Generative AI: The Definitive Review

[ChatGPT subhead here]

In my secondary school days, way back in the closing years of the 1980s, we were lucky enough to have a design element in our handiwork curriculum. Naturally there was woodworking, some metalwork, making stuff out of Perspex with hacksaws and drills, but the same department had a fully-equipped studio for the sole purpose of teaching technical drawing. Our tutor, one Mr Jagger (not related) was notorious for maintaining a thermos full of whiskey, but more frequently admired for a piece of A2 paper pinned to the wall, showing an astonishingly precise drawing of a very 1970s beach buggy drawn from a rear three-quarter perspective. The lines, clear and confident, met with perfect harmony. There were no overlaps or eraser smudges to be seen and the penwork over the pencilling, consistently opaque as correctly executed in one single pass, looked utterly indelible. This was, we learnt, one of Mr Jagger’s degree exam pieces. He explained that for every union of line or curve on the paper, a single mark would be deducted for the slightest mistake. And there were far more unions to grade than points on offer. Most memorably, we came into his class one morning in the summer term of 1989 to find him sat at one of the student drawing tables. Here he took ten minutes, a set of two French curves, a ruler and his pencil to expertly draw out the Batman logo. We were stunned at the simplicity of the task, yet befuddled by our own, sloppy executions. But then this was the magic of technical drawing technique - the tools and the knowledge go a long way to creating magical drawings for those without the gift of freehand artists. Having been deeply enamoured with Head over Heels, Knightlore et al, I was delighted to find that a single set-square can have you rendering isometric projections with alarming speed once you grasp the methodology. It warms my heart that you can still take GSCEs in technical drawing to this day, but I’m reminded of the fact that rooms of drawing tables are now the sole reserve of education. Once upon a time, they were one of the most vital parts of industry and manufacture. We all know what happened, but we never really lamented the loss. Aside certain Formula One designers, who draws with pencil and pen in the CAD era? Where there were once thousands, tens of thousands, there must now only be a handful. And what was once part of industry is now artisan craft. But yet, to see a truly skilled technical drawing is as captivating as fine art for the right kind of techno-fetishist. We forget that in a society where nearly everything that’s made has drawings to describe it, there was once a time where every single drawing came from a human hand.

What does this have to do with generative AI? Well, it’s not so much what is lost but what causes it to be lost. CAD put paid to entire floors of draughtsmen (and their apprentices) and, naturally, we’re all looking at the same threat to the videogame development industry’s orthodoxy from generative AI (and the far greater evil of growth-based profit motives). It’s fascinating to watch the industry react to generative AI’s arrival and gradually inevitable integration. This year’s GDC played home to some definitive assaults in this direction, most notably Ubisoft’s grand reveal of AI-generated NPCs - though it seems it’s the LLM-derived dialogue of those NPCs that’s causing the bigger upset. This strikes me as slightly weird. Why should dialogue be the line that should not be crossed, as if it’s some definitive moral boundary? I’m prompted to remind people that prior to 2002, there were probably a handful of extremely talented 3D modellers who excelled at making trees, but then the SpeedTree middleware took off and all of a sudden, great-looking, realistic trees were trivial to generate, leaving that small cadre without a market¹. And perhaps that should have been a red flag for anyone looking to pursue a 3D modelling career - if foliage can be procedurally generated with perfectly acceptable results, then it’s likely that at some point, anything organic will be.  And of course today, there are middleware and modeller plugins to procedurally generate buildings, cities, vehicles, planets, animals. It’s part of a tragedy of human obsolescence for videogame development, but raw capitalism also signposts, with grim predictability, that the largest populations of expense will be the ones to be cut whenever a cheaper alternative appears. Perhaps we should all be wondering if we’re the legions of draughtsmen staring into the face of AutoCAD and if not now, when will we be? In some part, it’s the industry’s own fault - a rush into near-decadence to mass-produce vast universes of 3D models and textures initially demands a human labour approach to brute-force the problem. Universities rush to provide degrees and courses to train the workforce to serve a hungry industry and we see this all expressed in the proportions AAA studios show between the coding and asset teams (I’ve seen 2:1 claimed in favour of asset generation). But we also see the fruit of this happening in the games themselves - NPCs being procedurally generated from a set of human-authored components is commonplace wherever believably large populations are required. The seeds of doom are all around us.

From a corporate perspective, there’s nothing particularly precious about narrative that means it also can’t be drastically rationalised and de-staffed with modern AI. Of all the discourse about Ubisoft’s magic NPCs, this transgression of the sacred art of narrative writing seems to have stirred up some serious indignancy. Nathan Grayson, writing for Aftermath, tours the AI NPCs available at 2024’s GDC to report that this first generation of embryonic attempts are, well, embryonic. But perhaps more ideologically, it’s Brendan Keogh’s comments on Twitter² that betray a deeper, protectionist stance. “Why would I want to listen to dialog nobody wrote?” is Brendan’s closing line. For me, this is much the same as the time-worn quote about fitting subjects for videogames, wherein a Redditor asked “who wants to play a funeral?” The answer, of course, is it depends on the funeral, just as it depends on the dialogue more than who wrote it. Refusing AI-generated speech purely on principle surely betrays a certain prejudice³. If we are to take the converse, namely that all dialogue worthy of a listen must be authored by a human, then we’re excluding a ludicrously large possibility space of future videogames. Because while AI might be a bit rubbish now, it’s a fairly safe bet that it won’t stay that way.

Can we imagine a Blade Runner open-world game where we chase replicants through a city of 100,000 people? We’ve already enjoyed over a decade of open-world games that could procedurally generate the models for those 100,000 people, so why not permit the generation of texts to make them all individually interactive? With sufficiently-developed tech, it’s not beyond reason to imagine AI powering a replicant trying to escape the city, with us taking the role of a Deckard stand-in trying to track them down. We could ask each person in the streets if they’ve seen our quarry, or even if they’ve seen anything unusual. We could maybe lie, or ask if they need anything to curry favour with them. If scripted, there’s no way a team of the most talented dialogue authors could write a believable response tree for each of the 100,000 NPCs by hand, but with sufficiently developed AI it’s trivial. If you can believably do one, then you can probably do a million. A fully authored, scripted version of this would be extremely limited in comparison to a dynamic, AI-run equivalent with far more value coming from the increased, unforeseen emergent properties of this setup than anything you could realistically plan if doing it all manually. Now, you can argue that you don’t need 100,000 different dialogues, but I’d say knowing that they’re there if you want them enhances the gameworld and therefore the immersive experience. To discount all that as a matter of pride or principal seems remarkably blinkered to me. What’s more, I can’t say I’ve ever been convinced one jot by the scripted environmental dialogues of NPCs in any videogame so far, because it’s always trivially easy to find the limit of what was possible for the developer to implement, or worse still, NPCs turn out to be the equivalent of shop dummies and simply have no ability to speak because the resources for that simply aren’t available. And this is precisely because when dialogue is written by hand and spoken by humans, it exists as a finite set of audio files in the game media. It is fundamentally limited by its very nature, whereas generated dialogue via speech synthesis is practically infinite. The contrast in capability here is chasmic and the increase in possibilities is fucking huge. In fact, you might even say that AI dialogue, if convincing, would be a generational leap⁴. And it fits with the trend across the board to increase procedural content, be that in terms of foliage, buildings, models etc. After all, do we expect gameworlds to improve over time or are we happy to have them simply plateau at some point, way before they get anywhere near that ideal 1:1 correlation with objective reality? Do we really think gameworld development is some mechanical turk problem where just throwing more modellers, texture artists, narrative writers, audio teams at a project is the only way to increase complexity and march towards that goal? How many more dev staff will it take? Won’t it just get thirstier and thirstier as demand for more complex worlds increases? Or, as has happened across other disciplines, do we develop increasingly sophisticated automated and semi-automated techniques to help make the effort in some way sustainable?

Now I regrettably agree that generative AI is potentially disastrous for employment in the current model the industry uses, but when can you ever trust industry to give a single fuck about the morality of its modes of employment? We can maybe stick a thumb in the dike with some hope for unionisation or governmental legislation, but that won’t stop the onward march of AI’s development. In some ways it is absolutely cancerous to the existing employment model. But if anything, and for a grander sociological view, it perhaps highlights the weakness inherent in the system. This is the problem of an industry whose thirst for people was always going to dwindle, as is the way of all industry, but also of an educational sector that looked to make bucks from eager enthusiasts wanting employable videogame development skills in a time of an employment gold rush. Procedural content generation’s march is well underway - remember SpeedTree is twenty-two years old. Really, the warning shots were being fired way back then. I have the greatest of sympathies for anyone affected by AI encroachment on their hard-earned skills, but I’m wearily cynical that capitalism simply could not give a fuck, something that we should all have been well aware of.

In the face of AI, I am reminded of the wonderful documentary Farewell Etaoin Shrdlu, which details the very last days of hot metal typesetting at the New York Times in the late 1970s. It details the old process as it transitions to the new, digital age and absolutely fascinating as it is, it’s hard to believe the sheer effort, the sheer number of bodies required to make a single page of a daily newspaper. Because every bit of text, every image, is represented by a chunk of metal, it takes crazy amounts of space to just store an issue. Massive tables, page-size frames, all rendered completely redundant by a move to the digital. Interestingly, the staff interviewed are calm and sanguine, fully accepting of the changeover - and these aren’t the reporters or writers of the text. They’re the facilitators that casted the text into objects and arranged them accordingly. Of course, their roles no longer exist and now, the very existence of daily printed newspapers seems increasingly under threat. LLMs have been touted as being the ultimate replacement for human reporters for an industry of local newspapers that’s been in a painfully visible death spiral for decades. But there’s a generational lesson here - these nut-and-bolts processes for making newspapers were virtualised as the information age took hold, just as the dizzying manifold reams of hand-drawn pencil-and-ink schematics by draughtsmen were absorbed into CAD files and plotter-scribed hardcopies, creating a lost generation of careers we no longer recognise, let alone mourn. And the cost was in the number of people required to accomplish the end product. We have no reason to expect information-age products can avoid the same fate. And so perhaps it’s time to really recognise the arrival of generational AI as the most obvious of canaries in the coalmine for our generation of computer-based workers. Sure, you can point and laugh at how shit it is now, but the chances are it’s not going to be shit forever. But at least I can fantasise about capitalism collapsing under AI’s weight, hoping that before it all explodes, I at least get to make a few dream games by simply saying prompts, or have a wild time playing other people’s best efforts. Via YouTube and now TikTok, there was a completely unforeseen social democratisation in viewable content that for newer generations, increasingly edges out the noble arts that older generations see as immutable and sacred. Them youngsters may now watch films or TV shows as background filler for the gaps between content scrolling, but that’s what’s simply evolved. And that old, trad-media content has evolved to take that viewer behaviour into account. There’s no deliberate push for dual-screening, it’s simply what has thrived under a certain set of selective pressures. And the upshot is a devastating reduction of available work in the Film and TV sectors - something I saw beautifully illustrated in a mega-stitch of struggling broadcast workers explaining their lack of work, funnily enough on TikTok. I’d like to see the endgame of generative AI as a kind of social democratisation of anything that can be created by computer. An evolutionary leap under the same selective pressures. And I think you have to see it in those grand, universal terms. It’s all going to be vulnerable eventually. Images, moving images, games, music. None of it is safe, and we need to get used to this. But there’ll always be the folk-art survivalists to keep history alive. I look forward to visiting some old heads in a shed somewhere, booting up vintage PCs for some honest-to-god manual modelling in Maya or 3DS Max, just like it was in the good old days. All the time glancing at my phone for the latest 100-hour open-world RPG releases in this current hour. I suppose in the very worst of generative AI content-tsunami dystopias, there could still be more treasure released in a month than the traditional industry could make in a decade. And, of course, no matter how good the content, how fun the games, being discoverable will still be the biggest hurdle.

[21]

1. Is there a similar argument to be made about motion capture and its effect on 3D animators? I think maybe, yes, perhaps, vaguely.

2. I actually take more issue with his first statement that thinking of videogames as 'worlds' is fallacious when according to him, they are first and foremost human-authored texts. I don't think there's much textural content in Pong, but there is most certainly a world. And that's the first fucking mass-market videogame, after all. Brimming with my own unreasonable prejudice, I’ll simply accuse this mindset as being part of a wider problem that casts the videogame’s cultural worth as being derived from their value as vessels for narratives, hence an urge to see them as nothing more than texts. I suggest the inverse - all authored texts are in fact worlds, though to be more pretentiously precise, I prefer the term simulated continuity, thank you very much.

3. One presumes Brendan is fine with procedural trees and NPC models, so why draw the line at dialogue? What is the special characteristic that confers protection on speech but not the speaker? And surely more people’s careers are under threat of joblessness from asset generation than generated text?

4. Something we can't help but notice has been notably, almost painfully, absent in the most recent round of new hardware.