Back at the start of April, it emerged that Take-Two had laid off its AI team as part of a restructuring initiative.
Layoffs are, sadly, not a new occurrence. Layoffs in AI, a technology that most game publishers are heavily investing in, are more notable. Most reporting, that of GamesIndustry.biz included, referenced Take-Two CEO Strauss Zelnick’s previous assertions that the firm was embracing generative AI – removing AI experts from the payroll appeared to be counter-intuitive. The seeming contradiction speaks to the challenges that a long-established specialisation in game development faces as a new, highly disruptive technology with the same name forces its way onto the scene.
The affected staff weren’t doing things with generative AI at Take-Two. Rather, the team was founded in 2019 at Zynga as an “R&D innovation group”, former head of AI, Dr Luke Dicken, PhD, tells GamesIndustry.biz. This is back before ChatGPT helped generative AI to become equated with the notion of AI more generally in the public consciousness, when in reality it is a much wider field. It was also before Take-Two acquired Zynga in 2022.
The venture was a skunkworks in the basement of Zynga’s San Francisco headquarters, exploring how AI, as a broad technology, could be used in game development.
“My life’s work is looking at what games can be and pushing on that harder,” Dicken says.
“One of the big touchpoints for me is tabletop RPGs like D&D. That is now a 50-year-old game, and it has retention like you don’t get in video games. My thesis is that the human intellect managing the game experience is what makes a good TTRPG good, and it also provides a strong model for what AI in games could be.”
Much of what a dungeon master does in a game like D&D is social profiling and matchmaking – figuring out who is going to play together, sit together, and so on. From there, it comes down to managing how someone wants to play the game; sometimes you’ll give in to their desires, other times you’ll go against them.
“Good DMing starts with understanding what people systemically want from a game,” Dicken says.
“What if I could figure out what players want and change the game for them on the fly?”
From there, Dicken saw that AI might be a good way to make video games – in particular, the mobile games that Zynga specialised in – stickier and more cost-effective in terms of acquiring new users. So the team developed a method for customising a game using a machine-learning profile system that analysed around 40 different metrics to see how a person interacted with the game.
“My thought was, ‘What if I could figure out what players want and change the game for them on the fly?’,” Dicken explains. “Someone comes into FarmVille, we do an analysis on how they are playing the game, see that they don’t like farms, and you can do an asset swap and suddenly it’s CityVille. It’s the same game, but now has a slightly different texture and vibe.”
Based on this premise, the team released a game: 2020’s Spell Forest. Not only did this prove Dicken’s thesis correct (“We could measurably impact core business KPIs by fiddling with how someone plays and what their experience was”), it also put the group on Zynga executives’ radar. Dicken’s team was suddenly supporting a number of different teams at the company.
“We wanted to change the conversation around AI,” Dicken says of his team’s mission in early 2021. “And the monkey’s paw curled.”
The generative AI takeover
In 2022, OpenAI released ChatGPT to the public. This was the first mass market generative AI tool and, as Dicken remembers it, “those early days were the absolute wild west.”
While there was a lot of hype around this new technology, and many people were very impressed by what it could do, no one really understood how ChatGPT worked at first. For example, it was revealed that the data users input into the tech could be used as training data.
“That set off every alarm bell at every corporation on the planet,” Dicken recalls.
Zynga’s management saw that it already had an AI team sat in its basement, so it handed the group governance for the use of generative AI for the entire company, a task that mostly took the form of educating staff on “concerns and challenges.” All generative AI tools had to go through the team for approval.
The AI team was 25-strong, but only three or four members of staff were working on the genAI side; the rest were continuing their work on AI writ large.
“You have to know how best to pick your battles”
After being handed the reins for genAI at Zynga, the team was later given oversight of Take-Two’s tech as a whole. But ultimately, the group was retired earlier this year as the company transitioned many of its governance responsibilities to other teams.
“Generative AI is not something that I have ever been particularly passionate about,” Dicken says. “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. This is an incredibly polarising topic on both sides. My position, more than anything, is that you have to know how best to pick your battles. Some of the excesses of genAI are so egregious that you need to make sure you’re able to push back.”
Serious concerns
You might expect someone who, until very recently, had the job title ‘head of AI’ to at least have a loosely positive view of generative AI, but Dicken is forthright in his concerns about how it can be used.
He has ethical concerns about how large language models (LLMs) have been trained, citing the revelation during the discovery process of a lawsuit that Midjourney had been keeping a list of artists whose work had been stolen to train its models. Dicken says he knows artists on that list, as well as writers whose work has unknowingly been used to train LLMs.
“I want to be able to look my friends in the eye and know that I am not making their life worse,” he says.
“If you are a good coder, these systems can also make you a mediocre coder”
Another concern Dicken expresses is whether these models are even that good at the jobs they are meant to do, pointing to the technology underlying LLMs and how literally average their output can be.
“At some level of abstraction, LLMs are a next word predictor system,” he says. “That means it is statistically always going to be biased towards the mean of what it sees in its data. If this stuff has to exist, I’d like it to generate good stuff rather than mediocre stuff. If you don’t know code or are a bad coder, AI can make you a mediocre coder. But if you are a good coder, these systems can also make you a mediocre coder. It’s regression to the mean as a service. That’s concerning to me.”
There’s also the small matter of change management; Dicken says that it does not take much to change how some of these models operate. Small changes can have a massive impact on an LLM, fundamentally altering it.
“The fact you don’t have control over that scares the shit out of me”
“Let’s say that we have a task writing email marketing,” he says. “We have a pipeline that can kick out marketing for a game, we’ve tested a variety of models, and we’ve found the one that is better at sticking to the brand voice. It doesn’t take much change to the training data or the way it has been pre-trained. One little change somewhere can have a cascading effect across the neural network. What was good in that one use case isn’t now. You are effectively outsourcing whatever the use case is to someone you gave the test to one day, and, in an anthropomorphic context, a different person might come in tomorrow. The fact you don’t have control over that scares the shit out of me.”
There are also concerns about the business fundamentals of the AI industry. Dicken cites writing by tech columnist Ed Zitron, who argues that the economics underpinning the current AI boom don’t stack up.
“On the one hand, you have all the ethical and moral questions,” he continues. “Then you’ve got your legal questions, and now you have to ask if this is good for business. The answer seems to be no on all three, and yet here we are.”
Room for improvement
The lasting impact of generative AI on games and game development is yet to be seen. The tech has already been employed across a number of titles, with titans of the industry saying that soon it will be a fixture of how tomorrow’s hits are made.
Though he has obvious concerns about generative AI, one positive that Dicken has seen thanks to the technology is a greater willingness to be imaginative about how tech can improve games.
“Five years ago, you’d say you have an algorithm that will be really beneficial for accelerating level generation content in a mobile title,” Dicken says.
“It has made people more receptive to conversations about what traditional techniques could have done for them years ago”
“Back then, people looked at us like we had two heads. Now, the hype of AI has created an environment where I could tell you that AI is going to be the thing that moves your game to quantum computing, and people will nod and say: ‘Yeah, we want AI in the game’, and I think: ‘Great, love that, wish you had a more nuanced take on this, but sure’. 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.”
However, Dicken says that the hype cycle around generative AI isn’t anything new. There has been an ebbing and flowing of excitement around artificial intelligence for decades.
“The problem is that because the hype people come in, overblow what it is, try to eke out all the funding they possibly can, but it doesn’t deliver on the hype,” Dicken says. He’s concerned that if the generative AI bubble pops, it will leave a bad taste in the mouth of everyone in the industry. Rather than realising that more traditional AI techniques can help with development, people might just want shot of the technology as a whole.
“My worry is that generative AI is poisoning the well,” he says. “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.”
Untangling the knot
Dicken sees ongoing use of generative AI as something that developers and publishers will need to navigate based on a series of often-contradictory standards.
“It’s so contextual,” he says. “If you’re a tiny startup and you are going out of business in six months, why wouldn’t you use every advantage available? If you want to genAI all the things, it’s important that you understand the ethical implications of that. It’s a really hard one to wrestle with. If you have to pay minimum wage to bring in an artist, it’s already been an accepted practice to bypass minimum wage in your region and outsource it to somewhere where there’s a cheaper cost of living and pay below minimum wage for the same output. In some ways, how is this different? But at the same time, I think it is. Maybe that’s because when you are outsourcing to a lower cost-of-living area, you’re still paying a person.”
Ultimately, Dicken comes back to the three areas he mentioned earlier that need to be considered: ethical and moral, legal, and business. For smaller companies, the business conversation is probably what you need to consider; larger firms likely have to look at legal concerns.
“All of it is a really intertwined ball of string, and it’s hard to figure out where the start is to untangle the knot,” Dicken says. “At the same time, why do we have to untangle this knot? Other than a whole bunch of people with money in San Francisco say we have to. Some of this is about what you value as a game studio. Do you want the best output ever? You probably won’t get that from really permissive use of these systems.”
He concludes: “The morally correct answer is no genAI. The business correct is just enough genAI, and where you draw that line is going to be values-dependent.”
