It’s been almost impossible to avoid the conversation about Generative AI in the games industry over the past couple of years. No matter which side you fall on personally – from starry-eyed evangelists, via the cautiously interested and the ethically concerned, all the way to the hard-bitten harbingers of dystopia – the discussion itself has never been off the agenda.
Much of the discussion, however, has involved more heat than light – for the simple reason that nobody really knew to what extent AI would actually be useful in game development. Lots of people made sweeping predictions – in line with their own opinions about the technology, for the most part – but there’s not been much in the way of actual hard data arising from genuine experiences of trying to incorporate AI tools into creative workflows.
Some of the most basic facts about AI’s potential and its role in the industry, in fact, have been incredibly vague. In what areas, and by how much, might it boost productivity? What tasks can and can’t it do? How much additional work is required to supervise and verify its outputs? Even more fundamental; how much will it cost?
Productivity gains from using AI tools are real – but they’re inconsistent, highly task specific, and extremely reliant on human supervision
In the past couple of months, it feels like some answers have started to emerge – not just for the games industry, but for every sector that has dipped its toes in these waters. Companies have been working with various types of AI tools for a while now; some have integrated it more deeply into their workflows than others, but almost every company in a tech-related sector has at least dabbled. Consequently, the body of actual experience and data around what AI can accomplish is growing rapidly – and with it has come a much-needed injection of reality to the debate.
The growing consensus seems to be that productivity gains from using AI tools are real enough – but they’re inconsistent, highly task specific, and extremely reliant on careful, skilled human supervision. There’s an undercurrent of cautious optimism about the productivity-boosting potential for certain tools, in certain areas – but little sign that the dream of massive wins at little or no cost might materialise.
Cost, in fact, is the second factor that’s changing the conversation materially. Less than a year ago, developers who voiced concerns over the actual utility of AI often struggled to get a fair hearing from senior management. The evangelising salespeople who had sold them on the dream of AI had also furnished them with ready-baked dismissive counterarguments; concerned developers were just Luddites (a moniker readily embraced by those who actually know the history of that unjustly maligned group) trying to ringfence their job security, or simply failing to understand that the breakneck progress of the technology would sweep away their arguments about quality and reliability.
The AI bills, both metaphorical and literal, have started to come due
Today, there’s a far more receptive ear for those concerns in a lot of boardrooms, and it’s largely because the AI bills – both metaphorical and literal – have started to come due. The past few months have seen an industry-wide shift towards token-based billing instead of flat rates, accompanied by a move away from the subsidised introductory rates many corporate clients had been enjoying. Microsoft’s Copilot services were one of the first major dominoes to fall, but the trend is more or less universal. The actual costs of AI are increasingly being passed through to customers rather than being soaked up by tens of billions of dollars of private funding, and executives who last year wanted AI integrated into every facet of their company’s business are belatedly starting to wonder about the long-term costs involved.
Those changes, of course, hit hardest of all for use cases that burn huge numbers of tokens – use cases like generating complex assets, or working with extremely large codebases, scenarios which are rapidly becoming far more expensive as realistic, unsubsidised pricing takes hold. As those bills mount up, AI needs to have a business case that’s based on more than hype, vibes, and all-consuming FOMO; high costs need to be matched by real, measurable productivity gains and cost savings on the other side of the equation.
This has taken some of the puff out of the irrational exuberance that characterised AI discussions at management level in many companies. As the data has started to roll in from actual use cases, it can charitably be described as a mixed bag.
As is often the case with divisive new technologies, it does at least seem that AI is not as bad or as useless as its loudest detractors claim. There’s genuinely useful stuff in amongst the froth of countless startups trying to turn every application into a chatbot, and it does seem that a lot of developers have quietly integrated certain AI features into their workflows. Many good programmers, for example, eliminate a lot of repetitive donkey-work with code completion tools in their IDEs, while automated code reviews do a reasonable job of catching some bugs – both solid, albeit incremental, productivity improvements.
These solid little productivity gains are a long way from the dream that executives were sold
Similarly, some artists find that image editor tools based on deep learning models do a solid job of speeding up tedious parts of their workflows. AI tools also do a reasonable job of some managerial slog, like transcribing and summarising team meetings. These things are not nothing – they’re solid little gains that free up staff to spend more time applying their skills to more interesting and complex tasks.
Those gains, however, are a long way from the dream that executives were sold. Those developers who have tried to use more complex AI tools in their workflows, often being pushed to do so from senior echelons of their companies, generally seem far less enthused by the experience. Setting agentic AI tools loose on game codebases reportedly runs into hard limits very quickly; the codebases are too big, too complex, and too specialised, and any code produced by the agents needs to be extremely carefully vetted by senior developers – a dull and time-consuming task.
Generative AI in art workflows, meanwhile, seems to be really poor at maintaining visual consistency in assets it produces, and that’s even before the as-yet-unresolved question of whether anything it creates can be copyrighted (most legal interpretations still lean towards “no”, a huge problem for any studio using those assets in games). Like the programmers, artists who have had to work with generative AI tools report that senior art staff have to devote large amounts of time to supervising and checking the output; rather like having a very fast but extremely unreliable junior staff member on your team (one whose salary demands keep soaring from month to month, at that).
Generative AI is like having a very fast but extremely unreliable junior staff member on your team
AI optimists will argue that these tools will get better over time – the core technology underlying them is still young, after all – but the escalating cost and rapidly approaching horizon for hefty token price subsidies may make the ongoing improvement of the technology itself somewhat moot. Almost every attempt to meaningfully improve these tools relies on essentially adding more token processing; “thinking” models, for example, are essentially a whole bunch of LLMs talking over and back to one another in order to formulate the final response to the user, which means that each of their responses burns an order of magnitude more tokens than you see on screen.
Even if the underlying costs of processing tokens and training new models fall (though the countless billions of capital expenditure being mooted by AI-related companies suggests that costs are a long way from actually falling), those savings are immediately swallowed by new attempts to abstract away the non-determinative, error-prone nature of AI systems by adding more layers of processing, more AIs checking each other’s work, and bigger, more computationally expensive context windows.
What emerges most clearly from talking to people with actual experience of these tools is that there has been a sharp enthusiasm gap between developers themselves and executive-level decision makers. That’s historically been a huge red flag for any new technology, because successful, meaningful technology adoption is almost always initiated from a bottom-up pattern.
Artists and engineers in this industry are quintessential early adopters; they’re enthusiastic to try out new technology
Artists and engineers in this industry are quintessential early adopters; they’re enthusiastic to try out new technology and will lobby their leadership to invest in it if it’s genuinely useful. When you see the anti-pattern of that emerging – executives pushing a buzzword-heavy new technology at reticent development staff – it’s never a good sign. None of us, I suspect, need reminding that the last time we saw that exact pattern in the industry, it was for NFTs – an ill-fated borderline scam of a technology whose six months in the undeserved limelight now feel like a fever dream.
The games industry has always embraced new technological trends and it’s unlikely that any early misgivings over the ethics of AI will stop actually useful tools based on the technology from seeing widespread adoption across development studios. The radical transformation that some people expected, however, is not materialising – and now that the bills are coming due, even the most ardent evangelists are starting to back away from a cost to utility calculation that looks far less rosy than they hoped.
Meanwhile, there’s also a growing (though sometimes rather grudging) acceptance that rather a lot of consumers utterly loathe AI-created art or music, which is a nail firmly hammered into the coffin of many possible use cases. If AI had turned out to be a true revolution in productivity and development cost, that’s a cultural battle some executives might have been willing to take on – but with little evidence emerging of any such revolution, nobody’s going to want to get their hands dirty in that fight. As the dust slowly settles, it’s looking increasingly likely that AI tools will be useful in the long term – but in a vastly more limited and far less impactful manner than the past few years of hype suggested.
