⏱️ Lectura: 11 min
A village fair poster made with artificial intelligence is recognizable from a distance: rounded typography, watercolor flowers, borders with an old-paper texture, and a pastel background nearly identical to the poster for the fair next door. John Hartnup, a developer and author of the personal blog ‘Ere I Am, JH!, documented the phenomenon on June 7, 2026, after seeing dozens of nearly identical posters circulating on Facebook for spring fairs and beer festivals across the UK.
📑 En este artículo
- TL;DR
- What happened: the AI-generated poster experiment
- Context and history: the problem isn’t capability, it’s the prompt
- Technical details: the twelve styles ChatGPT offered
- How to start testing it
- Impact and analysis
- What’s next
- Frequently Asked Questions
- Which AI model generated the posters in the experiment?
- Why do AI-generated posters usually all look the same?
- Does this technique work with Midjourney or Stable Diffusion?
- Do you need to know graphic design to apply this technique?
- Does this only work for event posters?
- What about the text that appears inside the generated poster?
- References
His conclusion was simple: repetition, not quality, was the real problem. To prove it, he ran an experiment with ChatGPT, using the same details any real organizer would put in a commission, and then forced the model to abandon its default response for AI-generated posters.
TL;DR
- On June 7, 2026, John Hartnup documented the repetitive style of AI-generated fair posters in the UK.
- His first attempt with ChatGPT reproduced the same pastel, bunting-covered look he was trying to avoid.
- Explicitly asking for ‘a completely different design aesthetic’ produced a modernist Bauhaus/geometric poster.
- ChatGPT identified its own output as a ‘Bauhaus-inspired geometric minimalist poster’ when asked to name the style.
- The model provided a menu of twelve alternative styles, grouped into four design categories.
- The options included risography, cut-paper collage, graphic brutalism, Memphis Design, and Japanese minimalism.
- The technique works the same way in DALL-E 3, Midjourney, or Stable Diffusion: naming the design movement changes the result.
- The real gap isn’t in the model, it’s in the design vocabulary the average user has.
What happened: the AI-generated poster experiment
Hartnup asked ChatGPT for a poster for a fictional spring fair: date, time, location, free admission, a tombola, craft stalls, and performances by a samba band and a dhol band. He added explicit style instructions: a clean, unadorned design with an eye-catching spring graphic, avoiding a pastel, airbrushed, or oil-painting style, and no images of people.
The result didn’t meet the brief. Despite the instructions, ChatGPT returned a variation of the same crafty fair poster Hartnup wanted to avoid: soft illustration, muted colors, the default aesthetic that dominates thousands of AI-generated community posters. The negative constraint, telling the model what not to do, wasn’t enough to pull it out of its usual template.
The second attempt changed tactics. Instead of describing what he didn’t want, Hartnup asked ChatGPT to use a completely different design aesthetic, one of its own choosing, and to treat the first result as an example of what not to do, not because it was wrong, but because the goal was for the poster to stand out from similar ones. The result was a poster with geometric shapes, bold sans-serif typography, and a high-contrast palette: a complete turnaround from the first attempt.
Context and history: the problem isn’t capability, it’s the prompt
The phenomenon Hartnup documented has a name in recent tech conversation: ‘AI slop,’ the flood of AI-generated visual content that looks generic through repetition, not through poor individual quality. Any single poster can pass a quick glance, but cumulative exposure (twenty fair posters with the same look in the same feed) turns an acceptable result into a signal of laziness.
Image generation models don’t fail for lack of variety in their training data. They fail because a short, generic prompt like ‘poster for a spring fair’ collapses toward the statistically most likely response, the one the model most frequently associates with that combination of words. Naming a specific design movement pulls the model out of that zone of maximum probability and forces it to search a different corner of its style space.
Technical details: the twelve styles ChatGPT offered
When Hartnup asked ChatGPT what the style it had just used was called, the model answered precisely: modernist, with Bauhaus influence, and a touch of contemporary flat illustration on top. It broke it down into three recognizable components: the Bauhaus tradition (asymmetric hierarchy, bold sans-serif typography, a limited high-contrast palette), geometric minimalism (basic shapes, almost no texture or shading), and the influence of the Swiss Style, or International Typographic Style (clean grid, information-centered layout).
Asking for a name wasn’t the end of it: Hartnup also asked for a full menu of styles to choose from next time. ChatGPT returned twelve options grouped into four categories, each with a concrete description of how it would look applied to the same fair poster.
| Category | Included styles | When to use it | Limitation |
|---|---|---|---|
| Clean but with character | Bauhaus, Swiss Style, magazine editorial | Events that want to look curated, almost museum- or gallery-like | Can feel cold for a very local or family-oriented audience |
| Graphic and illustrative | Risography, cut-paper collage, botanical illustration | When the event has a genuine craft component and you want to keep it without falling into the generic | Risography needs to simulate grain and ink misregistration, an effect that’s hard to control |
| Bold or unusual | Graphic brutalism, 90s rave, Memphis Design | Youth events, electronic music, high-energy festivals | May alienate an older or more conservative audience |
| Understated and distinct | Japanese minimalism, monochrome with an accent color, signage or wayfinding | When the goal is to stand out through contrast, not visual volume | Requires a lot of white space: doesn’t work if there’s too much information to fit |
💭 Key takeaway: the model didn’t need extra training or a design plugin: all that variety was already available, it just had to be requested by name.
How to start testing it
The technique can be replicated in any conversational image generator (ChatGPT with gpt-image-1, Midjourney, or Stable Diffusion) in three steps. First, an initial prompt with the event’s real details and a simple negative constraint. Second, if the result is generic, explicitly ask for a different style and name the previous one as a reference for what not to repeat. Third, ask the model itself for the technical name of the resulting style so it can be reused without iterating again.
Initial prompt:
"Produce a poster for a spring fair.
April 21, 11am to 3pm.
Mill Beach Park, Honeyford. Free admission.
Tombola, cakes and drinks, samba band and dhol band,
craft stalls, circus workshop.
Clean design, no ornamentation, eye-catching spring graphic.
Avoid pastel, airbrushed, oil-painting styles, or images of people."That first prompt isn’t enough on its own. The second step is what really changes the result:
Refinement prompt:
"Make another version using a completely different design
aesthetic, of your choosing. Treat the previous version as an
example of what NOT to do (not because it's wrong, but because
we want ours to stand out from similar posters)."To automate this in your own pipeline, for example generating poster variants for several events in a row, it’s enough to call OpenAI’s image API and pass the already-identified style in the prompt, without repeating the refinement step every time.
On macOS and Linux, with curl:
curl https://api.openai.com/v1/images/generations \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-image-1",
"prompt": "Bauhaus-inspired geometric minimalist poster for a spring fair, bold sans-serif typography, high-contrast palette, no people",
"size": "1024x1536"
}'On Windows, the same call with PowerShell:
$headers = @{ Authorization = "Bearer $env:OPENAI_API_KEY" }
Invoke-RestMethod -Uri "https://api.openai.com/v1/images/generations" `
-Method Post -Headers $headers -ContentType "application/json" `
-Body '{"model":"gpt-image-1","prompt":"Bauhaus-inspired geometric minimalist poster for a spring fair","size":"1024x1536"}'⚠️ Heads up: image generation models still make typographic errors when rendering text inside the poster. Check every word (date, time, address) before printing or publishing, because a misspelled detail on the final poster is worse than a generic style.
To confirm the style actually changed and isn’t just a cosmetic variation of the same template, a simple check works: ask the model to describe the style of each generated version in one sentence. If two answers describe essentially the same thing in different words, for example ‘soft, warm illustration’ versus ‘friendly, colorful design,’ the prompt never left the generic zone.
flowchart TD
A["Initial prompt with event details"] --> B["Result with generic style"]
B --> C["Ask for a completely different aesthetic"]
C --> D["Result with defined style (Bauhaus)"]
D --> E["Ask for the style's name"]
E --> F["Save the prompt as a reusable template"]Impact and analysis
Hartnup’s case matters beyond a single fair poster. Any small business, nonprofit, or community group that now uses AI to generate graphic material (flyers, podcast covers, social media banners) faces the same problem with AI-generated posters: the default output tends toward a recognizable visual average, and that homogeneity dilutes brand identity right when producing images became cheaper than ever.
The real limitation isn’t in the model, it’s in the user’s vocabulary. Naming ‘Bauhaus’ or ‘risography’ requires knowing graphic design history, something a community fair organizer or a social media manager doesn’t necessarily have. The solution Hartnup found, asking the model itself for a menu of styles, resolves exactly that gap: it turns the image generator into its own reference catalog.
The trade-off shows up in the typography. None of the bold styles (90s rave, Memphis Design) solves the most commonly cited problem with text in image generators: spelling and kerning errors in the poster’s own words. Changing the visual style doesn’t fix that technical limitation, so any AI-generated poster needs a manual text review before it goes to print.
What’s next
It’s likely that prompt libraries or reusable style packs will emerge, collections of already-tested instructions for Bauhaus, risography, or brutalism, as a direct response to this problem, similar to the presets that already exist for AI-generated photography. It’s also foreseeable that image generation tools will start exposing more explicit style selectors in their interface, instead of leaving all the vocabulary work to the user inside a free-text prompt.
📖 Summary on Telegram: View summary
Try it yourself: build the refinement prompt from this article with your next event’s details and compare the result against the first, generic attempt.
Frequently Asked Questions
Which AI model generated the posters in the experiment?
ChatGPT, using its built-in image generator (based on gpt-image-1), with no plugins or external tools.
Why do AI-generated posters usually all look the same?
Because a generic prompt like ‘poster for a spring fair’ triggers the statistically most likely response in the model’s training data: pastel illustration, rounded typography, and bunting-covered borders.
Does this technique work with Midjourney or Stable Diffusion?
Yes. The principle is the same in any generator: naming a specific design movement reduces the prompt’s ambiguity and pulls the model out of its default response.
Do you need to know graphic design to apply this technique?
It helps, but it’s not required. Asking the model itself for a menu of styles, as Hartnup did, solves the problem of not knowing the technical names.
Does this only work for event posters?
No: it applies equally to podcast covers, video thumbnails, or social media banners, any piece where a generic AI style is a branding problem.
What about the text that appears inside the generated poster?
Image models still make spelling and spacing errors when rendering text. It’s best to check every word before publishing, or consider adding the text afterward in a separate editor.
References
- Ere I Am, JH!: John Hartnup’s original article with screenshots of each attempt and the full list of twelve styles.
- OpenAI: the official page for the image generator used in the experiment.
- Wikipedia: history and principles of the Bauhaus school, the style reference ChatGPT reproduced.
- Wikipedia: the Swiss Style, or International Typographic Style, the other influence the model cited.
📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
0 Comments