What I Got Wrong About AI
Eight years ago, I stood on a stage at the Promised Land Art Festival in Poland and gave a talk about the future of game characters. I tried to predict where the industry was going: the role of AI, the pipeline improvements we'd see, and what artists would be doing five to ten years out.
Looking back at that talk now, I got some things right, and I got some things quite wrong. Some of the misses are things nobody could have predicted in 2018. But some of them I probably should have seen coming, and didn't. And the ones I didn't see are the ones really affecting artists and games today. So let's talk about that.
The Talk
The talk was at the 2018 Promised Land Art Festival, organized by CD PROJEKT RED in Łódź, Poland. It's a great international event, hundreds of artists and professionals from across games, film, and visual arts, with plenty of speakers from studios you'd recognize. They invited me to talk about the future of game character production. I went with my then-colleague Matthias Develtere, a fantastic hard surface modeler and a great travel buddy.
The structure was simple: how we made characters a decade ago, how we made them in 2018, and a section on where I thought we were going.
In that future section I had a slide on AI. I talked about generative adversarial networks, machine learning, and I mentioned Adobe's Scribbler demo, which was state of the art at the time. The framing was that AI would be assistive. It would help us work faster, generate options, fill in the details, but leave the artist in charge while the tool did more of the busy work.
That was my prediction. Let me start with what I got right.
What I Got Right
I built the talk around three pillars: automation, simulation, and reality capture. Looking back, all three held up.
- We did get more automation. UV tools, retopology, baking, all of it got better.
- We got more simulation, with cloth and even strand-based hair simulated routinely in production now.
- And we got more reality capture, where photogrammetry is a baseline and performance capture pretty much everywhere.
The pipeline I described as "tomorrow's pipeline" is recognizably today's pipeline: concept, proxy, high resolution, game model, UVs, baking, texturing. A character artist today is doing essentially the same chain of operations I was doing in 2018, just with better tools at each step.
I mentioned VR sculpting, which existed in early forms back then and has matured into real production tools now, though it hasn't replaced the desktop apps the way some people thought it might. I mentioned AI assistance as a coming thing, which it was. And I closed the talk with the line that intent matters, which has held up as well, maybe more than I knew at the time.
So on the technology level, I called it. The trajectory was real, and the pieces I named were the pieces that showed up. If you'd watched that talk back then and made some career bets based on it, you wouldn't have been wrong about where to invest your skills.
So far, so good. Now for the trickier part.
What I Got Wrong
I got two things wrong, and they're connected:
The first is my assumption about what AI tools would be used for. I was imagining tools that would take the tedious parts off my job:
The technical drudgery. The parts that are necessary but not creative. That's what I, as an artist, wanted automated. Nobody gets into character art because they love unwrapping UVs or building LODs.
That is not what the current generation of AI tools went after. The big generative models, the image generators, the 3D generators, went after the creative part: the concepting, the designing, the sculpting. They went after the things artists actually want to do, and left most of the technical drudgery exactly where it was.
And here's the part that gets lost in a lot of the conversation about this: the output is mostly not usable in production.
A generative image of a character may look impressive at a glance, but look closer and the details tend to fall apart. The classic hands that don't make sense. Armor that doesn't connect to anything. A belt that turns into a strap that turns into nothing. It's a mashup of thousands of scraped images, averaged into something that reads as a character from across the room but collapses the moment you actually need to build from it.
The 3D output is even worse for production. Generated meshes come with topology that no animator could rig, UVs that are unusable, geometry that's fine as a silhouette maybe, but useless as an asset. In the current state of things, what you get out of 3D generation tools is perhaps a decent block-out. A placeholder. Which, going back to my 2018 talk, is a real step in the pipeline, but one of the faster and least expensive ones.
It's not where the time goes, and it's not where the cost is. I'm aware that many models are pushing for more game-ready topology etc, but without knowing the constraints of an engine, the intended use of an asset and whatever pipeline considerations are needed an AI model will only ever be able to generate very average models that most likely will not be performant in production (or optimized for the game/use case).
So generally the tools were built to automated the part I didn't want automated, produced output that mostly can't be used in production, and did it by speeding up steps that was never really a bottleneck.
Why It Went This Way
So why did it go this way? Why build a tool that replaces the artist instead of a tool that helps the artist?
I think the honest answer is that it sold better.
Automating the creative process, "you won't need artists, or you'll need far fewer," is a great story to tell an investor. Replacing salaries and shortening expensive production timelines looks fantastic on a balance sheet. "We made a tool that helps your existing artists unwrap UVs faster" does not raise a venture round. While "We made a tool that replaces your art department" certainly does.
So the money went toward the more dramatic version. The version that promised to remove the humans rather than help them. And we ended up with a generation of tools optimized for a pitch deck rather than for a production pipeline.
The scraping question sits underneath all of this. These models are good at producing derivative work because they were trained on an enormous amount of work scraped from the internet without permission or payment. Work by me, work by everyone I know. Nobody asked, and that's a real problem that deserves its own conversation. But even setting the ethics aside for a second, the tools that resulted are aimed at the wrong target.
They were built to replace the work, not support it.
What I'd Predict Now
So here's where I've landed eight years on, and what I'd predict today.
There's enormous room for AI tools to do the thing I actually wanted in 2018: tools that take the technical work off the artist's plate and leave the creative decisions with the human.
I've been testing some of these. There's a tool called Tractive that does AI retopology. I've been using it since their beta, the first version released to outside artists, and it's recently opened up to early access. It's non-generative, which is the important part. It doesn't touch your sculpt or your style. You still drive the edge flow and the intent, and it handles the tedious rebuild of the mesh underneath, like ZRemesher guides but with more control.
Just the other day we saw another contender in the same arena, with Cosy Blanket Pro being announced, also for automating retopology.
There's better automated and semi-automated UV mapping than we've ever had, with tools like RizomUV and Unwrella (which, full disclosure, I still haven't checked out myself). And then there are Houdini-based pipelines that automate huge amounts of the routine work between the high-poly sculpt stage and the engine-ready asset.
To give another personal example: I've used Claude Code to automate some scripts and workflows. I'm an okay Python coder, but writing code isn't really what I find engaging or where I want to spend my effort. So I let AI handle the parts I'd normally grind through, to get me to a working tool a bit faster.
But here's an important caveat. The AI helps me, a competent-enough coder, move faster on things I already understand. It does not replace an actual technical artist or tools programmer.
A good TA or tools programmer wins out over an AI tool any day, because they bring judgment and context the model doesn't have. They know what to build, why, and how it fits the wider pipeline. It's not a substitute for the skill. Same pattern as everything else here: it's useful when it supports a human who knows what they're doing, and it becomes very hollow when it tries to replace them.
There was a talk at GDC earlier this year from Embark Studios about exactly this: their character pipeline for Arc Raiders and The Finals, built on Houdini and USD, designed around a procedural mentality that spans the full range from manual to automated. The whole point of it is to maximize artistic freedom by minimizing repetitive work.
That's the sentence. That's the thing I was trying to describe in 2018: Minimize the repetitive work, maximize the artistic freedom.
That's the future I actually believe in. Not a tool that generates a character from a prompt and hands you an unusable mess, but a pipeline that takes your sculpt, the thing you made with your hands and your eye and your judgment, and removes the friction between that and the finished game asset.
The difference matters. One of these is aimed at replacing the artist. The other frees the artist to spend more time being an artist. The technology underneath is similar. The intent behind it is the opposite.
As game artists, we've always lived in a very technical space. We have to understand the pipeline, and we have to work within the constraints of the platforms we're developing for. If we can get from our DCC applications to the game engine in a more problem-free way, that's something to strive for.
The Honest Accounting
So that's my honest accounting, eight years later.
I got the technology trajectory mostly right. Automation, simulation, capture, AI: they all showed up. I got the direction wrong in one specific way: I assumed the tools would be aimed at helping artists, and instead they were aimed at replacing artists. And I underestimated how much of that choice would be driven by what sells to investors rather than by what actually works in production.
But here's the thing I keep coming back to. Over a long enough timeline, production reality wins over the pitch deck. The studios that lean on generative output to skip the creative work are going to ship worse, more derivative art, and it's going to show.
The fundamentals are the bet that always pays off. If you know how to look, how to design, how to build a character that actually works, you can pick up any of these tools and use the good ones without being replaced by the bad ones. The artist who understands the whole pipeline is the artist who decides which parts are worth automating.
That's what I'd tell my 2018 self, The future isn't a tool. It never was. The future is the artist who knows what tools to use, and when.
Good luck out there!
I'm a character artist with 20+ years in AAA games (MachineGames, Arrowhead, Avalanche, and others). I make videos about the craft and the workflow over on YouTube. If you're a student or junior artist, I also run a small Discord for feedback and craft talk: https://discord.gg/8gjGzc8rE9