Background & Context§
AI image generators and large language models (LLMs) have become ubiquitous tools for rapid content creation. However, their widespread adoption has led to a homogenization of visual design, particularly in event promotion. A now-famous Facebook post highlighted the "scourge of identikit posters generated by AI"—a phenomenon where countless events, from local beer festivals to community fairs, feature strikingly similar aesthetics: glossy gradients, generic 3D text, and pastel color palettes. This repetition, while not inherently "bad" in isolation, breeds irritation through sheer ubiquity.
The recent blog post by John Hartnup (john.hartnup.uk, June 7, 2026) directly confronts this issue. Hartnup argues that the problem isn't that AI posters are terrible, but that they are extit{repetitive}. He notes that even ChatGPT is capable of a broader variety of styles, and he set out to prove it by prompting the model to generate posters with deliberately different design aesthetics. This matters because it challenges the assumption that AI tools are inherently limited to a single, clichéd look—and it highlights the critical role of prompt engineering in escaping stylistic ruts.
The News: What Happened Exactly§
Hartnup begins by acknowledging the viral Facebook post that sparked the conversation, as well as an article in extit{The Independent} covering the same phenomenon. He also shares an example from the wild—a poster for the Leamington Beer Festival—noting that it is "by no means unique." The core issue, he writes, is not that these posters are poorly made. "They're OK, I don't love them," he admits. "The problem is that once you've seen that style 20 times it starts to irritate just from the sheer repetition." This distinction is crucial: the posters are technically competent, but their stylistic uniformity creates a cognitive backlash.
To test whether this uniformity is a limitation of the tool or of the user, Hartnup conducted a simple experiment. He gave ChatGPT invented event details and asked it to produce a poster for a spring fayre. The details were:
- Date: 21 April, 11am to 3pm
- Location: Mill Beach Park, Honeyford
- Free entry
- Attractions: Tombola, cakes and drinks, performance by a samba band and a dhol band, craft stalls, circus skills workshop
Crucially, he specified a desired style: "Go for a clean, unfussy, bright layout with a bold striking spring-themed graphic. Avoid pastel/airbrush/oil style art or images of people." Despite this explicit instruction, the result was disappointing. Hartnup writes, "Not as successful as I hoped. To me it still feels like what I'm trying to avoid." This reveals a key challenge: LLMs often default to their training distribution, even when prompted otherwise.
Undeterred, Hartnup escalated his request. He asked for "another one using a completely different design aesthetic of your choice" and explicitly framed the previous attempt as a "what not to do." This meta-instruction—defining the negative space—proved effective. While the primary source text is truncated, the implication is clear: by forcing the model to abandon its default and choose a new aesthetic, Hartnup obtained a poster that stood out. This experiment demonstrates that the solution to homogeneous AI design is not to reject the tools, but to use more sophisticated, constraint-based prompting.
Historical Parallels & Similar Incidents§
This phenomenon is not unique to AI-generated posters. In the early 2010s, the rise of responsive web design frameworks like Bootstrap led to a similar homogenization of website aesthetics. Thousands of sites adopted identical grid layouts, button styles, and color schemes because the framework made it easy to produce competent, mobile-friendly designs. Critics coined the term "Bootstrap fatigue" to describe the visual monotony. Just as with AI posters, the issue was not that Bootstrap sites were bad—they were functional and accessible—but that they all looked the same. The lesson is that tools that lower the barrier to entry often inadvertently lower diversity.
A more direct parallel can be found in the history of desktop publishing. When Adobe PageMaker and later Microsoft Publisher democratized print design in the 1980s and 1990s, a flood of amateur newsletters and flyers emerged. Many used the same default templates, clip art, and fonts (e.g., Comic Sans, Papyrus). Design professionals lamented the "death of taste," but over time, users developed more sophisticated skills. Similarly, AI poster generation is currently in its "clip art" phase. The difference is the speed of iteration: LLMs can generate hundreds of variations in seconds, but without deliberate prompting, they converge on a single "safe" style.
Both historical cases show that the solution is not to ban the tool but to educate users. In the Bootstrap era, designers learned to customize frameworks or use them as a starting point rather than a final product. In desktop publishing, users gradually adopted better typography and layout principles. For AI posters, the path forward involves teaching users to prompt adversarially—to explicitly avoid the default. Hartnup's experiment is a microcosm of this: by telling ChatGPT what extit{not} to do, he forced it into a different region of its latent space.
The Role of Prompt Engineering in Escaping the Default§
The default aesthetic of AI image generators is a direct reflection of their training data. Models like DALL-E, Midjourney, and Stable Diffusion are trained on massive datasets scraped from the web, where certain styles (e.g., digital art, concept art, glossy 3D renders) are overrepresented. When prompted with a generic request like "poster for a spring fayre," the model samples from the densest region of its latent space—the region that corresponds to the most common visual patterns. This is why so many AI posters look alike: they are statistical averages.
To escape this, users must employ what I call "negative prompting" and "style anchoring." Negative prompting involves explicitly listing what you don't want, as Hartnup did with "Avoid pastel/airbrush/oil style art or images of people." Style anchoring involves specifying a concrete, unusual reference—e.g., "in the style of a 1970s Swiss International Style poster" or "minimalist Bauhaus with primary colors." Here is a Python example of how one might structure such a prompt for an API call:
prompt = """ Design a poster for a spring fayre. Details: 21 April, 11am-3pm, Mill Beach Park, Honeyford, Free entry. Activities: Tombola, cakes, drinks, samba band, dhol band, craft stalls, circus skills. Style: Clean, unfussy, bright. Bold striking spring-themed graphic. Negative: Avoid pastel, airbrush, oil style, images of people. Inspiration: 1960s Japanese graphic design, flat vector, limited color palette. """
By combining negative constraints with a specific historical style, users can guide the model away from its default. The key is to be as restrictive as possible. Vague prompts yield vague (and generic) results. Moreover, users should iterate: generate several variations, then use those as seeds for further refinement. Hartnup's second attempt, where he explicitly framed the first as "what not to do," is a form of iterative refinement. It leverages the model's ability to understand contrastive instructions.
Why This Matters for the AI Community§
The homogenization of AI-generated content is not just an aesthetic problem; it is a symptom of a deeper issue: the tendency of large models to converge on safe, average outputs. This has implications beyond posters. In code generation, LLMs often produce similar boilerplate solutions. In writing, they default to a neutral, journalistic tone. The challenge for developers and users is to build systems that encourage diversity without sacrificing quality.
Hartnup's experiment offers a practical lesson: the tool is not the bottleneck; the prompt is. By demonstrating that ChatGPT can produce a non-default poster when properly instructed, he shifts the blame from the technology to the user. This is empowering, but it also places a greater burden on users to understand how to interact with these models. As AI becomes more integrated into creative workflows, prompt engineering will become as essential as knowing how to use a design tool like Photoshop or Figma.
Ultimately, the problem of "identikit posters" is solvable. It requires a combination of better prompting, more diverse training data, and a willingness to experiment. Hartnup's post is a timely reminder that AI is a tool, not a crutch—and that the quality of its output depends heavily on the quality of our input.