural Substrate, or DSNS, which generates original chess puzzles by combining features drawn from paintings and musical scores using stochastic methods. ANGELINA, a system developed by Michael Cook, creatively develops video games in Java. Its Mechanic Miner subsystem generates short segments of code that act as simple game mechanics, evaluating them by playing simple unsolvable game levels and testing whether a new mechanic makes the level solvable. Sometimes Mechanic Miner discovers bugs in the code and exploits them as new mechanics.

• 7. Hallucination as a Creative Tool

  ›

  Language models like GPT and LSTM are used to generate texts for creative purposes, including novels and scripts. These models hallucinate from time to time, presenting erroneous material as factual. Some creators have turned that tendency to their advantage, treating the unintended outputs as raw creative material.

  Ross Goodwin's 1 the Road uses an LSTM model trained on literary corpora to generate a novel inspired by Jack Kerouac's On the Road. The model received multimodal input captured by a camera, a microphone, a laptop's internal clock, and a GPS throughout an actual road trip. Brian Merchant described the resulting novel as pixelated poetry in its ragtag assemblage of modern American imagery. Oscar Sharp and Ross Goodwin created the experimental sci-fi short film Sunspring in 2016, written with an LSTM model trained on their scripts and on science fiction movies from the 1980s-1990s. Rodica Gotca critiqued the film's overall lack of focus on narrative and intention.

  Researchers highlight the positive side of hallucination for generating novel solutions, provided that correctness and consistency can be controlled. Jiang et al. propose a divergence-convergence flow model for harnessing hallucinatory effects. In the divergence stage, potentially hallucinatory content is generated freely. In the convergence stage, intent recognition and evaluation metrics filter the hallucinations for those that are useful. They categorize hallucinations into factuality hallucinations and faithfulness hallucinations, each divisible into smaller classes including factual fabrication and instruction inconsistency.

  Luciano Floridi, in 2025, offered a literary-theoretical framework for AI-assisted writing that he called Distant Writing. In his model, the human author functions as a designer and curator who develops narrative structures through iterative selection and what he calls Socratic maieutics, meaning prompting, while taking full intellectual responsibility for the resulting work. Floridi coined the term in explicit analogy to Franco Moretti's concept of distant reading, which had reframed literary analysis as the large-scale, algorithm-assisted study of textual corpora. A pre-LLM antecedent for Floridi's idea existed in visual art: Nicolas Bourriaud's Postproduction, published in 2002, argued that artists increasingly function as programmers and navigators of pre-existing cultural material rather than as original creators.

• 8. Criticism and Open Questions

  ›

  Traditional computers fundamentally transform a discrete, limited set of input parameters into a discrete, limited set of outputs using a limited set of computational functions. Critics argue that everything in the output must already have been present in the input data or the algorithms. The mathematician Chaitin made this same argument mathematically. Similar observations come from a Model Theory perspective. All of this criticism converges on the claim that computational creativity may look like creativity but is not real creativity, because nothing new is created, only transformed by well-defined algorithms.

  Mark Riedl argues that human creativity and computational creativity differ in several dimensions at their current state. One challenge he identifies is the educational, moralizing aspect of stories. The lack of intention in AI models hinders them from making morally responsible choices, which appear frequently in human creative work.

  Michele Loi and Eleonora Vigano identified a threat running in the opposite direction. Society's overreliance on algorithms for decision-making could constrain utility functions and discourage people from exploring riskier solutions. Drawing on John Stuart Mill's notion of openness to experiments of life as a factor in creativity, they warned that this dynamic could decrease the diversity of human exploration and thus human creativity itself.

  Philip Hutchinson introduced the concept of self-innovating artificial intelligence, or SAI, to describe how companies use AI in innovation processes to enhance their offerings incrementally by continuously combining and analyzing multiple data sources. As AI becomes a general-purpose technology, the spectrum of products developed with SAI will broaden from simple to increasingly complex. Jordanous and Keller, in a more technical vein, extracted 694 creativity-related words from empirical studies spanning 60 years of psychology and creativity research, clustered them by lexical similarity, and identified 14 key components of creativity as the basis for a framework called SPECS, or Standardised Procedure for Evaluating Creative Systems.