Game AI Expert Explains Where AI Actually Belongs in Game Development—And The Risks

Artificial intelligence has become one of the most divisive topics in game development, with discussions often centered on automation, cost reduction, and replacing existing workflows. But according to AI researcher Luke Dicken, that conversation misses a lot of context.

Having spent years working on AI research and development at Zynga before helping shape AI governance across Take-Two Interactive, Dicken argues that AI's greatest potential lies not in producing more content faster, but in enabling richer, more adaptive player experiences that weren't previously possible.

In our conversation, Dicken reflects on the evolution of AI in games long before the generative AI boom, discusses where current AI tools genuinely provide value, examines the economic realities of deploying AI at production scale, and explains why technologies inspired by tabletop dungeon masters could ultimately prove more transformative than today's content generation tools.

First, could you give us an intro? What do you do, and what are some highlights of your career and your work with AI?

Luke Dicken: I fell in love with AI for games when I was a teenager, which was…longer ago than I want to admit here. There was this game called Creatures in the mid-90s, and it was awesome. It was like a Tamagotchi turned to 11; you could train these little fuzzy guys and teach them things and play with them, and you could breed them.

What I found out was that the studio that made this was a spinout from the University of Cambridge, and all of the systems behind it were applied research from what was called “Artificial Life” - that was enough to get me hooked. I turned that passion into a bunch of degrees in AI, including a Master’s degree in robotics and a doctorate in AI for efficient long-term behavioural reasoning in games.

I joined Zynga as a Data Scientist in late 2014, which gave me an amazing opportunity to take all these fairly dry academic systems and try to apply them in the real world to have an impact on real players, which was incredibly exciting!

By 2019, I’d led the Central Analytics group through some fairly gnarly technical reimagining, and I then got the opportunity to found a forward-looking R&D team, focused on the intersection of game design and development with classical AI technologies like Procedural Content, Player Profiling, and Mixed Initiative Tooling. This was a few years before GenAI showed up and changed how we talked about everything “AI”.

During that time, we shipped games built around novel tech stacks (some of which we patented) and built tools to support a bunch of internal teams to create and ship content faster using those same classical AI technologies. And then ChatGPT and Stable Diffusion came on the scene, and the conversation around AI changed almost overnight.

Those early days of GenAI were an absolute Wild West of an entire sector trying to go fast no matter the consequences or potential risks, and we rapidly found ourselves tasked not just with R&D for Classical AI but also governance for GenAI throughout Zynga. We spent a lot of time trying to get that right and be more than just “The AI Police” - ensuring that everyone’s voices were heard, that we were as discipline-inclusive as we could be, and that we were really transparent about the downsides as well as the upsides of these technologies. That approach seemed to resonate with folks, and got us attention from the leadership team at the corporate layer in Take-Two, and in short order we were being asked to set up similar systems for the entire enterprise.

As part of that process, the whole of the Zynga AI team got picked up and moved into the mothership, which gave us a greatly expanded remit across all of Take-Two, but the core of our focus and interest stayed on R&D - figuring out how to harness cool technologies to make games more interesting and engaging.

However, after a little more than a year, what had seemed initially like a golden opportunity for us to have a huge impact on the entire company turned into the whole team being laid off rather suddenly. I can’t, shouldn’t, and won’t speculate as to the reasoning behind the decision. That was a few months ago, and I’m pleased that a majority of the folks have been able to find new positions, which in the current climate feels like an incredible testament to the strength of talent we’d cultivated. I’ve been able to take that time to reflect on a lot of things, such as the state of the industry, the state of AI generally, and what I see as the critical areas ripe for innovation.

In the game industry, AI is not really that much of a novelty anymore. We've seen it in animation a lot; we've heard about it in GDC presentations for years, so what happened in the last couple of years that the conversation has changed so significantly?

Luke Dicken: AI has been a thing since the 1950s. There’s a few different ways of defining it, but the one that I favor is “anything that allows a machine to make a decision” and that encompasses stuff like robotics (“I have collided with something so I must back up”) to machine learning and data science (“People who like X and Y are probably going to be interested in Z”) and computer vision (“Pixels shaped and colored like this typically represent a traffic light”) amongst a bunch of other sub-fields. There are a lot of algorithms and techniques that have quietly done some sophisticated things over the years, and it’s ALL Artificial Intelligence. I like to say that Intelligence is an attribute, not a quality, as people make dumb decisions constantly and we wouldn’t say they lack the attribute of intelligence.

The big change of the last couple of years has been that a couple of algorithms have become very prominent, those being Large Language Models (LLMs like ChatGPT and Claude) and Diffusion Models (like Midjourney and Stable Diffusion). Under the hood, they’re both Neural Networks with a bunch of sophisticated sauce to make them work better than a vanilla NN would. Both these technologies erupted onto the scene in late 2022, and overnight the conversation shifted.

If I were being cynical, which I often am, I would say that the real shift has been that ML and those more classical types of AI techniques have tended to be pretty normal, or even almost boring. They’re algorithms people come up with, and then you use them, or you don’t.

GenAI is this wild “Software as a Service” model (in most cases) that’s backed by companies that have marketing budgets literally measured in tens of billions of dollars. Algorithms should be invisible; they should be boring. I have no idea what brand screwdriver my contractor used on a recent renovation project (and maybe they used a socket set instead?) because I don’t care; I just want the job done. But that doesn’t sell more screwdrivers. The hype that’s been anchored to these systems is outlandish, and I do think that there are some interesting capabilities to the algorithms underpinning them, but there’s no way there’s a multi-trillion-dollar industry here. So there needs to be relentless PR and new features and buzz to drive adoption, not because the tool is objectively so amazing, but because this is the first time the algorithm itself has been the entirety of the business.

That’s a real problem coming and going, because it’s making having a nuanced conversation like this one almost impossible (and makes me exceptionally grateful for an opportunity like this). There were over 70 articles written about my team being laid off from Take-Two, and most of them were celebratory, like it was some massive blow in the moral struggle against GenAI, not even bothering to understand the team’s focus or that the vast majority of the staff had been working in the field for a lot longer than the current hype cycle has been going. But equally, I get messages on LinkedIn from folks who “are also ALL IN on AI!” and I’m like…what does that even mean?

It’s exceptionally weird to me that a couple of companies have managed to ignite a culture war and bring the economy to the brink of collapse off the back of a handful of marginally interesting algorithms, but that’s pretty much what’s happened and I think it’s the biggest factor here in why we’ve had this shift from kind of subtle behind the scenes AI to aggressively in your face AI over the past couple of years.

Tell us a bit about your vision of what AI could really be for game developers, where it really enriches the experience by making it more meaningful and less repetitive? I really liked this analogy with the dungeon master reading the room. And I believe in games there were some experiments like this, going as far back as Left4Dead.

Luke Dicken: The Dungeon Master analogy is one I’ve been using for well over a decade now. Back then, we’d gather at GDC to talk about “Game AI” and as a practitioner community, we spent the vast majority of our time talking about making the bad guys shoot back with squad coordination, cover selection, audio barks, and that kind of stuff. The DM analogy really helped bring into focus that when you’re talking about a truly exceptional Dungeon Master such as Matt Mercer, Brennan Lee Mulligan, or Aabria Iyengar, for example, the decision-making inside a combat encounter is such a tiny portion of the experience they’re creating. 

I truly believe that TTRPGs are the purest expression of good, engaging play, and that the human intelligence managing the game is what makes these experiences exquisite. If that’s accurate, then the hallmark of meaningful intelligence, whether it be human or artificial, in games should look similar: a managed, personalized, and crafted experience, catered to the players at the table.

The good news is that when you decompose all the elements of a Dungeon Master, they’re all separately fairly well understood problems. First off, you want to put together players that will mesh well and fit the tone of the table you want to run: that’s social matchmaking with some gameplay-aware heuristics. A hallmark of a good DM is understanding your power fantasy and finding ways to either deliver on it or meaningfully defy your expectations: that’s player profiling and personas and a whole bunch of machine learning techniques. There are so many of these sub-systems within the “Dungeon Master” framing that you can take and look and find an AI algorithm for, whether it's improv or pacing or economy design or loot table creation or name-a-thing. We’ve had the puzzle pieces to replicate each of these for a while, but nobody has ever really sat down and put the entire jigsaw together.

That’s, of course, not to say that there’s no good previous examples to look to; the aforementioned Left4Dead is a great one where an AI Director attempted to match the intensity of the player experience to a pacing curve emulating horror movies. That system could modify which zombies were spawning and where, what supplies were available, and even the weather effects to create that tension.

It was a masterful system, but it also demonstrates my point a little as well, in that the tech that drove that whole system was a “Hierarchical Concurrent State Machine”, an algorithm that had been shown to be highly applicable to this kind of management in a paper published by the University of Iowa as part of a driving instruction system in 1994. Left4Dead was published in 2008. These algorithms exist; they’re just languishing in research papers, underutilized.

There are tons of great examples of AI technologies being used in isolation in games. Versu made fantastic strides in Interactive Narrative, Spelunky and Minecraft showed how far procedural content could be pushed, Dwarf Fortress demonstrated how strong simulationist approaches could lead to incredible emergent gameplay. There’s no way I can talk about all of these, but fortunately I don’t need to, as Dr. Tommy Thompson does a fantastic job of covering all of this and an awful lot more over on his website and the associated conference at AIandGames.com

What is the actual state of generative AI in game production right now? What does it genuinely deliver versus what studios are promising their executives and investors it will deliver? Have you seen any good examples that are "worth" the hype?

Luke Dicken: Well, this is a bit of a loaded question. I’ve definitely seen some good examples; as much as I might come across as an unmitigated hater here, there are things that these algorithms can do that are somewhat useful. But “that are worth the hype?”... that’s a harder one when the hype is this extreme.

The hype machine wants you to believe in magic buttons that automate entire pipelines, and that just doesn’t exist in a meaningful and generalized way. That said, a lot of our work isn’t creative or novel, and a system that spits out a statistically average solution to a problem is fine in some cases. I think that’s why we’re seeing so much traction for the code generation tools, because unit tests are unit tests; they don’t need to be sophisticated or novel, and it’s fairly easy to get that kind of well-understood boilerplate code out of a GenAI system. A lot of documentation of code is a pretty rote description of what the process is doing, and I’ve seen some awesome systems help developers annotate their output far faster than they could have done personally. There’s tons of these little boring mundane things that are part of the job, but in no way the creative dynamo that determines the success or failure of the project; it's just boring and unsexy applications, so there seems like there could be some measurable, albeit somewhat incremental, value in those.

More generally, I like to think of AI as having two primary use cases: when you need something faster than a human could make it, or when you need more of something than a human could make. That holds for the old techniques and the new. If you think of a game like No Man’s Sky, the core of that game was a massive universe that was almost unique on a per-player basis. That would be wholly impractical for human authoring.

It’s kind of the same for GenAI. If you’re just using it to do the same old things a bit faster or cheaper, I’m not sure it really lives up to the hype. But where it gets interesting is when you apply it to scale issues humans simply can't match.

For example, I was talking with a startup recently (I’m not sure how public their progress is, so I won’t name names) that wants to build a nexus for all information related to a development project across JIRA, git commit messages, standup updates, emails, etc. to spot inconsistencies and misalignment and try to ensure that these get flagged early. That kind of analysis of a firehose of data seems like an incredibly good fit for a GenAI system, not least because most of the input is unstructured in a way that would make it hard to analyze with more traditional systems. That’s a really promising use of the technology because you just couldn’t do that effectively without this tech.

When we think about a general rubric for production use today, the right mindset is to think of GenAI as automated scaffolding. It can’t build a house itself; when it tries, weird things tend to happen, and you almost certainly don’t want to be liable for any house it does try to build. But as a way of providing some basic accelerated framing for teams, there are some really valid, at least from a technical point of view, ways these algorithms can be deployed.

Is the cost structure of deploying AI at production scale something the industry has genuinely reckoned with, or is it still being treated as a future problem?

Luke Dicken: Hah. No. Not even close. We’re in this crazy era of Big Tech trying to "make fetch happen" so hard that they’re burning giant piles of money subsidizing these systems. If you look at the investigative reporting being done by folks like Ed Zitron, the math just isn’t mathing. OpenAI is reportedly burning over a billion dollars a week in net losses, and that seems to be after some very “creative” accounting practices have massaged that figure.

Even despite this subsidization, we’re seeing wild headlines like Uber burning through their entire 2026 GenAI budget in four months or ServiceNow’s CEO saying that people are losing track of GenAI costs. This entire thing is an incredible expense today, and that’s at the subsidized rate. What happens when the major model providers have to move their businesses into actually acting like a business and generating a profit? What happens when the sweetheart deals they’ve negotiated with the hyperscalers (who are also significant investors in the providers) expire, and those folks also need to get paid?

We’ve already seen this happen when GitHub Copilot moved from a subscription model to a usage-based model, and users saw what their actual costs would be. In some extreme cases, we’re talking about a 1,000x jump in price, but 10-20x seems to be incredibly common.

This is a five-alarm fire that ought to be keeping every business leader up at night. They’ve allowed themselves to be lulled in by these low prices, and they’ve not been particularly diligent about quantifying the ROI of their implementations, which means they have no idea whether a new price still represents good value or not for their business. That’s terrifying, not least because most rigorous studies are showing a huge discrepancy between self-reported value of AI deployment and actual measurable value.

That startup I mentioned earlier that’s monitoring for misalignment sounds cool on paper, but how much is that capability worth to you? What value is it delivering and how much would you actually pay for that? If you can’t answer that, how can you determine whether it’s something you should be embedding into your production pipeline?

A regular counter-argument here is that over time, these technologies are becoming cheaper to operate, with a specific emphasis on “per-token cost of inference”. The costs associated with these technologies are deliberately obfuscated, but for the sake of argument, let’s say that that is true. The problem is that this overlooks the overall trend, which is that every new model typically consumes more tokens per task. They now run loops; agents prompt other agents. It’s like arguing that vehicles are getting better miles per gallon while simultaneously dragging everywhere further apart so you have to drive further. Micro-level efficiency isn’t the point when macro-level costs are growing so rapidly.

So many people advocating for GenAI right now are kicking the can on the cost concerns, whether intentionally or not, and that’s going to bite them hard, probably within the next 18 months. That’s not necessarily an argument for not going near these systems, but I think it’s important to do so with eyes very wide open and paying attention to the value so you know what costs you can bear, and making sure your adoption is modular enough that you don’t get held hostage by a price hike impacting a critical pipeline you can’t rework.

There is a concern that training AI on AI-generated content will gradually degrade the quality of models over time as original human creative work becomes harder to find. How seriously should the games industry be taking that risk, and is anyone actually planning around it?

Luke Dicken: Training is something that I didn’t really touch on when talking about costs, but it’s a really expensive process. It takes a huge amount of data and a massive amount of number crunching to train these models; it’s wildly expensive. One of the biggest misunderstandings about the nature of this process is that it’s hard to make it better. It’s not a case where a bit more data gets you a bit better result. Instead, it’s very much a situation of diminishing returns where you need a massive amount of new data to get a bit better result, and a massive amount of extra processing. It’s part of why these companies need such large datacenters and such frequent cash injections, and that’s not something that will be easily overcome; it’s been a fairly accepted feature of how Neural Networks function for decades.

Now another major problem comes when you start from a place of having trained your LLM on the entire collected works of humanity. Setting aside the legality and ethics of having done that, let’s just think about the technical side. You’ve already consumed everything, but you need more. So what do you do? Well, if you have a machine that can make "new" content, you start spitting out stuff to feed back into the machine and train on what’s called “Synthetic Data”.

If you think about it in terms of photocopying a photocopy, you can immediately see why this might be bad. The errors start to accumulate, and over time the entire thing becomes unreadable. For LLMs, that process starts to do two things: it reinforces and exacerbates the crazy amounts of bias prevalent in the output of these kinds of systems, and it starts to codify as fact “hallucinated” outputs.

That’s a pretty big problem. These systems have picked up an awful lot of historical material that reflects significantly less inclusive eras of society, and the efforts to straighten that out have been minimal and hamfisted at best. In fairness, it’s a very hard problem, and why responsible folks might have taken a second to reflect before unleashing these systems on society.

But importantly, this isn’t really a games industry problem or something that we can address except to voice our concerns as current or potential customers. Candidly, I don’t think the current recklessness among GenAI companies makes it likely this will be addressed, and it seems likely that we’re just going to see these technologies start to exhibit systemic degradation, and when that happens, hopefully we’re not so deep down the descent into Idiocracy that we can’t find our way back out.

Studios are making long bets on where generative AI will be in three to five years, not where it is today. What would it actually need to become before the gap between the promise and the reality closes enough to justify those bets?

Luke Dicken: Three to five years feels like a long time given the pace that GenAI as a field is moving right now, but it really isn’t because the majority of the stumbling blocks it faces are external to algorithmic improvements and the kinds of iteration that these companies can make. There are broadly three buckets that I group the issues into: Legal, Moral/Ethical, and Business.

Legal is the easiest to deal with; there’s obviously been an awful lot of concern about the fact that the training data for these models allegedly flagrantly violates copyright and tramples on the rights of artists and authors. Unfortunately, my best guess today is that on a five-year horizon we’ll see a lot of court rulings siding with corporate interests on this one. That’s going to have some interesting repercussions since right now I know some large companies don’t realize that their IP is deeply at risk of not being an effective moat or something they can even protect on an ongoing basis; in a world where anyone can train on anyone else’s source material, some of the AI-based offerings that large companies with first party IP are putting together will have to compete with third parties training on the same material.

But there’s an overlooked side to the legal concerns too, which is that it’s not currently possible to get copyright protections or even necessarily assert ownership over the output of these systems. For internal documents and code, that might not be such a big deal, but it’s certainly a massive concern for GenAI images and videos. Imagine creating a game with a character that people resonate with and finding out that you don’t own that character and other games can just use it… I don’t think that one gets resolved on a three-to five-year horizon, but it feels like a big gap that needs to get bridged.

Moral/Ethical concerns are some of the most commonly discussed. Whilst we’ve got some hints as to what the legal position on training data might end up being, I think there are some bigger questions to be asked as to what it ought to be, but that’s something everyone’s aware of and trying to work through. There’s a lot of other issues here as well, but the one that I truly worry about is what we’re going to allow these technologies to do to our human talent pipelines.

If leadership looks at GenAI purely as a cost-cutting tool, the obvious place to start those cuts is junior talent, eliminating a lot of entry-level tasks that we use to train our staff on. People don’t just wake up one day as Technical Directors or incredible concept artists; they learn their craft by doing. If we burn down the bottom rungs of the career ladder for a short-term margin bump at the next earnings call, we’re all but guaranteeing there'll be a talent drought on that five-year timeline. That’s got some incredibly widespread societal ramifications that I don’t even want to try to come to terms with. It’s bad.

Business concerns are in some ways also technical concerns, because that’s where I put a variety of aspects of these algorithms that make them a nightmare to integrate into a serious production environment. Obviously, we’ve talked about the unpredictable costs and the difficulty of measuring value, but an example of another pretty significant issue is the way that they produce answers. These systems are inherently non-deterministic, meaning that for a given input you don’t necessarily get the same output. Worse still, the nature of that non-determinism is not immediately obvious to a user.

“What is two plus two” and “Can you tell me the result of adding two and two” are, to an LLM at least, very different questions (incidentally, this is also why weird things like adding “And be careful to get the answer right!” to your prompt can sometimes get you different answers. It’s not following an instruction; you’re changing the nature of the input you’re sending). In the context of evaluating the efficacy of a tool in a pipeline, you need the result to be stable.

This situation gets even worse when you consider that even minor changes to a model’s weights (which are often tweaked by the providers) can have a significant impact on the quality of the output. In AI circles, this is referred to as “Agentic Drift”; in practical terms, you’re giving an interview to the “person” that OpenAI or Anthropic have sent today, and hoping that they send the same person to do the job tomorrow after the PoC process is completed. There are a few solutions proposed to Agentic Drift, but most of them involve another AI system burning even more tokens to check that the first system worked right. These aren’t really things that can be fixed; they’re inherent to the nature of how these algorithms work. Maybe on a three- to five-year horizon someone could come up with something, but right now it feels like a lot of these long bets are being used to build the foundations of studios on constantly shifting sand.

Game development costs have ballooned to a point where the traditional AAA model is under serious strain. AI is being positioned as part of the answer to that cost problem. Is that a realistic solution, a convenient narrative, or something in between?

Luke Dicken: Well, I think we should look at why development costs have ballooned in order to know what the solutions might be. Sure, certain types of games have gotten bigger over time and player expectations in some genres have gotten higher, but how many projects do you hear about that have gone through multiple resets and pivots, where leadership has wandered off at a tangent, for years in some cases! It doesn’t matter if you’re moving at 50 miles an hour or 100 miles an hour if leadership is taking the project in the wrong direction. How much of that “ballooning strained cost” was actually spent building the thing that shipped? It’s really easy to start pointing at artists and engineers and saying “AI can do their jobs faster and that will keep costs down”, but is that where the bottleneck actually is?

I’d argue that in a lot of cases it isn’t; it’s that teams are building stuff that frequently gets thrown away and their time and value is being wasted. GenAI can make those inefficiencies less impactful, I guess, but the bottleneck isn’t the people doing the work; it’s the people telling them what work to do. No amount of automated asset generation is going to fix a broken steering wheel.

Right now, GenAI presents a convenient narrative, but as we’ve talked about already, the costs of that are also getting out of control. As the technology continues to not really deliver on the promises it was sold on, we’re already seeing companies resort to rehiring humans that they’d dismissed in favor of AI, and there are reports of various companies starting to set policies about what tasks should and shouldn’t be sent to AI systems. It really seems like rather than fixing ballooning costs, GenAI might actually be contributing to making it worse in some places, which is a truly astounding achievement.

And remember that this is what’s happening while these tools are still being subsidized! It’s really hard to paint a picture of how this is going to actually reduce costs in a meaningful way in the long run.

If the AI investment currently flowing into generative tools were redirected toward AI that genuinely understood and responded to individual player behaviour, what could games become that they currently are not? I think NVIDIA actually did some Deus Ex-style experiments in that area. And there are, for example, mods that let you talk with characters in Skyrim in real-time and stuff like that.

Luke Dicken: So let’s take this in reverse order. The NVIDIA tech demos are really interesting, but they’re tech demos, not games. Sid Meier has a really useful definition of a game that works well to highlight this point: “A game is a series of interesting choices”. What’s great about this is that it really highlights the wrongheadedness of these tech demos in that if you wire an NPC’s dialogue tree up to an LLM, it gives you more choices, but that doesn’t make them interesting. Riccitiello, in one of his final interviews as CEO at Unity, talked about the potential for AI in games and imagined this future where you could play a game like FIFA and walk around the stadium, talk to the fans, and have them come play with you. What the f**k is the point of that? From a pipeline and design perspective, that is an enormous waste of R&D resources on features that don't align with player intent. Good R&D isn't about adding endless noise; it's about amplifying the core player experience.

Games, or any media really, are as much about what’s not included as they are what is. Crowbarring more in haphazardly has never been the way that games are successful, because more isn’t always better. If you want mundane “more”, the real world is right there. So yes, NVIDIA has a fancy tech demo set in a Cyber-renaissance style that evokes Deus Ex, but what about it is meaningfully better for our players than Deus Ex? What’s the “couldn’t be done without GenAI” aspect that’s core to the gameplay experience?

I still have to negotiate a dialogue tree with the hotel receptionist to sneak upstairs, but now that tree has infinitely more branches for me to try to get through. It’s a cool tech demo, but as a core mechanic of a Deus Ex-style game, I think that gets old rapidly and is frustrating as hell soon thereafter.

Ultimately, the question very few people are asking is: “What does this, from a design standpoint, allow us to do that’s beneficial to the player experience and that we couldn’t already do before?” That’s where the value actually lies.

Shipping faster, productivity gains, all that stuff is going to net out in the wash, and as I’ve mentioned, we’re not really going to be able to get a good sense of the true ROI here until the real cost starts to hit. Ten years from now, our toolset will probably look different than it has historically, and collectively we’ll have worked out what works and what doesn’t. There’s some incremental value to be captured in “figuring it out first,” but it’s frankly minimal. But figuring out what new player experiences we can unlock, what else games can be by harnessing these technologies? That could lead to new genres, new markets, new modalities. That could be transformative for a business!

To get there though, we have to fundamentally rethink how we architect our development infrastructure. Good doesn’t look like burning capital on tools that aim to automate the status quo but with fewer employees, nor is it content for the sake of content. I think that more than anything, the real thing that’s needed here is investing in R&D pipelines designed to empower teams to experiment creatively in as cost-effective a way as possible. Technology has to be in service of art to be effective, and I think that’s something that’s been overlooked throughout this GenAI hype cycle.

The studios that realize that the real goal shouldn’t be to squeeze the human workforce but instead to harness any technology to amplify that workforce’s creative reach—those are going to be the ones that define the next decade of play.

What is the thing that gives you the most optimism, and what is the thing that concerns you most about the direction the industry has chosen?

Luke Dicken: We talk about the industry like it’s a monolith, but it’s not. A lot of the GenAI enthusiasm is coming from the corporatist side of the industry (I’m being specific here to not use a term like AAA since this cuts across the kinds of products being generated and is much more about corporate structure). Big Business as it operates today needs better margins; they need huge returns, they need to be in service to the hypergrowth dogma that’s captured all of capitalism —numbers must go up!

GenAI lets them tell a good story today, though it’s deeply questionable how sustainable or accurate that narrative ends up being. It sucks, and it’s self-defeating, self-destructive, and ultimately I think this earnings-call-driven short-termism will be the undoing of a lot of things. That’s concerning.

Where the optimism lies for me is that this industry wasn’t built by corporate interests. It was Weird ‘Lil Guys making stuff that captured people’s imaginations first, and there’s still a lot of us around! I think a lot of people have forgotten that before Minecraft was part of the Microsoft juggernaut, it was just a couple of folks in Sweden with a neat idea they wanted to explore, and it’s just one example of a real, common trend. Corporations rarely make “the first…” of anything; they buy it up pretty often, or they clone it, and they’d very much like you to forget that they didn’t actually make it themselves in the first place. But true innovation can’t be stopped, and it doesn’t rely on them. That’s pretty exciting.

Luke Dicken, Data Scientist and Game AI Expert

Dr. Luke Dicken has shaped the architecture of modern gaming over more than 15 years, spending a decade driving studio innovation at Zynga and most recently serving as the Head of AI for Take-Two Interactive. Alongside his corporate track record, he spent seven years supporting the next generation of industry talent as Chair of the IGDA Foundation.

This interview was conducted by Kirill Tokarev, with copy-editing and publishing by David Jagneaux.