Valve Explores AI for NPC Reactions, Not Creative Writing
A small group at Valve has been experimenting with generative AI for video game applications, but the company has no formal initiative to deploy the technology across projects, according to Half-Life 2 and Portal writer Erik Wolpaw.
Wolpaw said the experiments focus on one specific area: AI-generated dialogue for non-player characters reacting to unpredictable player behavior. He dismissed generative AI's usefulness for creative work like novel-writing or comedy.
Where AI Shows Promise
The technology excels at one thing: accepting whatever players do and adjusting responses accordingly. Wolpaw imagined a Grand Theft Auto scenario where AI characters react as "straight men" to player chaos, simply rolling with whatever happens.
"It's very good at just going along with whatever insane thing you say and kind of adjusting to the flow of that," Wolpaw said in a recent podcast interview.
Game writing has always required simulating character reactions to player input. Left 4 Dead, Wolpaw noted, uses conditional matrices-if X happens and Y happens, play this dialogue line. AI could expand those possibilities beyond the limited verbs in traditional games (jump, punch, shoot) to actual speech.
Current Limitations
Wolpaw acknowledged the technology remains expensive to deploy at scale and noted that generative AI fails at being funny or especially creative. These experiments carry no official Valve endorsement beyond the fact that employees are conducting them on company time.
The work isn't attached to any particular game project.
Broader Concerns
Wolpaw's observation about AI "going along with whatever insane thing you say" extends beyond games. Military planners use similar generative AI models supplied by Google, Microsoft, and OpenAI to process intelligence data for targeting decisions. Researchers have identified a risk called "sycophancy"-the tendency of AI to align outputs with user preferences, even incorrect ones.
The distinction between game-focused models and military-grade systems matters, but the underlying design principle remains the same: generative AI models are built to be responsive and adaptive. That responsiveness creates both opportunity and risk depending on context.
For writers specifically, understanding AI's actual capabilities in dialogue and narrative work matters more than hype. Wolpaw's assessment-that the technology is poor at creativity but useful for reactive systems-provides a clearer picture than most industry claims.
Your membership also unlocks:
