Can Artificial Intelligence Create Gardenscapes Levels?

Green AI robot designing a Match-3 puzzle level in a futuristic garden-themed laboratory with colorful blocks and digital planning screens.

Explore how AI could generate, test, and refine Gardenscapes levels, while human designers still shape pacing, purpose, and player experience.

Gardenscapes constantly needs new levels. Another 50 are added every week, each with different boards, objectives, blockers, move counts, and combinations of elements. As the game continues to grow, one question becomes increasingly reasonable: does a human really need to design every new level, or could artificial intelligence now handle part of that job?

We already know quite a bit about how a Gardenscapes level is designed. The question here is different: can an AI system bring all of those pieces together and turn them into a real, playable Match-3 level?

Based on today's technology, the answer is yes. But that does not mean we have evidence that Playrix is using AI to create current Gardenscapes levels. Those are two very different claims.

What Does It Actually Mean for AI to Create a Level?

Simply filling a grid with colorful pieces and a few obstacles is not enough. A real level has to work.

There must be enough playable space to make meaningful moves, the objectives have to be achievable, the elements need to interact correctly, and the number of available moves has to match the level of difficulty the designer is aiming for.

On top of that, the board cannot just be technically solvable. It has to create a specific kind of experience. The player may need to open up part of the board first, remove a blocker, activate a generator, or create Power-ups before gaining access to the main objective.

That makes the challenge far more complicated than simply generating a random board.

AI Can Already Generate Match-3 Layouts

Research into procedural and automated game creation has reached the point where machine-learning systems can be trained on existing levels and then used to generate new ones.

One notable example comes from EA SEED research into Match-3 level generation. Its system can create new levels using characteristics learned from levels originally designed by humans.

The designer does not necessarily have to request one exact layout. Instead, certain characteristics can be defined, and the system can generate many different possibilities.

For a Gardenscapes-style system, those inputs could theoretically include:

the size and shape of the board,
which elements are allowed to appear,
which objectives must be completed,
how open or restricted the starting board should be,
the number of available moves,
and the desired difficulty level.

From those specifications, the AI would not have to produce just one level. It could generate hundreds of possible versions and then move on to the part that really matters: deciding which ones are worth keeping.

AI Can Reject Bad Levels on Its Own

Imagine that the system generates 500 different boards.

Some may open up far too quickly. In others, the objective may practically complete itself. Some could restrict the opening moves so heavily that luck has too much influence over the outcome.

A human tester does not need to sit down and play all 500.

AI agents can act as automated players. Research published in IEEE Access showed that reinforcement-learning agents can playtest Match-3 missions and help evaluate their difficulty.

In the most complex mission examined in that study, the agent's performance came within roughly 5% of human performance.

In practical terms, one system can generate a level and another can immediately start playing it over and over again.

Thousands of Attempts Before a Real Player Ever Sees the Level

This may be where AI has its biggest advantage.

A human tester has limited time. An automated system can replay the same level an enormous number of times and collect data from every attempt.

It can measure how often the level is completed, where the agent tends to fail, how many moves remain after successful attempts, and which starting conditions produce dramatically different outcomes.

For Gardenscapes, this is especially interesting because older public presentations from Playrix showed how Gardenscapes difficulty was measured using failed attempts and then compared with the target originally set for a level.

With AI, part of that measurement could theoretically happen before the level ever reaches real players.

Could AI Deliberately Create an Extremely Difficult Level?

From a technical standpoint, that is even more interesting than simply generating a board.

A system can be given a specific difficulty target and search for layouts that come close to it.

It does not need to know in advance exactly which arrangement will work. It can generate variations, test them, and keep the versions whose results come closest to the target.

In 2025, a reinforcement-learning system for Match-3D puzzles was presented at the IEEE Conference on Games. It generates levels according to specific difficulty targets, and the researchers reported that the model was integrated into a real commercial mobile puzzle game.

Match-3D is not Gardenscapes, and its mechanics are different. What matters here is something else: the idea of giving AI a target difficulty and allowing it to generate a level around that target has already moved beyond pure research and into an actual commercial game.

Huge Blockers Are an Ideal Problem for AI

Consider a board where large objects take up a substantial portion of the playable area.

The position of a single blocker can completely change the opening moves. One extra free tile may make it possible to create a Power-up. Move that same object by one tile, and the opportunity may disappear.

A human designer can recognize those differences and make adjustments. AI, however, can systematically test an enormous number of variations.

It can move a large object, rerun the simulations, adjust the number of moves, test again, and compare the results.

For an optimization system, those tiny adjustments are not tedious. They are simply more data.

But Some Things Are Much Harder to Measure

An AI system can determine that a level has a certain win rate. That does not mean it knows whether the level is actually good.

Two levels may have almost identical statistical difficulty while delivering completely different experiences.

In one, the player may gradually understand how the board needs to be opened and feel that every move is bringing the objective closer. In the other, the player may simply be waiting for the right pieces to fall into place.

That distinction matters because older explanations from Playrix showed that element placement followed deliberate design logic. A playable area was supposed to serve a purpose, and an objective was expected to add something meaningful to the level.

That is much harder to capture with a single difficulty number.

The Most Realistic Use of AI Does Not Have to Replace the Designer

AI may be more useful as a tool for generation, testing, and selection than as the final creator.

A designer could define the central idea of the level and choose the elements that should be used. The system could then generate multiple variations. Automated agents could eliminate the versions that create problems, leaving the designer to choose the most interesting ones.

Instead of starting with an empty board, the designer could begin with ten versions that already work.

The biggest benefit would not necessarily be that “AI created the level.” It would be that thousands of small tests, variations, and iterations had already been completed before a human designer ever had to work on it directly.

So What Do We Know About Gardenscapes Itself?

There is currently no public confirmation from Playrix that it uses generative AI or reinforcement learning to create regular Gardenscapes levels.

That means we cannot look at an unusual level and claim that AI created it. Neither the complexity of a board nor extremely high difficulty is evidence of AI involvement.

The important distinction is between “this can be done” and “we know this is being done.”

On the first point, the answer is now clear: current technology can generate puzzle layouts, playtest them automatically, and even steer their creation toward a specific difficulty target.

On the second point, when it comes specifically to Gardenscapes, we still do not have public evidence.

Conclusion

The interesting question is no longer whether a computer can place pieces and blockers on a Match-3 board. It already can.

The real challenge is whether AI can consistently create levels that are not only technically sound and appropriately difficult, but also have a clear idea, good pacing, and interesting decisions for the player.

That is where the human level designer still has a different role: not simply deciding whether a level can be beaten, but whether it is worth playing.

For now, the robots can keep running a few thousand simulations behind the scenes. The next impossible-looking level, though, probably will not tell us who designed it.