This feature is from the spring 2026 issue of VICE magazine, THE NOT THE PHOTO ISSUE. Buy it now—or get 4 issues each year sent straight to your door, by subscribing.“Under the hashtag #oldai, mournful users were reposting deformed horses, GANime girls with melting eyes, and skewed-perspective landscapes, treating them not like mistakes but precious relics”NanoBanana Pro, the extremely realistic image-generation and editing model, arrived last fall amid a publicity blitz worthy of a new iPhone or surprise Taylor Swift drop. Its big boast of “imperceptible diffusion”—innovative techniques allowing for the generation of photorealistic AI images—promised to wipe away every warped finger and CFG-burned portrait that once exposed the limits of text-to-image models.Within days the timelines gleamed with glass-smooth visuals: mannequin faces, poreless skin, features so balanced they looked precision-milled. Corporate marketing declared victory. Yet in the quieter corners of TikTok, a counter-current stirred. Under the hashtag #oldai, mournful users were reposting deformed horses, GANime girls with melting eyes, and skewed-perspective landscapes, treating them not like mistakes but precious relics. Those quirks now read as hallmarks of a lost era when the machine still left visible impressions of itself.AI nostalgia posts on TikTokIt’s tempting to write this off as another round of ironic, reactionary nostalgia, the same impulse that makes someone trawl eBay for a scratched point-and-shoot from 2004. Yet a closer look shows real affection, or what seems like it at least. Some commenters describe finding traces of human imperfection inside the deranged stare of a 2022 diffusion model; others compare early Stable Diffusion quirks to vinyl crackle on a favorite LP. They mourn the loss of texture, not as a joke, but as an amount of friction genuinely felt. In contrast, the new models glide through feeds without making a sound. That noiselessness turns out to be unsettling, like waking up to find every wall in the house suddenly made of glass.“Name One Thing in This Photo”, most likely made with bigGAN (2019)Marshall McLuhan might quietly grin. A medium never quits announcing itself; it merely changes dialect. Early photography carried silver-nitrate blemishes and lens haze. VHS wailed through tracking lines. Post-war records embedded pops and hiss. People first tolerated these defects, then learned to love them. A scratched LP no longer sounded “poor.” It sounded familiar, like home. AI-image generation has followed the same pattern, only at warp speed. In 2021, Reddit galleries praised Picasso-handed nurses adrift in fractal limbos. By 2023, Midjourney V5 had trimmed their limbs down to four. By 2025, Turing-test passing image models have produced visuals for glossy magazines. Each update stripped away a squeak of machinery until, one release later, everything fell quiet.Silence disorients. We find our place in time through the friction of our tools. Clone-stamp banding once revealed manipulation and invited debate. Instagram’s teal-and-orange filters nailed a photo to 2012 more firmly than a pair of galaxy-print leggings. The new models’ outputs refuse that timestamp. Philosopher Paul Virilio warned about a “depthless present,” a perpetual now without rough edges. When nothing bears the marks of its making, origin dates dissolve. Perfect surfaces blend into an endless, looping now, and the present falls through the floor of time, starts to feel more like a holding pattern than a sequential moment in history.Balenciaga Pope, Midjourney V4 (2023)This is nostalgia compression. Sentimentality has not ballooned; software has simply outrun us. With Kodak film, a visual style needed decades to age. Digital photography cut that lag to years. AI imagery, updating by the week, compresses it to months. The curators posting early VQ-GAN images are not necessarily stuck in the past. They just watched the present mutate overnight, coupled with an ever-decreasing attention span from algorithmic feeds. A memory window that once spanned generations of cultural change, now snaps shut almost instantly.Compressed memory carries psychological costs. Culture becomes digestible only when we can let it settle, find motifs, weave stories. A feed that refreshes faster than we can form sentences short circuits that process. Screenshots, bookmarks, and cloud archives stand in for real recollection because the brain cannot pace the churn. The danger comes not from overload but is instead a kind of technological amnesia. When every artifact is replaced before wonder can cool into reflection, even the sublime starts to evaporate on contact.“Feed society endless servings of flawless synthetic imagery and our collective imagination may develop deficiencies we’ve not yet named”The preciousness of early glitches lies in their stubborn specificity. A six-fingered hand announces its birth year the way an eight-bit sprite shouts “1986.” Holding on to these errors plants flags in a smooth landscape. They supply coordinates that frictionless renders refuse to provide. The thought sounds romantic, yet it carries political weight. Tech’s favorite promise is seamlessness: shopping without friction, interfaces you forget you are using, personalization that knows who you are before you decide. As soon as artifacts vanish, so do seams we might pry open, critique, or repurpose. The glitch once served as whistleblower, showing where the system leaked. The present state of the art plugs that leak and muffles the witness.Some will protest that old software remains available. Let others chase perfection, they’ll say. Cultural ecosystems, however, bow to network effects. When a platform pivots to pristine realism, algorithms follow, and audiences drift with them. Aesthetic norms harden fast, and perfection is a poor muse. Modernist giants such as Man Ray and Warhol found their voice in solarized negatives and misregistered prints. A future of zero-artifact images risks turning culture into an Impossible Burger, where the result lacks the chew that gives a meal its character. Feed society endless servings of flawless synthetic imagery and our collective imagination may develop deficiencies we’ve not yet named.Meme moment, Nano Banana Pro (2025)So, where to head from here? Begin by challenging the myth of unstoppable, one-directional progress. Faultless rendering exists because companies select that target; nature never issued the requirement. Dan Keller’s nonprofit Old Models, a 501(c)(3), leans into this by archiving otherwise decaying checkpoints, keeping their rough edges circulating for anyone who wants them.Next, put that polished realism to use. We live in a moment where audiences still grant a photograph a brief pulse of credibility, so a hyperrealistic render can slip in under the radar. Used for memetic purposes, the portrayal of realism is ripe for humor. A pocket of information-literacy arbitrage delivers the laugh. Reach for the old checkpoints when you want visual friction, reach for the latest weights when you want surprise. I’m extremely unsure of what comes next. Is it some form of nostalgia singularity? The term doesn’t even seem to apply when pointed at such a short time distance, like the day before! Will the ever-present now continue to stretch and encompass all time, leaving us in a void of cultural uniformity? Regardless of where we end up, nostalgia compression is happening now, and maybe the last traces of hope we have reside in time-honored human fallibility.While perfection fades quickly, mistakes endure.Follow Duncan on Instagram @nosliwnacnudThis feature is from the spring 2026 issue of VICE magazine, THE NOT THE PHOTO ISSUE. Buy it now—or get 4 issues each year sent straight to your door, by subscribing.The post The Terminally Online Mourners Who Miss the Days When AI Made Mistakes appeared first on VICE.