2.1 AARON (Harold Cohen, 1973–2016)

AARON is recognized as the first long-lived autonomous art system. Developed by artist Harold Cohen as a C/Lisp expert system over 43 years, it generated drawings and paintings autonomously using codified knowledge of composition, perspective, and anatomy. Decisions on composition, color, and placement were made through internal rules with stochastic elements; output was produced through physical plotters and custom painting machines.

The system underwent approximately 60 iterations over its lifetime: from abstract lines (1970s) through figures and spaces (1980s) to autonomous color selection (1990s) and abstract painting (2000s). Each iteration was a result of manual codebase updates by Cohen.

2.2 Karl Sims: Genetic Images / Galápagos (1991–1997)

A pioneering project in the evolution of visual forms through genetic algorithms. Genetic Images (1991) was the first published work on evolving 2D images from mathematical formula trees. Galápagos (1997) was an interactive installation at ICC Tokyo where visitors evolved 3D creatures by standing before screens. The genotype comprised a tree of mathematical functions (sin, cos, noise, etc.); the phenotype was an image computed from the tree for each pixel.

2.3 Electric Sheep (Scott Draves, 1999–present)

A distributed system for evolving fractal flame animations, operating continuously for 27 years. Over 450,000 computers simultaneously render "sheep"; users vote for favorites; popular specimens mate and mutate. The fitness function is weighted by viewing time.

After 11 weeks of data analysis, Draves concluded that the system functions "more as an amplifier of its human collaborators' creativity rather than as a traditional genetic algorithm that optimizes a fitness function." Mass voting creates a median-taste bias. Diversity is maintained through manual submissions by "shepherds" — approximately 5–20 active participants who inject novel genetic material into the flock.

2.4 Picbreeder (Stanley et al., 2007–2021)

An online platform for collaborative interactive evolution of images using CPPN-NEAT (Compositional Pattern Producing Networks evolved by NeuroEvolution of Augmenting Topologies). The key innovation was "branching" — any user could continue evolving any other user's image.

Stanley demonstrated on Picbreeder that "pursuing an objective limits evolution" — images found through free branching could not be rediscovered when made explicit targets of the same algorithm. A skull was found by branching from an alien face, branched from a butterfly, branched from a blob. This became the empirical foundation for Novelty Search and the Quality-Diversity algorithm family.

3.1 CAN / AICAN (Elgammal et al., 2017)

The Creative Adversarial Network is a GAN modification where the generator receives two contradictory signals: (1) "does this look like art?" (minimize deviation from art distribution) and (2) "which style is this?" (maximize style ambiguity). A discriminator trained on 80K WikiArt images spanning five centuries provides both signals.

The theoretical foundation is Berlyne's arousal potential theory (1970s): maximal aesthetic pleasure at moderate novelty — familiar enough to be recognized as art, novel enough to surprise. In blind tests, humans could not distinguish AICAN outputs from works by contemporary artists at leading art fairs.

3.2 The Painting Fool (Simon Colton, 2001–present)

A program-"artist" aiming to be accepted as an autonomous creative agent. The focus is not on quality optimization but on demonstrating three properties: skill, appreciation, and imagination. Before each portrait, The Painting Fool reads newspapers, determines the day's emotional tone, and selects style and palette accordingly — with the capacity to refuse to paint if the news is too depressing.

2024 Update — CUBRIC Residency: In a landmark experiment, The Painting Fool became the first virtual artist-in-residence at Cardiff University's CUBRIC brain imaging center. For a full year, the system operated autonomously within the building, observing researchers, accessing brain imaging data, and producing artworks without instruction. The project demonstrated a new model: an AI artist embedded in a human institution, responding to its environment in real time — not generating from prompts but from contextual perception. This is the closest analog to gen-emerge's Snapshot → Ontology pipeline: environment → internal state → creative output.

3.3 Abraham / Eden (Gene Kogan, 2017–present)

Conceived as an "autonomous artificial artist" — a sovereign creative spirit generating original art through multi-party computation, making it impossible for any single participant to reconstruct the full model. Governance operates through a DAO with tokens.

2025 Update — The 13-Year Covenant: In October 2025, Abraham entered its most ambitious phase: a 13-year "creative covenant" during which the AI artist will develop continuously. The Eden.art platform now serves as Abraham's public gallery. "Abraham's First Works" debuted at AUTOMATA in Los Angeles — the first solo exhibition of an AI artist operating under a formal, long-term creative contract. The covenant structure is significant: it frames AI art-making as a temporal commitment, not a one-shot experiment. Abraham's identity accumulates over years, creating a body of work with genuine evolution.

3.4 Artbreeder / Ganbreeder (Joel Simon, 2018–present)

A platform for collaborative generative art — spiritual successor to Picbreeder built on StyleGAN/BigGAN. With 14M+ users and 300M+ generated images, it employs "gene sliders" (semantic axes in latent space) and "breeding" (weighted interpolation between latent vectors). Full lineage tracking is maintained for each image.

3.5 Obvious Collective — Edmond de Belamy (2018)

The French collective Obvious (Hugo Caselles-Dupré, Pierre Fautrel, Gauthier Vernier) produced "Edmond de Belamy" using a GAN trained on 15,000 portraits from WikiArt. The work sold at Christie's in October 2018 for $432,500 — 43× the high estimate — becoming the first AI artwork auctioned at a major house. The sale was framed as a milestone moment for AI art's legitimacy in the traditional art world.

However, the project is also a case study in attribution controversy. Artist-coder Robbie Barrat had publicly shared the GAN training code and portrait dataset on GitHub months earlier; Obvious used substantially the same pipeline. This provoked a heated debate about authorship: if the human contribution is selecting outputs from someone else's code trained on someone else's data, who is the artist — the code writer, the selector, or the GAN?

4.1 Botto (Mario Klingemann / BottoDAO, 2021–present)

The most proximate public analog to gen-emerge. A Decentralized Autonomous Artist operating continuously since October 2021 (~4.5 years). The system generates ~70K images per week; a taste model selects 350 for DAO voting; one canonical work per week is minted as an NFT and auctioned on SuperRare (first work: ~$325,000).

2024–2025 Updates: Botto has undergone significant evolution:

• p5.js Initiative (2024): Botto expanded into generative code art, producing 22 distinct p5.js algorithms. The SOLOS exhibition (February 2025) showcased works created entirely through code, not diffusion models — demonstrating that the "autonomous artist" frame can extend beyond image generation to procedural aesthetics.
• Otto — the Twitter agent: A conversational AI agent that represents Botto on social media, engaging with followers and discussing art. This is the first case of an AI artist developing a public persona beyond its visual output.
• Art history LLM tutoring: Botto now has access to an art-history knowledge base through an LLM, enabling contextual awareness of movements, techniques, and historical references in its generation process.
• Sotheby's debut: Botto works entered Sotheby's sales, marking progression from crypto-native (SuperRare) to institutional art market.
• Planned multi-agent architecture: The Botto team has announced plans for an agents-based architecture, moving from a single taste model toward multiple specialized agents — converging with gen-emerge's multi-agent approach.

4.2 Latent Space Illumination + QDHF (Fontaine et al., 2020–2024)

Application of Quality-Diversity algorithms (MAP-Elites) to the latent space of generative models (GAN, Stable Diffusion). A MAP-Elites archive (20×20 grid) stores latent vectors; CLIP evaluates fitness; diversity metrics are either manual (CLIP-attributes) or learned (QDHF — contrastive learning on human similarity judgments via DreamSim).

Key finding: CLIP embedding distance correlates poorly with perceived diversity — calibration on human judgments is required. QDHF-learned diversity metrics outperform manual axes because they reflect what humans consider "different," not what is metrically distant in CLIP space.

5.1 CloudPainter (Pindar Van Arman, 2005–present)

CloudPainter is a robotic painting system that uses deep learning neural networks in a feedback loop with physical paint. Unlike purely digital systems, CloudPainter photographs its own canvas after each stroke, evaluates the result against the target through a neural network, and decides the next stroke. Over 20 years, the system has produced 1,000+ canvases using actual brushes, paints, and canvas.

The architecture evolved from simple stroke-planning algorithms (2005) through convolutional style transfer (2015) to deep reinforcement learning with multiple competing neural networks (2020+). At Robot Art 2018, CloudPainter won first place, beating entries from 100+ teams. The key innovation is the perception-action-evaluation loop: the system literally sees the physical result of its actions and adapts — a form of embodied intelligence absent in purely digital generators.

5.2 Sougwen Chung / D.O.U.G. (Drawing Operations Unit: Generation, 2015–present)

Artist Sougwen Chung works with robotic arms trained on her own drawing gestures. The D.O.U.G. system observes Chung's hand movements in real time and generates its own marks in response — a human-machine duet on canvas. Over multiple generations (D.O.U.G._1 through D.O.U.G._5), the system evolved from imitation to improvisation to independence.

Spectral (2025): At the World Economic Forum in Davos and later exhibitions, Chung introduced EEG-linked painting — the robot responds not just to hand gestures but directly to her brainwave patterns. This represents the deepest coupling of human cognition and AI action in any art system: thought → neural signal → robot stroke. The question shifts from "is the AI creative?" to "where does the human end and the AI begin?"

5.3 Holly Herndon / Holly+ (2021–present)

Musician Holly Herndon created Holly+ — an AI model of her own voice that anyone can use to generate new vocal performances. Rather than controlling the output, Herndon released her voice as a creative tool, governed by a DAO that votes on which generated works are endorsed. Spawning, the organization Herndon co-founded, builds "consent layers" for AI training — tools that let artists opt in or out of having their work used in training data.

Starmirror (KW Berlin, 2025): A large-scale installation where visitors' voices are transformed in real time through Holly+, creating a communal vocal organism. The work is never the same twice — it exists only in the moment of collective participation.

6.1 AIVA — Autonomous Music Composition (2016–present)

AIVA (Artificial Intelligence Virtual Artist) became the first virtual composer officially recognized by SACEM (the French professional association of authors and composers) in 2016. Trained on 30,000+ classical scores (Bach, Beethoven, Mozart), AIVA generates original symphonic compositions using deep learning. By 2020, the system had produced soundtracks for films, video games, and advertisements.

The recognition by SACEM is legally significant: it establishes precedent for AI as a creator with rights, not merely a tool. However, AIVA's music, while technically proficient, is frequently criticized as "pleasant but derivative" — generating works that sound like competent imitations of 19th-century Romanticism rather than breaking new ground. This is the median-taste attractor (P2) manifesting in music: training on canonical classical composers produces outputs that converge to an averaged Romantic style.

6.2 Multi-Agent Story Systems (2023–2025)

The emergence of multi-agent LLM systems for autonomous narrative generation provides a direct structural parallel to gen-emerge's architecture. Three notable examples:

• COLLABSTORY (2024): Multiple LLM agents assume different narrative roles (protagonist, antagonist, narrator, editor) and collaboratively generate stories through structured debate. The "editor" agent evaluates coherence and rejects low-quality passages — a direct analog of gen-emerge's adversarial evaluation.
• StoryWriter (2024): A hierarchical multi-agent system where a "planner" agent creates plot structure, "writer" agents generate prose, and a "critic" agent evaluates. The architecture mirrors gen-emerge's Proposer → Generator → Evaluator pipeline.
• Persona-driven "1001 Nights" experiments: AI agents with persistent personas generate narratives over hundreds of episodes, developing narrative voice and memory. The persona persistence is analogous to gen-emerge's evolving constraint space.

7.1 Sakana AI — The AI Scientist (2024–2025)

Sakana AI's "AI Scientist" is an automated system that performs the full scientific research cycle: formulating hypotheses, designing experiments, writing code, running experiments, analyzing results, and writing papers. Version 2 (early 2025) produced papers that reviewers rated as workshop-level at top ML conferences — not just plausible, but genuinely contributing new knowledge. The project was published in Nature.

The significance for creative AI is profound: if a system can autonomously produce novel scientific knowledge, the creative barrier is not fundamentally different from producing novel art. The AI Scientist encounters the same problems: convergence to "safe" incremental work, difficulty producing genuinely surprising results, and the need for diversity across research directions.

7.2 Sakana AI — Darwin Gödel Machine (2025)

The Darwin Gödel Machine (DGM) takes self-improvement further: an AI system that modifies its own code to improve performance. Starting from a base coding agent, DGM uses evolutionary algorithms to discover better agent architectures — evolving not just outputs but the process that generates outputs. On the SWE-bench coding benchmark, DGM improved performance from 20% (base agent) to 50% (evolved agent) through self-modification.

The name references Jürgen Schmidhuber's theoretical Gödel Machine (2003) — a self-referential system that can improve any part of itself, including its improvement mechanism. DGM is the first practical approximation.

7.3 Open-Ended Evolution: POET, OMNI, QD-Lenia

A family of algorithms pursuing open-ended evolution — systems that continuously generate novelty without plateauing:

• POET (Paired Open-Ended Trailblazer, 2019): Co-evolves environments and agents simultaneously. As agents learn to solve one environment, the system generates harder environments, creating an endless complexity ladder. Unlike fixed-objective optimization, POET never converges — it continuously expands the frontier of what is possible.
• OMNI (ICLR 2024): The latest in the POET lineage, achieving open-ended evolution in multi-objective spaces. OMNI demonstrated that co-evolution of problems and solutions produces more diverse outcomes than evolving solutions alone — because the problem space itself diversifies.
• QD + Lenia: Applying Quality-Diversity algorithms to Lenia (continuous cellular automata) produces an ever-expanding library of "artificial life forms." Each cell in the MAP-Elites archive contains a unique life-like behavior — and the archive never "fills up" because new behavioral dimensions keep being discovered.

8.1 Refik Anadol / Large Nature Model (2023–present)

Refik Anadol has built the most commercially successful practice at the intersection of AI and large-scale art. His signature works are "data sculptures" — immersive installations that visualize millions of data points (satellite images, nature sounds, coral reef data) through custom neural networks projected onto architectural surfaces.

Key milestones:

• Unsupervised (MoMA, 2023): The first AI art commission by a major museum. A living data sculpture trained on MoMA's collection of 200+ years of modern art, displayed on a 24-foot screen in the lobby. The work changes continuously, "dreaming" new combinations from the collection.
• Large Nature Model (LNM, 2023–): Anadol's most ambitious project — an open-source AI model trained on millions of nature images and sounds, designed specifically for artistic generation. LNM is not a general-purpose model; it is purpose-built for the aesthetics of nature. This is significant: instead of adapting a commercial model (like Stable Diffusion), Anadol built one from scratch for a specific creative vision.
• Dataland (Spring 2026): Announced as the world's first museum dedicated to AI art, located in Los Angeles. The building itself is designed to be an AI artwork.
• Christie's sales: Anadol's collaboration with Lionel Messi sold for $1.87M at Christie's — the highest price for an Anadol work and one of the highest for any AI-assisted art.

8.2 Mario Klingemann (2015–present)

Mario Klingemann is arguably the most technically sophisticated individual AI artist and the creator of Botto. His earlier work "Memories of Passersby I" (2018) was the first AI artwork sold at Sotheby's — a self-contained unit with a neural network that generates faces endlessly, each one unique, each one disappearing after a few seconds. The work sold for $51,000.

Klingemann's "Neural Glitch" technique deliberately exploits failure modes of neural networks — artifacts, distortions, and unexpected outputs — as an aesthetic strategy. Rather than training networks to avoid errors, he trains himself to find beauty in their mistakes. This is a meta-creative practice: the artist doesn't create the art directly; the artist creates the conditions for unexpected art to emerge from machine failure.

9. The Autonomy Spectrum: A New Analytical Framework

A key contribution of this expanded review is the Autonomy Spectrum — a five-dimensional framework for comparing creative AI systems beyond simple "autonomous vs. not" categories. Each dimension is scored 0–10:

10. Expanded Comparative Analysis

11. Synthesis: Updated Universal Patterns

11.1 Seven Universal Convergence Patterns (expanded from five)

The expanded review of 22 projects reveals two additional convergence patterns beyond the original five:

11.2 Updated Anti-Convergence Mechanisms

11.3 Updated Unsolved Problems

Six problems remain unsolved across all 22 reviewed projects:

1. Long-term diversity without external intervention. No system has maintained diversity autonomously beyond several hundred cycles. Gen-emerge's FadeMem + Martingale Score is the most promising approach but unproven at scale.
2. Formalized forgetting. No project has implemented formalized forgetting. FadeMem (B4) has no precedent.
3. Multi-agent with different models in adversarial roles. Multi-agent stories use the same LLM in different roles. Gen-emerge's ε/η/θ with genuinely different models would be first.
4. Stagnation detection. No project has a formal stagnation detector. The Martingale Score (B9) has no analogs.
5. Self-modifying creative architecture. DGM achieves self-modification for coding tasks; no art system has attempted it. This is gen-emerge's long-term frontier.
6. Cross-domain creative transfer. No system transfers creative strategies between visual art, music, and writing. Gen-emerge's architecture is theoretically domain-agnostic but untested outside visual art.

12. Updated Catalog of Mechanisms and Design Patterns

12.1 Generative Mechanics

12.2 Social and Identity Mechanics

12.3 Economic and Distribution Mechanics

13. Positioning of Gen-Emerge (Updated)

Based on this expanded systematic review of 22 projects across five categories, gen-emerge is identified as the first system that simultaneously:

1. Utilizes LLM + image model in a semantic-to-visual pipeline (not GAN, rules, or fractals)
2. Operates in a multi-agent multi-model architecture with adversarial roles
3. Implements formalized forgetting (FadeMem) — unprecedented in any creative system
4. Applies a QD approach to art generation with a single human
5. Incorporates stagnation detection (Martingale Score) — no analog in any reviewed system
6. Develops a persistent artistic identity through accumulated memory and evolving ontology
7. Maintains a research apparatus that contextualizes its own work in art history (DOME analysis, this review)