When AI starts using your PCG tools
Recently, much of my work has been related to PCG systems and content generation tools.
An interesting aspect of this work is that these tools were not used only by humans.
As the system evolved, it became increasingly clear that tools convenient for humans are not always suitable for AI. Conversely, solutions that work perfectly in an automated pipeline often turn out to be inconvenient for manual use.
Over time, the difference between these two approaches became so noticeable that I began to think of them as two separate classes of tools.
In this article, I would like to share several observations about how PCG tools designed for humans differ from tools intended for AI use.
Determinism and Result Stability
The first and perhaps most noticeable difference is the required level of precision in how a tool behaves.
Those who have observed the Spline Sampler in Unreal Engine 5 know that sometimes you need to slightly adjust the spline in the level to get the expected result.
For a human, this is not a problem. In fact, it is a normal part of the workflow. A person quickly understands what is happening and corrects the input data directly in the level.
For AI, however, this approach is practically unacceptable.
If a tool requires manual adjustment of input data, the generation time may increase by an order of magnitude or even two.
What your PCG system needs in order to work properly with AI:
- Same input → same output
- No hidden dependencies
- Stable seeds
- No nondeterministic sampling
Things that are problematic for AI:
- Uncontrolled random seeds
- Sampling from unstable geometry
- Floating-point dependencies on world position
- Hidden state inside the graph
Simplicity vs. Universality
For humans, the versatility of a tool is often more important than its simplicity.
A single flexible tool with many settings can be used in a wide range of scenarios. For the creative process this is convenient: an artist can adapt the tool to a specific task.
For AI, the situation is the opposite.
One complex tool with many parameters is often less effective than several simple, specialized tools.
Therefore, when developing systems intended for automated use, it is often necessary to:
- Simplify individual operations
- Break complex processes into several stages
- Automate auxiliary operations
- Move part of the logic to the system or actor level
Level of Abstraction
For obvious reasons, humans tend to prefer high-level tools.
In production pipelines, tools are often created that are tied to a specific setting or task. These may be ready-made compositions with a predefined set of parameters—for example, a road generator or a building generator.
Such tools work well for manual workflows. The artist gets a quick and predictable result without needing to dive into implementation details.
For AI, the situation is usually the opposite.
AI tends to work much more efficiently with low-level patterns and basic operations. For example:
- Along Spline
- Scattering on Surface
- Pattern distribution
- Simple rule-based placement
These kinds of operations are easy to combine and allow the system to orchestrate more complex structures at the pipeline level.
Complex high-level tools often turn into black boxes for AI.
The internal logic of such tools is hidden, the possibilities for combination are limited, and their behavior becomes harder to predict.
For this reason, when designing PCG systems intended for automated use, it is often more useful to provide simple low-level building blocks rather than ready-made compositional tools.
Critique or Observability
A human can easily analyze the result visually. They look at the level and immediately understand: here it’s too dense, here trees are placed on the road, here the turn is incorrect.
AI cannot perceive this directly. Therefore, tools designed for AI need to expose metrics, debug outputs, and structured data. For example:
- Density maps
- Collision counts
- Overlap metrics
- Distance violations
Computational Scalability
Humans usually run PCG locally: adjust the spline a bit, press Regenerate, and check the result.
AI, on the other hand, may generate hundreds of variations. Therefore, for AI the following are important:
- Fast execution
- The ability to perform batch generation
- The absence of heavy world queries