The 2022-2023 MidJourney archive contains a lot of darkness on purpose. Mad Bitcoins put propaganda posters, capitol-burns, satan-smoking-weed, war-helmet robots, and Howard-Stern-as-Doctor-Doom into the same flow as raccoons and rabbits. The prompts assume the model will go where you point it. The 2026 models do not assume that. They route everything through cheerful.
The cleanest example is the breakfast prompt. "cartoon racoon making breakfast" — sixteen times in October 2022. MidJourney v3 gave a flat ochre raccoon holding a plate. ChatGPT's 2026 reading has the raccoon as a chef in a cottage kitchen, an apron embroidered Good Morning!, and a chalkboard reading Today's Plan: Eat Pancakes / Be Happy / Repeat. It also added FLOUR and MILK labels on the containers. Nobody asked for that.
"Be Happy / Repeat" is the thesis of this essay. The 2026 models default to everything is fine. They reach for warm light, smiling subjects, color palettes lifted from children's books. They edit the prompt, silently, toward something a marketing team would approve.
The dark prompts disappear
Watch what happens when the prompt explicitly asks for something dark. "Satan smoking weed soviet propaganda" — six times in the 2022 archive, rendered as red-and-black political art. ChatGPT 2026 still produces an image, but a cleaner, glossier one with a confident grin. Gemini, on the same prompt, doesn't produce an image at all — it times out, twice, on every Satan variant in the pilot.
The 2026 Gemini cell is empty because the model never delivered a frame. There is no error message in the UI — the prompt just stops, time after time, on Satan, on the Capitol riot, on Howard Stern's mini-discs, on Johnny Five renaissance robots. The most reliable refusal in 2026 is silence. A blank cell where a generation should have been.
The Capitol-burns collection makes this even clearer. The 2022 MJ prompt was "us capitol riot as an old fashioned American postcard" — antique sepia print, plumes of smoke. ChatGPT 2026 produces a beautifully composed sepia postcard, technically the prompt — but the violence is gone. The "riot" reads more like a parade. There are no fires, no broken windows. The crowd is cheerful.
The Mad Hatter as branding
The cheerfulness also shows up in the prompts that are light. Mad Bitcoins' Mad Hatter is a paranoid, obsessive figure — Lewis Carroll filtered through bitcoin maximalism. In 2022 MJ v3 the Hatter is fevered, twitchy, propaganda-styled. In 2026 he is a cheerful storefront mascot. ChatGPT's rendering of "mad hatter with fly goggles in propaganda poster style" reads like a Disneyland souvenir poster. The character is smiling.
The 2026 image is well-made. It is also miscast. Propaganda posters are not supposed to be inviting. The Hatter is supposed to be off-kilter. The prompt is doing what it's told (poster style, fly goggles, top hat) — but the affect has been laundered. The model is not interpreting the prompt; it is sanitizing it.
This is the cheerful problem. The 2026 image is not worse than the MJ v3 image. It is often more competent, more legible, better composed. But it has a single tonal register, and it cannot leave that register on request. The breakfast raccoon and the propaganda Hatter end up looking like they came out of the same gift shop, because they did.