Bing Gordon on AI and games: agents are the new outsource, but building the system underneath is still the hard part

Bing Gordon on AI and games

Bing Gordon, co-founder of Electronic Arts and author of Everybody Wins, argues that AI's most immediate value in game development is eliminating grunt work — the equivalent of going "from white out to word processors." The people who benefit most are indie creators who currently spend half their time fighting tools like Adobe Photoshop pixel by pixel rather than designing games.

Agents are the new outsource

Gordon draws a direct line from EA's 1990s expansion to today. When EA moved from console to CD, it increased its graphic artist headcount by a factor of ten, mostly with people who didn't play games. That kind of non-fan production work was eventually outsourced to Eastern Europe and Asia. AI agents, he argues, fill that same role now — and the analogy is deliberate, not flattering. It's a cost and capacity argument, not a creativity one.

“The promise of AI for games is a couple fold. One is more iteration... Better people get more iterations out. I think the AI is going to be able to create physical assets with world, not quite yet, but almost. But the system underneath it is always the hard part... Electronic Arts in the nineties when we went from console to CD, we increased our graphic artist headcount by a factor of 10. And most of those people did not play games. For some of this stuff, agents are the new outsource.”

The system underneath is still the hard part

Where Gordon is more skeptical is on the deeper promise. He's working with companies trying to operate as mostly agentic studios, and his blunt assessment is that getting from a rough first draft to a shippable product has taken a year longer than anyone expected. The bottleneck isn't generation — it's system design. Game mechanics layer on top of each other the way Magic: The Gathering's math does, requiring empirical tuning that took Wizards of the Coast's team a full year and Pokémon two. AI can iterate faster, but it hasn't yet shown it can build in layers reliably.

The prose problem in games illustrates this precisely. Players ignore quest text unless it contains a gameplay clue. To prompt AI to write useful in-game language, you first have to know what the underlying system requires — and at that point, Gordon says, it's almost easier to write the prose yourself than write the prompt.

NPC characters remain the long-term prize

Gordon's original vision for virtual worlds traces back to working at Disneyland as a kid and wanting to actually talk to the Abraham Lincoln audio-animatronic, walk the park with him, hear his stories. That ambition — genuinely interesting dramatis personae — is what he thinks AI could eventually deliver. Current models aren't close. But it's the use case he's waiting for.

Hallucination as feature

One area where Gordon is genuinely enthusiastic: controlled hallucination for creative ideation. His framing is that you let the model "hallucinate its butt off," collect the shitty first drafts, and then apply human judgment to find the gems worth building into actual game systems. The danger isn't hallucination itself — it's hallucination that doesn't loop back into coherent mechanics.

Iteration as the underlying theory

Running through all of Gordon's AI commentary is a consistent thesis about iteration. He cites Franz Johansson's The Medici Effect, which finds that the most successful patents come from people who file the most patents. David Kelley's IDEO method — 100 writable ideas in an hour — reflects the same principle. Better creative people get more good iterations out per unit of time, and AI's real promise is compressing that cycle, not replacing the judgment about which iterations matter.

The honest near-term picture Gordon paints is narrower than the hype: faster asset production, cheaper iteration on mechanics, and agents absorbing the non-fan production work that used to go offshore. The grander vision of living, talking game worlds remains a design problem that generation alone can't solve.