Varför Take-Twos före detta AI-chef är klar med att försvara generativ AI

Take-Two laid off its entire AI team in April as part of a restructuring, a move that read as odd given how openly the wider industry has been chasing generative AI. The catch is that most of the people let go were not working on generative AI at all. The team began at Zynga in 2019 as an R&D innovation group, three years before ChatGPT made the public equate AI with large language models, and stayed in place after Take-Two acquired Zynga in 2022. Its former head, Dr Luke Dicken, laid out his concerns in an interview with GamesIndustry.biz.

The group ran as a skunkworks in the basement of Zynga's San Francisco headquarters, testing how AI as a broad field could feed game development. Dicken's reference point was tabletop RPGs. He argues a good dungeon master reads what players want and bends the session to it, and that the same logic could shape a video game on the fly. The team built a machine-learning profile system that tracked around 40 metrics per player. The premise produced a game, 2020's Spell Forest, which Dicken says proved the thesis by moving core business numbers through changes to how a person experienced the game.

ChatGPT's 2022 release rewrote the team's job. When it emerged that user input could become training data, Zynga handed the group governance over generative AI across the company, mostly educating staff and approving tools. Of the 25-person team, only three or four worked on the generative side; the rest kept on with AI in the wider sense. Oversight later expanded to all of Take-Two before the group was retired this year and its duties spread to other teams.

Dicken does not speak like someone whose title was head of AI until recently. He frames generative AI as something he felt a moral obligation to manage rather than a technology he believed in.

"Generative AI is not something that I have ever been particularly passionate about. It was something I think there's a moral obligation to see managed as best as can be, but also on the understanding that for any big corporation in 2025/2026, no generative AI is the wrong answer that will get a lot of people's backs up."

— Luke Dicken

His objections run across three fronts: ethics, law, and business. He cites the lawsuit discovery showing Midjourney kept a list of artists whose work trained its models, and says he knows people on it. He questions whether the output is even good, calling LLMs a next-word predictor biased toward the mean. A strong coder, he argues, gets dragged toward mediocre work the same way a weak one does. He also flags how little it takes to break a tuned model, where one change to training data cascades through the network and ruins a use case that worked yesterday.

I see his point about regression to the mean as the most useful frame here, because it separates the marketing from what the tools measurably do. Take-Two's stock rose by 5% after GTA 6 pre-order announcement this week, and that gap, between a hand-built Rockstar product investors will pay a premium for and the averaged output of an LLM, is exactly the tension Dicken keeps circling.

There is one upside he credits to the hype. Five years ago, pitching a level-generation algorithm got him blank stares; now the same room nods along to almost anything labeled AI.

"It has made people more receptive to conversations about what traditional techniques could have done for them years ago. They are more inclined to believe things like that can exist."

— Luke Dicken

The worry is what happens when the bubble pops. Strauss Zelnick has pitched GTA 6 as a game that's "never been done before", the kind of ambition that takes years and human craft rather than statistical averaging, and Dicken fears the opposite outcome for the field: that a burst bubble sours the industry on AI entirely, traditional methods included. I think that fear is the right one to hold, since the backlash rarely sorts the useful tools from the overhyped ones. He invokes Ed Zitron's argument that the economics of the boom do not add up, and concludes the answer looks like no across ethics, law, and business while the spending continues anyway.

"My worry is that generative AI is poisoning the well. I don't think there is enough sophistication and nuance to retain the traditional stuff. For LLMs, we have already stumbled into the trough of disillusionment."

— Luke Dicken

How studios navigate it stays contextual, Dicken says. A startup with six months left will use every advantage it can. A larger firm has to weigh legal exposure. He lands on values: the morally correct answer is no generative AI, the business-correct answer is just enough, and where a studio draws the line depends on what it cares about. The same Zelnick interviews that explain why the development of GTA 6 took so long point to the other model entirely, a company spending 13 years to ship something it considers as good as it can possibly be.

Read also, Rockstar appears to be lining up a third GTA 6 trailer, having quietly stripped the "new" tag from Trailer 2 ahead of June 25 pre-orders, with insider DarkViperAU claiming the next trailer runs 20 minutes.