Self-evolving games: procedural generation and what is real

A "self-evolving" game usually means one of four things: levels built by an algorithm from a seed, text or art made by an AI model, difficulty that adjusts to the player, or a team that changes the game from what players do. The first three are decades old and work. Worlds generated frame by frame, like Google DeepMind's Genie 3, are research previews with minutes of memory and no learning study. Peabody (peabody.kids), free through the end of 2026, uses seeds, rotation and play-time data, and an AI only inside a sandbox.

Procedural generation: the old, solid part

Games have built levels from algorithms since the 1980s. The research names two strands. Procedural content generation via machine learning, surveyed by Adam Summerville, Julian Togelius and colleagues in IEEE Transactions on Games, trains models on existing levels and generates new ones; the survey lists small datasets and playability as the open problems. Experience-driven generation, set out by Georgios Yannakakis and Togelius in 2011, makes a model of the player's experience the objective and searches for content that should produce it.

Learning-specific evidence is thin. Danial Hooshyar and colleagues (2018), in the Journal of Computer Assisted Learning, used a genetic algorithm to generate early English reading game content tailored to each child's ability, and children playing the tailored content gained more than those playing untailored content. Language models can now write playable levels too: MarioGPT and similar systems produce playable Mario and Sokoban levels, evaluated for playability, not learning.

Generated worlds: the research-preview part

Google DeepMind announced Genie 3 on 5 August 2025: text-prompted worlds at 720p and 24 frames a second that a person can move through in real time. Its own announcement lists the limits: worlds stay consistent for a few minutes, the actions a player can take are restricted, several agents interacting is not reliable, text renders poorly, and it is a limited research preview. Roblox open-sourced Cube in March 2025 to generate 3D objects from text, and NVIDIA's ACE runs small language models that plan characters' behaviour in shipped games.

None of these has a study of what a child learns in them. For a web product, build-time generated assets with a person choosing what ships are real; a world dreamed up live for a five-year-old is not.

Adaptive difficulty: real, with one trap

Robin Hunicke's 2005 case for dynamic difficulty adjustment set the frame: too easy is boring, too hard is frustrating, and the adjustment must be invisible. Her cautionary example was racing-game rubber-banding, where trailing cars suddenly speed up, which players notice and resent. A 2024 systematic review found most studies report positive effects of adjustment on enjoyment and engagement, and a 2024 paper in Entertainment Computing argued the field leans too hard on flow theory. The supported version adjusts what is asked, item by item, not how the race is rigged.

Adaptive stories: twenty years of Crystal Island

The best-funded adaptive narrative game for learning is Crystal Island, an eighth-grade microbiology mystery from North Carolina State. Early versions taught less than a slideshow while motivating far more. Later work by Jonathan Rowe and colleagues (2011, 153 students) found that engagement predicted learning independent of prior knowledge, and that learning reached parity only with explicit guidance. Reinforcement learning over logged choices worked for pacing and hints but needed many sessions per policy. The lesson for a small product: a story can adapt among a handful of hand-written branches; it cannot invent a curriculum.

A template for generated levels

Nintendo's designers describe levels in four beats borrowed from four-panel comics, kishōtenketsu: introduce a mechanic safely, develop it, twist it, then conclude. Koichi Hayashida, director of Super Mario 3D Land, said a mechanic can be taught, developed, twisted and thrown away in about five minutes. It is an interpretive framework from interviews, not a Nintendo document, but it is a ready-made rule for any generator and for where a learning question should sit: first in a safe place, last as the proof.

How Peabody's games change

Peabody's games do evolve, in the old and checkable ways, and the learning never comes from a generator.

• Seeds, not dreams. Invasion rebuilds its island from a fresh seed every match: a new coastline, one to three mountain ranges, two to five lakes, a cold side and a hot side. Build World grows each island from seeded noise, with the generator's version frozen so a saved world never shifts.
• Rotation on a clock. The worlds behind the kid side and the home's layouts turn on a ten-minute clock, so the place looks new without anything being invented live.
• What kids play decides what stays. The games hub ranks games by seconds of real play over the last 30 days, per day the game has existed. A game kids drop sinks, and a game that never gets good can be moved to Trash games by hand, still playable. Play time is counted anonymously, with no child attached.
• The questions never come from a generator. Every learning question in every game comes from the practice brain, ordinary code with a mastery ladder per subject, three answer boxes and a rule that a question answered right is never asked again. A test called the learning invariants checks this before any change to a game can land.
• AI inside a sandbox. In Peabody Studio a child describes a game in their own words and an AI model writes it against a small game kit. The child's words are screened first, the code is checked before it runs, and the game runs in a sealed frame with no network and no storage. When a child's game asks a question, the practice brain supplies it.
• Agents that propose, a person who merges. A child's feedback from a game goes to an inbox, and an issue Aaron, who runs Peabody, has planned can be handed to a coding agent that builds it into a pull request (the Workshop, still in its first, bugs-and-tweaks-only stage). A change that touches the safety code, the AI policy or the learning invariants is handed back instead of opened, and only Aaron merges.

There is no difficulty rubber-banding. When a child misses, the practice brain stops adding new items for that round, and after more misses it tells the game to ease. Nothing in the race is rigged.

What this does not show

• No generated-content learning trial is large. The one tailored-content study cited above is small; MarioGPT-style work measures playability.
• Engagement is not learning. Crystal Island shows the two can come apart.
• Peabody has not been tested in a trial. Seeds, rotation and play-time ranking keep games fresh; nothing here claims an effect on scores.

More on the platform side is in what a self-evolving learning platform means, and the rules are on the science page. Start free or browse the games.

Questions parents ask

Are AI-generated games safe for kids?

It depends on the walls around them. A game whose code an AI wrote should run where it cannot reach the network or the child's data, and its words should be screened. Peabody Studio's games run that way, and their questions come from code.

Do procedurally generated games help children learn?

There is a small study where reading content tailored by an algorithm beat untailored content. Most generated-level research measures whether levels are playable, not what a child learns.

What is Genie 3?

A Google DeepMind research model, announced in August 2025, that generates a world a person can move through in real time from a text prompt. It holds a world together for a few minutes and is a limited research preview.

Does Peabody change its games by itself?

Not by itself. Islands come from seeds, backdrops rotate on a clock, games kids stop playing sink in the ranking, and a coding agent can propose fixes, but a person merges every change and a test guards the learning rules. No AI runs inside any game except the Studio, where a child asks for one.