"Hopefully a meteor will strike the earth soon": Game devs talk genAI witchhunts and the importance of honing your visual literacy

The era of generative AI has kindled an era of generative AI witchhunts. It's routine right now for videogame developers to fend off accusations of using the technology, even when they are full-throated in their disapproval of ChatGPT, Claude and their ilk. Recent examples include Sandustry, the 2D factory builder from Hooded Horse with strangely haunting physics, and the just-announced Humankind 2, whose creators Amplitude have rushed to assure players that their announcement trailer is genuine live action film.

The conversation reflects a shifting idea of what AI-generated imagery 'looks like' - shifting because, strictly speaking, AI-generated imagery has no fixed style or quality. It's derived from whatever data the model is 'fed'. The confusion is exacerbated by the fact that the technology is changing, with new or updated buzzwordy tools being added to the pile. Established hallmarks of generated imagery go out of date fast. Meanwhile, generative AI's advocates muddy the waters, conflating tools and framing their capabilities in ludicrously open-ended, investor-baiting terms. Still, some responsibility falls on regular users who've become accustomed to chucking around claims of "AI slop".

Over the past few weeks, I've been chatting to game developers and artists who've been falsely accused of using genAI for visual art in particular, while also trying to understand why, exactly, their images and videos attracted suspicion. It's been a funny journey, both depressing and at times uplifting. The difficulty of identifying generated imagery at a glance is a political and social priority that goes far beyond videogames – genAI fakes are routinely used to manipulate people and foment conspiracies, including racist and xenophobic narratives about the "great replacement". We need to become better at spotting the snakeoil salesmen in our midst. But more positively, learning how to spot genAI is also the act of improving your visual and technical literacy. It's a process of becoming more intimate with art that captivates us.

Image credit: Failbetter Games

Hannah Flynn, comms director for Fallen London and Sunless Skies developers Failbetter, is used to getting flak online. "Most of my job is concerned with putting our games in front of new people," she tells me over email. "Success in my efforts means attention on us, which has always been a double-edged sword (I'm grateful that it's been a while since someone asked me for feet pics, though I fear it's never far from happening again). It’s a new development, though, that if a video gets shared widely or we have a big announcement, we're guaranteed to get asked whether or not we're using genAI."

Failbetter have a statement on the website for their flagship literary text-me-do Fallen London, swearing off any and all generative AI usage in their games. "Some of [the] comments represent genuine concern from our community, who've maybe read our AI statement and are worried that we've reneged on it, or violated it accidentally," Flynn goes on. "I don't mind reassuring our community personally; reservations about the quality of genAI’s outputs aside, we take a lot of pleasure in our craft and no one here is interested in switching career to AI wrangler." More bothersome, she says, are the "social media fly-bys from people who don’t seem invested in our games or the studio".

New Blood Interactive CEO David Oshry has also been fending off generative AI accusations. With regard to the forthcoming Tenebris Somnia, a promising horror game that combines top-down pixelart with live action film, New Blood eventually created a behind-the-scenes video showing off the creation of some props to shut down claims that they had been coughed up by a chatbot. I emailed Oshry to ask whether this was a pre-emptive gesture, or a response to some specific allegations. "LOOK AT THE COMMENTS AND YOU TELL ME," he replied, sharing links to some previous videos.

Image credit: Saibot Studios

According to Oshry, allegations about generative AI have become "pretty much non stop" when sharing footage of New Blood games on social media, and especially Instagram - galling, given that he hasn't been shy about his disdain for the likes of Nvidia's face-plumping DLSS 5. "And we've noticed it's only very recently this has happened," Oshry continued. "Even a year ago we saw very little of this and mostly people just saying 'wow that looks so cool!'" Does he have any idea why Instagram users in particular seem more prone to this behaviour? "I think short form video platforms like Instagram and Tiktok have genuinely fried people's brains."

Still, Oshry is sanguine about the suspicion, noting that the feedback has actually rebounded to New Blood's benefit. "All the engagement has driven up our wishlists (15k added in the last week), but it is a bit exhausting," he tells me. As with Failbetter's projects, a significant proportion of the feedback is from drive-by posters, and New Blood's more committed followers have come to their defence. "As you can see in the comments people are quick to defend us and the game and set the record straight," Oshry goes on. "But we would rather be making cool content ABOUT THE GAME rather than making content proving it's NOT AI SLOP. It is what it is."

Image credit: Abandoned Sheep

Even when accusations are unfounded, some studios may prefer to change their art rather than deal with simmering mistrust. Abandoned Sheep, for example, are the developers of Schrodinger's Cat Burglar, a game about a self-duplicating cat navigating a maze of Portal-style testing chambers. One of the game's older pieces of key art, shown on the right, has been repeatedly called out as AI-generated, as studio director and former Jagex and Sony Cambridge developer Martin Binfield tells me.

"This piece is hand painted by local artist Tom Williams (as is the current key art) and was produced in mid 2022 - predating GenAI's ability to produce work to this standard," he says. Nonetheless, Abandoned Sheep ultimately found it more practical to replace the art than push back on the suspicion, not least because after four years, the art was in need of a refresh. "Honestly I don't begrudge those commenters for their vigilance," Binfield comments. "We as a studio have never and will never use generative AI, and I as a gamer also hate the idea that it might feature in a game I want to play. The 'AI art? No sale' commenters have their hearts in the right place."

All of the above games and artworks look very different, and it's tempting to throw your hands up and say there's no rhyme or reason to some of the accusations. "I honestly have no idea," says David Oshry of Tenebris Somnia, with characteristic forcefulness. "I think people may have honestly forgotten or are too young to know about practical monster makeup and VFX and think everything is either Marvel-style CGI or AI now??? It's baffling. Because as someone who sees a ton of AI slop on the internet every day, Tenebris Somnia definitely doesn't have "that" look. Or at least we thought so!" Asked whether he has any thoughts on the kind of game or aesthetic that gets misidentified as AI, he adds: "I am very old school and not tuned in to this bullshit and try not to be! I've seen people say that GTA 6 looks like soulless AI slop today so I think we've all just clearly lost the plot and hopefully a meteor will strike the earth soon."

Hannah Flynn is perhaps a smidge more optimistic than Oshry. She muses that Failbetter's "painterly art style" may read as generated, as may its penchant for deliciously 'aberrant' bodies that recall the disembodied limbs and knotted proportions of a slightly older variety of genAI photorealism. "It's often observed that AI images are trained on classical styles, which turns around on you when you're using a classical style to create by hand," she says. "And in fairness, we’re known for eldritch horror – some of our artwork genuinely does contain too many eyes or fingers, but I promise it’s on purpose!"

Image credit: Failbetter Games

Flynn adds that the conversation about generative AI is, in some ways, inspiring. "The meaning of made things is a sort of sub-theme" of Failbetter's current project Mandrake, a folkloric, Stardew-esque gardening sim invested with the studio's usual blend of whimsical and macabre. "It feels on point to talk about," she says, "and maybe leads to more people spreading the word!" Without having played Mandrake, I can't quite get into the weeds of this, but it certainly makes me curious to have a go.

If visual errors create grounds for suspicion, so too may immaculate execution. "From the comments, it seems that it was 'unnaturally polished'," says Martin Binfield of the key art for Schrodinger's Cat Burglar. "I presume things like the airbrushing and the reflections in the cat's eye were the aspects that set people off, as there's plenty of rough brush-stroking in the piece at large. This feels to me like a 'cursed by success' thing - Tom's art was super polished and high quality, and one presumes that gen AI has ripped off the highest quality art work for its algorithmic slop bucket. To be accused of being AI in some cases is to be accused of being 'too good to be true'.

"Of course, other times AI accusations are because of an incorrect finger count, but still," Binfield adds. "One comment talked about the cat having missing teeth - not sure which teeth in particular they thought were missing, and the cat is hardly anatomically correct anyway. I think the main trend in the responses I noticed was an absolute sense of being correct. It was rarely a question ('is this AI?') and more often a declaration ('no sale until you get rid of the AI'). I put this down to general internet-comment arrogance - indeed, in the instances where I replied and stressed that we weren't using AI they would generally comment 'that's great to hear! Sorry for getting it wrong!'. Next game, I'm going to draw the key art in crayon using my left hand so it was clearly human-made…"

Image credit: Abandoned Sheep

While talking to the developers about how their work has been misinterpreted, I've been thinking about my own, patchy visual literacy and sensitivity to indications of AI generation – I was embarrassingly caught out in a recent article on ASCII horror game The Carcass. I've been looking for resources to hone my savviness. Among the best I've found is this academic survey created by table-top game and book illustrator Janna Sophia Koppenwallner, which asks the user to sort the slop from a collection of book covers, archive photography, snapshots from news broadcasts, and more abstract 'fine art'. In the process, it feeds you tips: errors to watch out for, such as skin that has a certain Lovecraftian lustre, and contextual giveaways, such as a scarcity of biographical detail about the creator.

Koppenwallner is, in fact, writing a Masters thesis about human AI detection image accuracy. "Before I chose the topic I did a little bit of research on human AI detection and found multiple studies that claimed that it was no better than chance," she tells me over email (here's one example). "I think a part of me didn't want to believe that was true."

I should stress that Koppenwallner's data analysis is ongoing and not yet peer reviewed, with more research needed. Please scatter your salt shaker liberally over what follows. Still, her provisional findings are intriguing, and a bit discouraging. "The data I collected actually shows an average detection accuracy of 74%, and found that people who finished a short training module actually did just a bit worse than those who did not," she says, adding that the actual difference was "pretty marginal", however statistically significant. "From what I've gathered so far, people did get overly suspicious and over-corrected, misidentifying real photos and art as AI generated," Koppenwallner observes. This describes my experience of her survey: I chalked up more false positives than misses.

Image credit: New Blood Interactive

Koppenwallner herself has never been accused of using generative AI, partly because she's been posting her art online for 15 years, so there's ample evidence that she's cultivated her craft. She thinks artists who are just starting out today "might have a much harder time". I'm curious to know whether amateur art with obvious blemishes might be more at risk of being misidentified as AI-generated, and whether there are any artistic traditions or styles that are somehow innately more suspicious.

"There's no data in my survey about that," Koppenwallner began, "but anecdotally I've noticed that those artists being falsely accused of using gen AI had rather popular art styles, that look very clean, highly rendered and polished. Gen AI relies on immense datasets, and the more data a model has, the 'better' the output. Most models still fail to re-produce niche art styles, since there simply isn't enough data for them to draw on. But that might change in the future, and again, depends largely on the AI model." Koppenwallner adds elsewhere in our chat that every generative AI model has "a distinct visual language, which only becomes apparent when you compare a lot of their output". Still, she thinks that generated imagery trends more often arise from certain varieties of art being more popular.

As regards the risk of artists still mastering their techniques being branded genAI users, Koppenwallner offers a fascinating dissection of 'hand-crafted' versus generated error. "When it comes to mistakes in art, the failure patters AI models produce are very different from mistakes a human artist would make," she says. "AI tends to blend things together (like hair suddenly looping into a necklace), because the model doesn't actually know what it creates, it just uses a very good probability algorithm. Again - this wasn't part of my thesis and is purely anecdotal - but human beginners tend to struggle with proportions, colour and lighting (which usually the AI gets somewhat right), while most AI models struggle with anything that requires precise decision making, like lineart, details, textures and ornaments."

In her questionnaire, Koppenwallner encourages people to follow their own intuition – that nagging, inarticulate feeling that something about an image is Wrong. But she also echoes Martin Binfield in remarking that the feeling that something is 'too good to be true' can easily become sheer bias. Intuition has to be deliberately modulated through study. You have to cultivate your own taste. "There's actually a really good book by Robin Hogarth about training intuition," she says. "According to him, intuition is trained through feedback, but if the feedback is wrong our intuition misfires. That's how prejudices are created in the first place. Unfortunately, training correct intuition would require proper training and time to develop a certain level of expertise."

Image credit: Abandoned Sheep

Perhaps inevitably, Koppenwallner has found that fellow creative industry workers do better at her quiz than people in professions that don't involve making or sifting through visual art, but she thinks most internet users at least have a certain passive expertise with photographic media. "I still have to do some research in that area, but I believe that on average people are better at identifying photographs rather than art, since everyone, regardless of profession, is bombarded with photos on a daily basis."

We've hitherto discussed how certain varieties of art can be misinterpreted as AI-generated. Another layer of fiendish complexity, worthy of an article in itself, arises when we start talking about certain marginally older machine learning tools and ways of producing images, which overlap a little with what is now known as generative AI.

Abdou Bouam is a freelance artist whose work encompasses "a bit of everything really" – from 3D videogame props and scenes to advertising product renders, educational videos, and other "game adjacent" work. He recalls an instance when a client questioned some images he'd sent for feedback, calling attention to some blurry and indistinct patches, and asking if they had been generated or edited using AI. The blemishes were, in fact, created by older denoising methods.

A 10 part image from Abdou Bouam, showing different applications of path-tracing and denoising to create a cinematic horror scene. | Image credit: Abdou Bouam

I know even less about the technology of computer graphics than I do about the history of visual art. Thankfully, Bouam is willing to explain further. To summarise his thoughts, denoising is the automated removal of random variations and anomalies from an image. Amongst other things, it's used to clean up renders – that is 2D images of 3D scenes – that are created using pathtracing. Pathtracing is a way of creating physically plausible images of different objects and materials by firing out rays from the camera and bouncing them around until they hit a light source, then calculating how much light is transmitted back along the path to the camera. Each repetition of this process is called a sample, and the finished image is produced by averaging out the discrepancies between samples.

"The image starts very noisy, then as more samples are added and averaged, it converges to a final result that's physically accurate," Bouam tells me. "The only way to get a noise-free image is to render it for an infinite amount of time, which is obviously not feasible, so the better approach is to render it until the noise becomes so low that it can be denoised with little to no quality loss." The question, then, is how much time you can afford to spend bouncing rays around before reaching for the denoiser. The difficulty is that, as Bouam goes on, denoising blemishes can resemble symptoms of AI generation. "Details get lost, blurred, and smeared with a lot of artifacting that just feels like genAI."

The other difficulty is that some denoising tools like Nvidia Optix - first released in 2018, and used by many videogames that feature raytracing or DLSS - do fall under the umbrella of "machine learning". But they pre-date "generative AI" as we now understand it, and Bouam argues that they don't have the "same ethical concerns". Rather than relying on access to non-consensually obtained public data, for example, the models in question are created using archives of tailormade images, which "are purposely rendered with a lot of noise, and then compared to the noise-free ones that took ages to render," as Bouam explains. "It's not trained on stolen and unlicensed work. Even if you wanted to, you can't just take any picture off the internet and train with it, you have to use an actual production 3D scene with all the extra data it creates."

These elder denoisers run locally rather than on datacentres, Bouam continues, and save energy because they drastically shorten the time required to create an acceptable render. He also feels that they don't represent a loss of artistic control. "It doesn't change the way my scene looks, other than smears and splotches if the sample count is too low and the scene is challenging to render. But it doesn't change my lighting, or 'yassify' my characters, or anything else like that." In a nutshell: it's still his art.

Image credit: Nvidia

I hope all these insights and reflections offer some kind of steer, while navigating the rapids of a world in which generative AI is being shoved into every vocation and pastime. I'm conscious, however, that any takeaways from this piece may soon be obsolete.

Again, the tech keeps changing. Today's generated AI tools display fewer of the unearthly mutations that characterised the first iterations of Midjourney and ChatGPT. "Pure visual detection is getting harder and harder," Koppenwallner tells me in closing, "and I think most clues are in the context of an image these days (what am I seeing, why am I seeing this, can it be true that I will get a fully rendered painting in two days for $200, etc). It's one thing to be aware and to protect oneself, but I'd be very, very careful about calling someone out for using AI, based on a hunch."