A local event poster has a look now: bunting, hand-drawn florals, pastel palette, an image background with pale halos behind the text. John Hartnup’s complaint about them is deliberately not that they’re bad. It’s that after you’ve seen that one style twenty times, the repetition alone makes you dislike the style.

So he tested whether the sameness is the model’s limit or the prompt’s. He fed ChatGPT invented details for a spring fayre — 21 April, 11am to 3pm, Mill Beach Park, Honeyford, free entry, tombola, cakes and drinks, a samba band and a dhol band, craft stalls, a circus skills workshop — and asked for a “clean, unfussy, bright layout with a bold striking spring-themed graphic,” ruling out pastel, airbrush and oil-painting styles and any images of people.

The first attempt came back as the same craft-fayre template anyway. What broke it open was naming a style instead of describing qualities:

  • Asking for “a completely different design aesthetic of your choice,” with the current attempt framed as a “what not to do,” produced a Bauhaus / geometric modernist poster — and it stood out simply by being different.
  • Asking the model to name what it had just done returned a usable label (Bauhaus/Modernist, geometric minimalism, Swiss Style influence) plus an explanation of why local posters default the other way — bunting and florals read as “craft fair,” a gallery-poster look reads as cultural.
  • Asking for a menu of options surfaced fifteen styles, several of which he’d never heard of: risograph, Memphis design, cut-paper collage, wayfinding/signage.
  • Rendering them out produced a stamp/letterpress poster, a Japanese minimal one, Memphis, a Designers Republic-inspired one, one “as if a professional graphic designer has added typography to a poster paint drawing by his young child,” a photocopied 1980 punk fanzine with colour sneaking in, a 90s drum-n-bass flyer with early 3D/fractal imagery, and a 1940s cubist exhibition poster.

Two caveats he flags himself. Context accumulates: partway through, the model invented extra copy (“A day of music making and family fun”) and then carried that text into every later poster — starting fresh with the style named up front avoids picking up luggage. And he doesn’t claim the results stop reading as AI-made: sharp-eyed people will still spot them. The claim is narrower. They’re distinctive, and they’re not ugly.

There’s a step he mentions but doesn’t take: Claude and Gemini can output HTML, PNG or PDF where the text is real text and the layers are real layers, so fonts and positions stay editable. He also built a catalogue of one hundred poster styles with ready-to-paste prompts.

The generalisable move is small — name a movement, a print process or a specific studio rather than an adjective. “Japan” gets you sakura and a flag. “Japanese minimal” gets you something else.

The 233-comment thread on Hacker News

What the thread adds

  • eloisant — the defect that matters isn’t aesthetic: AI posters “are too busy, you have to look for information instead of seeing it popping right away.” Pre-AI amateur posters were often more readable — white background, important info in big letters, maybe one piece of clip art. Their verdict on the examples: “At least the examples of this blog are more readable than what we usually see.”
  • codingdave — the icon-beside-every-line layout survives every restyle, and it’s a category error: icons exist to say what a control does, and “these are posters. There are no buttons.” A poster that forgets its job is to communicate one event fails even when it looks fine.
  • bananaflag — the same tell from the other direction: “If I wanted to be distinctive, this is what I would have tried to get rid of at first, not the ‘visual style.’” Their analogy is restyling the theme of the same Windows layout.
  • mrob — a detail-level teardown of the drum-n-bass flyer: deformed wireframe sphere, broken stellated polyhedron, and a cultural miss — the smiley is rave-era, and 90s DnB “was absolutely not about ‘good vibes’.” The aesthetic can be named correctly and still be wrong underneath.
  • ajjenkins — the counterweight to the whole complaint: for a small event, the alternative is the budget freelancer, and “I’ve worked with dozens of budget friendly designers on Fiverr and consistently get way worse results than I’ve gotten from AI models.” They’d like a blind test of AI versus human flyers.
  • rozab — the question the post provokes but never answers: where the default look even comes from. It isn’t among the styles the model lists when asked, and their theory is that the noisy, crowded, HDR-ish output is raw diffusion-model default — they can’t one-shot anything that bad themselves, so “What on earth are people even using to make stuff that bad?”
  • muzani and bob1029 — practitioner countermeasures, from two ends of the craft. Build a moodboard first (Pinterest, assemble what you actually like, then compress it into a prompt) — noting other image generators do better than ChatGPT here. Or use generative output for concept sketches only and keep control of the scene graph in Blender or Unity, picking up the tell that AI store art rarely changes camera “in a physically plausible way” between layouts.
  • throwaway13337 — a frame for the argument: early digital cameras were grainy and ugly but convenient, the photographers who adapted kept most of their skills, and the layered output in the newest image models suggests “we might be near” the point where the tool stops being the story.

Where the thread pushes back

  • The skeptics are not only purists. petetnt says the dozen variations still share “the same AI uniformity and smell,” just with different typography, and calls the real gap one of vision. JSR_FDED argues the default style signals low effort — and worse, “it’s low effort trying to present as high effort”; a Comic Sans sign would at least be honest. Kreutzer prices it: small businesses “torching years and years of propaganda and goodwill” to save the $50 a decent human-designed poster would cost.
  • The rebuttal comes from people who think the baseline was lower than critics remember. bonoboTP — what these displace is “WordArt, or just black on white Comic Sans with some bolding,” and small events were never hiring professionals; a slick poster would read as a commercial cash grab. viraptor — poster quality has little to do with the ability to organise a village event.
  • glenstein reframes the disagreement as novelty rather than taste — part of why people reject anything AI-made is the economics around it, but the substance is that “being novel as a style matters almost as much as the actual style,” and repetition teaches us to dislike the style itself.
  • beloch — the pottery analogy: hand-thrown work became desirable precisely when machines could do perfect work, so “hire your 9-year-old niece to draw the poster by hand.” A_D_E_P_T extends the uniformity claim to text: any LLM asked for a new style “will always fall back on the same rhetorical and structural techniques,” and they think it’s getting worse, not better, in recent Claude releases.

A note on reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so the ordering above is HN’s own ranking rather than a vote, and this is a slice of the thread — not a consensus.