Dr Tommy Thompson is the founder and director of AI and Games, and was previously an AI researcher and college lecturer in recreation growth. Tommy co-founded Sport AI Occasions CIC in 2024: a non-profit organisation internet hosting occasions for recreation AI professionals.
This text is a part of AI Week.
Making an attempt to qualify the place I sit in the house of synthetic intelligence for video games isn’t simple. I have been in and round AI for over 20 years, and issues have by no means been extra difficult. I used to hypothesise in analysis papers that AI may have an actual influence on the games industry. Sadly, I have been confirmed proper for all the improper causes.
I now begin each introduction with: “I work in AI. Oh, however not like that.”
Synthetic intelligence and online game growth are two fields that profit from one another, whereas being poorly understood by these not concerned in both or each. It is this confusion and ambiguity that has pushed a lot of my work at AI and Games over the previous decade: advocating for attention-grabbing improvements, educating on rising greatest follow, and holding us to account.
We’ve seen super progress in helpful, sensible and significant adoption of AI in the games industry in the previous 15–20 years. However in fact, the headlines have centered on nearly every little thing else.
The growth (and bust?)
Since GPT4’s launch in 2023, our information feeds have been full of all types of pontificating about the “unrealised alternatives” generative AI presents. AI will probably be the great equaliser, a device to make games faster and cheaper. Games will react to customers in extra immersive and partaking methods, crafting bespoke storylines, with characters that make the common private, the finite infinite.
You may instantly see how that is interesting to an industry obsessive about slicing prices and rising recurrent consumer spending. These sunny uplands are however a handful of prompts away, whereas the very actual issues that exist with quite a lot of generative AI are pushed beneath the rug, be it violations of mental property rights, the inconsistent and often fabricated outputs, power consumption equalling that of major cities, lawsuits, and the cascades of slop now making the web even worse than it already was.
I think you’re sick of listening to it. But it’s inescapable. New AI fashions dominate the information, whereas buyers are asking studios or publishers their plans for “embracing” AI. Extra not too long ago, the shortage of DRAM provide led to by the rush to construct AI knowledge centres has resulted in prices rising and GPUs being delayed, and fairly quickly consoles will bear the brunt of it as effectively. Does anybody assume that the new Steam Machine will launch beneath $1,000? At this level, I am uncertain.
I am unable to and will not sit right here and defend generative AI as an industry: that’s, the assortment of huge tech firms mixed with start-ups resembling OpenAI and Anthropic that shovel AI to the plenty. It is a determined try to make a mishmash of unreliable instruments right into a worthwhile commodity. With over half a trillion dollars invested to this point, there’s little to no dependable path to a return on funding (ROI).
Sporting each my “former AI researcher” and “games industry skilled” hats, I can see these works from either side. Google DeepMind’s Genie 3 – the AI system that not too long ago tanked stock prices of games firms – is a big innovation in what we name “world fashions,” a breakthrough thought inconceivable just a few years again. But donning my recreation developer hat, I can see that it is largely ineffective. These fashions are astronomically costly, battle with consistency, and are extremely restricted in their constancy. There’s potential to a point (see Microsoft’s Muse, which is a way more subdued effort with good intentions), however as soon as once more we’re advised it will essentially change how games are made for the higher, by folks with little to no grasp of how games are literally constructed.
Vapid PR statements, company greed and desperation, mixed with an more and more vitriolic response on social media has meant there is not any significant discourse on AI. It robs us of getting an actual dialogue about the true breakthroughs, and what a significant path seems to be like for AI in games.
It is time the games industry regained management of the narrative.
Discovering that means in the noise
We’ve a protracted and storied historical past with AI, courting again to the very first non-player characters (NPCs) and procedural era strategies in the Eighties. The golden age of “recreation AI” in the late Nineteen Nineties and early 2000s introduced us games like Quake III Enviornment, Half-Life, The Sims, Halo 2, and F.E.A.R., which standardised a lot of the strategies we depend on for NPC design to today. In the meantime, in the previous 20 years, we have discovered how machine studying (ML) could be embraced all through manufacturing.
Traditionally, ML had little influence on gameplay, with solely a handful of games resembling Creatures and Black and White lingering in the gaming consciousness. However ML has confirmed much more priceless elsewhere, resembling in Xbox Reside’s TrueSkill matchmaking algorithms, which launched with Halo 2.
AI-driven animation was pioneered via Movement Matching, first in Hitman Absolution, and is now a staple in Unreal Engine 5. We use participant analytics throughout the largest titles in cell and AAA. Sweet Crush Saga developer King has created leading edge ML instruments in content material era, automated testing, and participant profiling – just for them for use as a handy excuse for layoffs a decade after they began constructing them. Cheat and toxicity detection helps shield our on-line communities by observing textual content chat, voice chat, and even gameplay.
Ubisoft makes use of ML-powered bots to emphasize check updates to Rainbow Six: Siege and For Honor. EA trains ML bots for for QA duties in Battlefield, NHL26, and EA Sports activities FC. Sq. Enix exams recreation balancing at huge scales with ML for its RPGs. The Snowdrop engine even has its personal ML-powered GPU profiler to assist help cross-platform growth that essentially modifications how games are being ported to consoles.
AI shouldn’t be coming to the games industry: we’re certainly one of the main sectors in embracing and making use of this expertise. So it is disheartening to see how the AI industry has began dictating coverage.
It isn’t black and white
We’re at a vital inflection level in the games industry: the market is unstable, hundreds of builders are dropping their jobs, and as a era of expertise is phasing out, new expertise is struggling to get in.
So what higher time for generative AI – a sub-discipline of ML that crafts “content material” from statistical inference – to come up. It guarantees to be a silver bullet that may ship virtually immediately, changing the want to rent expertise, slicing prices on manufacturing, saving studios the must scale up, bringing product to market sooner, and reacting to and partaking with gamers dynamically.
Unsurprisingly, the fact is much extra difficult. It is neither the balm for all of our ills, neither is it (when used correctly) a unending plagiarism engine. Very like ML earlier than it, generative AI shouldn’t be inherently evil: it is about the intent and goal of how it’s designed, constructed, skilled, and deployed. In any case, I can practice an LLM on my laptop computer utilizing ethically sourced knowledge if I do know what it’s I am attempting to do, and fine-tune it to my wants if the drawback itself is exact and particular.
We have already got a lot of examples of fine follow for generative AI in games if you realize the place to look. We see the likes of EA’s AgentMerge for consolidating redundant JIRA tickets, or Infinity Ward’s multimodal search engine for assets. Upscaling applied sciences resembling DLSS are derived from generative strategies, enabling high-quality ports of Cyberpunk 2077 and Star Wars Outlaws to the Change 2, or in the case of God of Struggle: Ragnarok, reducing texture sizes on the PlayStation 5. The truth is, the dev workforce behind Mass Impact Legendary Version used a extra primitive model of generative AI to do a first-pass on upscaling textures earlier than artists cleaned them up by hand. In the meantime, Twitch is utilizing LLMs for improved toxicity detection. New studios resembling BitPart, Lingotion, That means Machine, Uncooked Energy Labs, and Studio Atelico are exploring easy methods to craft new gameplay experiences utilizing ethically developed and fine-tuned generative fashions which might be usually working on-device, moderately than in the cloud.
“Generative AI won’t ever obtain the identical stage of high quality as a human in any inventive medium”
As a lot as I agree with the extremely unfavourable – and justifiably indignant – sentiment in direction of generative AI, we have to get previous this two-sided debate, and as an industry begin having a extra significant dialogue of the place this expertise is acceptable, if in any respect. Generative AI won’t ever obtain the identical stage of high quality as a human in any inventive medium, interval. So far as I am involved, this isn’t open to debate. However like different AI approaches, there’s worth once we recognise the issues our persons are attempting to resolve, we strategy them with the proper intent, and face them with out compromising ourselves in the course of.
It is about displaying resistance to what’s usually a determined try to flog generative AI with out goal. At a GDC roundtable in 2024, when discussing the explosion of AI firms promising the Earth, I said that “I do know a hammer in determined search of a nail after I see one” – a line that later resurfaced in Rez Graham’s fantastic GDC 2025 talk. I stated it as a result of I’ve seen all of this earlier than, albeit at a smaller scale.
AI advocates – be they researchers, builders, buyers, or enterprise folks – are extremely vulnerable to overestimating the competences of those applied sciences in real-world eventualities, even with the better of intentions. I’d know, as a result of I used to assume the identical manner.
For each acquire, the place is the loss?
We have to have a extra severe dialogue throughout the sector of fine follow in AI adoption. This could take the type of higher training at developer, writer, and commerce physique ranges, sharing success tales and failures, and recognising the place this tech merely doesn’t work when it faces harsh manufacturing actuality – a lesson being discovered the world over proper now in virtually each industry.
Plus, for each acquire proffered by generative AI, we should contemplate the losses incurred in consequence. What does it say about us as an industry if each artist, designer, author, voice actor, performer, group supervisor, and even programmer is apprehensive that they are going to get replaced by subpar automation?
It isn’t going to be a straightforward course of, given a lot of the massive AI gamers even have vital management over games studios and publishers. However neither can we sit right here and have one other yr of nothingburger PR statements, slop creeping into AAA merchandise (having one way or the other slipped via evaluation, once more), and a chipping away at the collective morale and livelihoods of an industry that simply desires to get on with making nice games.
“We have to take the narrative again, and begin displaying a path ahead that highlights how we will use AI responsibly”
It is certainly one of a number of causes my colleagues and I based our personal convention in 2024. We have to take the narrative again, and begin displaying a path ahead that highlights how we will use AI responsibly, earlier than it comes again to harm us – be it via laws, or in the eyes of shoppers.
Virtually two years in the past, I took to the stage at the London Developer Convention and delivered a chat stating that gamers are the true litmus check for generative AI’s future in our merchandise, and that our failure to point out worth to shoppers will result in anti-AI sentiment. I argued we might see gamers rally behind “AI-free” games, like one buys natural greens. To this point, we now have failed spectacularly, and whereas some in the viewers thought I used to be a bit mad, the indies have proven me right. That unfavourable sentiment is universal across demographics, and will solely do extra hurt than good in the long term.
Finally, we collectively must be working more durable to tell apart the good from the unhealthy, to coach and talk the place and how we use these applied sciences, and how to take action with out compromising ourselves in the long term. However equally, we have to have a protracted arduous have a look at the scenario we have ourselves into as an industry, if generative AI is being provided as the resolution.