A.I. May Help Design Your Favorite Video Game Character

Intrigued by the potential that generative artificial intelligence holds for video game design, the studio Blizzard Entertainment has trained an image generator on its own hit titles. By feeding assets like the combative orcs of World of Warcraft, the macabre dungeons of Diablo and the vivacious heroes of Overwatch into the machine, Blizzard can effortlessly produce concept art for new ideas.

Because generative artificial intelligence creates art faster than any human can, studios like Blizzard, a division of Activision Blizzard, are hopeful that the technology can cut out some design and development drudgery and make the creation of video games more fun.

Blizzard’s chief design officer, Allen Adham, told employees about the initiative last month in an email that was obtained by The New York Times. Its internal tool is called Blizzard Diffusion, a riff on Stable Diffusion, one of the popular image generators that enables anyone to turn text into art.

“Prepare to be amazed,” Adham wrote, adding, “We are on the brink of a major evolution in how we build and manage our games.”

Generative artificial intelligence, the technology behind tools like ChatGPT and Midjourney, uses considerable computing power to identify patterns in text or images and produce new content from the data.

Some researchers are wary of the technology, warning of copyright abuses, job displacement, and its potential to help the spread of false information. But video game developers, already relying on artificial intelligence so that nonplayer characters can make humanlike decisions, believe that harnessing generative A.I. can speed up the creative process in a labor-intensive industry plagued by delays.

There is a gaming A.I. division at Microsoft, which makes the Xbox console, and Ubisoft has built a tool called Ghostwriter that could produce basic dialogue for games like Assassin’s Creed. Several start-ups say their technology can make it easier to design the look of the nonplayer characters, known as NPCs, that give video game worlds heft.

Chris Lee, the former studio head of Halo Infinite at 343 Industries, said generative A.I. could improve game development by reducing the human toil required to make an enormous open-world game.

“Game developers have never been able to keep up with the demands of our audiences,” said Lee, who is now the head of immersive technologies at Amazon Web Services.

Halo Infinite was supposed to be the flagship launch game for the Xbox Series X in 2020, but its graphics were derided by fans as flat and ugly after an eight-minute preview was released. The studio ultimately delayed the game for another year.

The game’s developers were miserable, Lee said, because even working on place-holder encounters required slowly moving pixels frame by frame. “To load this giant world, it’s painful, it’s like specialized data entry,” he said.

Generative A.I. could also streamline quality assurance testing. At a recent conference for game developers, Kate Rayner, technical director for the Coalition, the studio behind Gears of War, talked about how A.I. could be used to catch bugs and glitches so players would see fewer crashes on launch day.

Many of the promises of generative A.I. are speculative, with Blizzard already abandoning machine-learning technology it had patented to create environmental textures like stone and brick.

Andrew Guerrero, Blizzard’s vice president of global insights, said the tool was taking up too much artist time to be effective. But he said another A.I. tool was helping fit cosmetic headpieces to player models in World of Warcraft.

“The goal is to remove a repetitive and manual process and enable artists to spend more time on creativity,” Guerrero said in a statement. “Our goal with A.I. has been, and will continue to be, to try to make creative work easier.”

The internal email about Blizzard Diffusion said it was being used to help generate concept art for game environments as well as characters and their outfits. It also mentioned possible tools for “autonomous, intelligent, in-game NPCs,” “procedurally assisted level design” and A.I.-assisted “voice cloning,” “game coding” and “anti-toxicity.”

Ghostwriter, Ubisoft’s A.I. dialogue tool, was a request by writers who faced the daunting and sometimes tedious task of filling open-world games with more than 100,000 lines of dialogue, the company said.

Many of those lines are the background chatter of characters that help simulate a living world; one mundane interaction may require a dozen or more variations.

In a promotional video for Ghostwriter, an employee begins with the prompt “I used to be an adventurer like you” (a nod to an infamous line in Skyrim) and hones several suggestions by the A.I., including “I was once the most talented and famous adventurer in the land” and “I remember when I was young and strong.”

These simple lines of dialogue have been a way for people to start careers in video game writing, and developers argued on social media that automating these tasks could threaten such jobs. Simon Johnson, an economist at the M.I.T. Sloan School of Management who has a new book about the impact of automation, said it was a bad idea for tech companies to invent algorithms that mimic humans.

“We should be focused on machines that help humans improve human capabilities, rather than displacing people,” he said.

Yves Jacquier, the executive director of Ubisoft La Forge, the research and development team responsible for Ghostwriter, said there had been a similar but unfounded fear that video game animators would be replaced when motion capture was introduced decades ago.

“While the future may involve more technology, it doesn’t take away the human in the loop,” Jacquier said in a statement. “Artists, writers and coders will always be at the heart of the development process, and while A.I. can now do a better job at assisting creators in their workflow, it’s the artistic vision and perspective of individuals that are essential in the creation of games.”

Another inescapable concern about A.I.-produced content is copyright. In one high-profile lawsuit, Getty Images has accused Stable Diffusion of scraping 12 million images from its photo database.

Employees of Activision Blizzard received an email this month from Michael Vance, its chief technology officer, that warned them not to use the company’s intellectual property with external image generators. (Microsoft is looking to purchase Activision for nearly $70 billion, but regulators want to block the deal.)

“These new tools come with new and unknown risks, and we will proceed carefully to avoid pitfalls,” Vance wrote in the email, which was obtained by The Times.

Some Activision Blizzard employees said the company’s A.I. tools have not always delivered as promised, pointing to those that struggled to catch bugs or interact properly with game environments.

“Leadership’s focus on A.I. doesn’t feel like it is solving a problem that individual contributors care about,” said Valentine Powell, a former World of Warcraft engineer who left Blizzard last August. “It feels like ignoring their problems and focusing on hype words that they think will sound impressive to shareholders.”

Smaller video game studios that do not have the resources to create generative A.I. tools are turning to start-ups for help.

Scenario, which raised $6 million in seed funding in January, creates image databases to turn text prompts into assets — such as a lizard in a spacesuit — that can then be incorporated into games.

Its image generator is being used by a few small games. In tests by The Times, it sometimes spit out animated characters with unrealistic-looking hands, a common weakness for image generators.

Another start-up, Didimo, said that Soleil Game Studios, which makes fighting games based on properties like Naruto and Samurai Jack, had created hundreds of nonplayer characters using its A.I.-powered generator.

“We just automated that process, allowing them to remove the mundane jobs, because it just gets boring after a while,” Sean Cooper, a Didimo developer, said.

Video games can take years to create, and it is early enough in the hype cycle that few published games have used generative A.I. But some companies like Niantic, the developer behind Pokémon Go, have experimented with the tools to create marketing materials, for instance writing a news release in Dr. Seuss’s tone of voice.

Niantic also used ChatGPT while working on Peridot, a new augmented reality game that leverages ’90s nostalgia for digital pets.

Kellee Santiago, Niantic’s head of production, said that it was expensive to fill a room with creative people, and that the technology condensed the time needed to generate ideas.

“I, for one, am happy to have the first six hours of a brainstorm taken care of for me,” she said.