Setting aside the hype, here is where teams we talk to are getting real value, and where they are not.
Working well right now:
- Tooling and pipeline. Batch renaming, asset triage, generating placeholder data, writing editor scripts. Unglamorous and genuinely saves days.
- First-pass content that a human then edits. Item descriptions, barks, localisation drafts, test data. The edit pass is not optional.
- Code assistance in the engine. Good at boilerplate, shader scaffolding and explaining an unfamiliar codebase. Still needs review; it will confidently write something that compiles and is wrong.
- Playtest analysis. Feeding session data in and asking where players stall.
Not working well yet:
- Runtime generative NPC dialogue in a shipped game. Latency, cost per session, and the fact that you cannot guarantee what it says. Most teams trying this end up constraining it so tightly it becomes a state machine with extra steps.
- Generated art going straight into a build. Style consistency across hundreds of assets is the problem, not the individual image.
- Anything where the output has to be exactly right every time. That is what code is for.
The pattern worth noticing: AI is good at the first draft and bad at the last mile, and games are almost entirely last mile.
One thing to be careful about commercially: if generated assets go into something you intend to sell or licence, get clear on the provenance and the licence terms of whatever produced them before it is baked into your build.
What are you using it for? Interested in what is actually working for people rather than what is being announced.
Working well right now:
- Tooling and pipeline. Batch renaming, asset triage, generating placeholder data, writing editor scripts. Unglamorous and genuinely saves days.
- First-pass content that a human then edits. Item descriptions, barks, localisation drafts, test data. The edit pass is not optional.
- Code assistance in the engine. Good at boilerplate, shader scaffolding and explaining an unfamiliar codebase. Still needs review; it will confidently write something that compiles and is wrong.
- Playtest analysis. Feeding session data in and asking where players stall.
Not working well yet:
- Runtime generative NPC dialogue in a shipped game. Latency, cost per session, and the fact that you cannot guarantee what it says. Most teams trying this end up constraining it so tightly it becomes a state machine with extra steps.
- Generated art going straight into a build. Style consistency across hundreds of assets is the problem, not the individual image.
- Anything where the output has to be exactly right every time. That is what code is for.
The pattern worth noticing: AI is good at the first draft and bad at the last mile, and games are almost entirely last mile.
One thing to be careful about commercially: if generated assets go into something you intend to sell or licence, get clear on the provenance and the licence terms of whatever produced them before it is baked into your build.
What are you using it for? Interested in what is actually working for people rather than what is being announced.