Creativity and Machine Learning: A Survey

TL;DR

This survey addresses the problem of distinguishing creativity in machine-generated artifacts from mere data replication by proposing a seven-way taxonomy of generative models (VAE, GAN, sequence models, transformers, diffusion, RL-based, and input-based methods) and linking each to Boden's creativity criteria. It introduces a structured evaluation framework, including LT 2.0, Ritchie’s criteria, FACE, SPECS, Creativity Implication Networks, generate-and-test, and Unexpectedness, to quantify novelty, value, and surprise across domains. The authors discuss practical implications, including the role of prompts, rewards, and human feedback, and highlight key challenges such as defining universal creativity metrics, copyright concerns, and the need for robust, domain-agnostic evaluation protocols. The work aims to guide researchers toward creativity-oriented objectives and better interpretability of machine-generated artefacts, with potential impact across art, design, and related creative professions. It also emphasizes that higher creativity often requires carefully crafted objective functions and evaluation strategies beyond standard likelihood or discriminator-based signals, fostering both exploratory and transformational outputs when appropriate cues and constraints are applied.

Abstract

There is a growing interest in the area of machine learning and creativity. This survey presents an overview of the history and the state of the art of computational creativity theories, key machine learning techniques (including generative deep learning), and corresponding automatic evaluation methods. After presenting a critical discussion of the key contributions in this area, we outline the current research challenges and emerging opportunities in this field.

Figures (1)

• Figure 1: A schematic view of the seven classes of generative learning methods presented in this survey. Top, left to right: Variational Autoencoder (\ref{['vae']}), with a decoder generating $\mathbf{x'}$ given a latent vector $\mathbf{z}$, and an encoder representing $\mathbf{x}$ into a latent distribution; Generative Adversarial Network (\ref{['gan']}), with a generator to produce $\mathbf{x'}$, and a discriminator to distinguish between real $\mathbf{x}$ and synthetic $\mathbf{x'}$; Sequence prediction model (\ref{['sequenceprediction']}), with a generator to output $\mathbf{x}$ one token after the other given in input previous tokens; Transformer-based model (\ref{['transformers']}), with a Transformer outputting $\mathbf{x}$ one token after the other given in input previous tokens, or a masked version of $\mathbf{x}$. Bottom, left to right: Diffusion model (\ref{['diffusion']}), with a model to learn an error $\mathbf{\epsilon}$, which is used to incrementally reconstruct $\mathbf{x_0}$; Reinforcement Learning (RL)-based method (\ref{['rlbased']}), with a generative model acting (i.e., progressively generating $\mathbf{x}$) to maximize a given reward function; Input-based methods (\ref{['inputbased']}), with an input optimized by a given loss. The input can be a vector $\mathbf{z}$ given to a generative model to obtain the desired output, or directly a product $\mathbf{x}$ becoming the desired output.

Theorems & Definitions (2)

• definition 1
• definition 2