Why Do All AI-Generated Posters Look Like This?

You’re probably familiar with this “style” already. Every company or event poster features strange-looking people, unnatural-looking buildings, and always the same font . But why? If generative AI is supposedly capable of “thinking” and being “creative,” why do so many of its results look exactly the same? The answer is complex and has to do with what chatbots define as “thinking.”

Gone are the days when chatbots like ChatGPT, Claude, or Gemini simply instantly generated text in response to a query. Some very fast models do this, but most use so-called ” thinking models ,” which is somewhat misleading. “Thinking” actually consists of generating a set of subqueries before executing the main one. And it is in this hidden, generated text that most of the “choices” (to the extent that we can call the statistical output of an AI model a “choice”) occur, determining the appearance of the design.

The hidden ‘technical challenge’ behind every AI-generated poster.

As a classic and unlikely example , if you gave an AI chatbot the instruction “Make a peanut butter and jelly sandwich,” it would start by generating step-by-step instructions for making the sandwich before attempting to make it (if that were possible). Much of this happens unnoticed in the background, and many proprietary chatbots won’t show you their “thinking” prompts as they work, even if you ask.

You may also like

In the case of an AI-generated poster, the “poster creation” process can be broken down into smaller steps, such as “creating a poster design brief.” To demonstrate this during testing of this material, I added additional prompts for ChatGPT to explain the “design briefs” it generated, which included details such as layout, colors, and typography.

The problem is that these instructions are still subject to the same flaws as any other AI-generated text. They can’t properly fact-check; they can’t go out into the real world and see what they’re describing; and they can hallucinate or use outdated formats that are easily noticed after a while.

For example, during testing, I asked ChatGPT to generate a Lifehacker-style poster for a nightclub. In each case, the generated specifications contained the basic details, but sometimes they added features I hadn’t requested. These included additional text, icons, and in one case, they even added the explanation I requested for the “specifications” at the bottom of the poster.

These errors make their way into the final product when the specifications are passed to the image generator, which ultimately creates the poster. In my case, ChatGPT showed me explanations of its reasoning (though even this may differ from what was actually sent to the image generator), but most people using a chatbot for this task won’t see these factors until the final result is ready.

Photo: Eric Ravenscraft/Lifehacker

“Choice” is still determined by statistical models.

While most chatbots won’t reveal the exact “logic” they’re using, most can be induced to describe the general outlines of how they arrived at certain results. These descriptions provide some insight into why some chatbots consistently produce similar results. Some patterns quickly emerge. For example, any request to create a poster for a nightclub almost always resulted in a specification specifying a color palette including “electric purple,” “electric yellow,” and “electric pink.”

The reason for this should be fairly obvious: when the statistical model underlying LLM needs to come up with an example for “colors for a nightclub poster,” the most obvious answers are neon, bright colors. If you don’t know anything about the atmosphere of a particular club (and don’t write it down and feed it to the chatbot), the AI ​​will have to fill in the blanks with something else. And that “something” will be based on the same statistical data it uses to predict any word.

This recurring pattern was ubiquitous in all the design proposals provided to me by ChatGPT. During numerous attempts to create a poster for the Lifehacker club, the generated brief added a general “concept” to the design, describing the fictional club’s atmosphere as “energetic” and “futuristic.” Every bot mentioned using images featuring a “silhouette” of a DJ or a dancing crowd (often both). All text was to be “bold” and “futuristic.”

Perhaps this “choice” of one font is a misinterpretation.

If there’s one undeniable sign that a poster was created by AI, it’s the brushstroke-like font that seems to be used on all the posters. This is, of course, a bit of a guess, but I have a theory about the origin of this particular design. And it’s worth noting right away that this seems to only apply to posters created specifically with ChatGPT.

What do you think at the moment?

While testing the design brief, I found that one of the most common factors was the use of phrases like “brush calligraphy” or “brush calligraphy font.” Brush calligraphy is a design term that refers to the use of various techniques, such as markers, brushes, or felt-tip pens, to write text. In digital design, this can manifest itself in fonts created to mimic the look of exquisite handwritten text .

However, in every test poster I created, ChatGPT seemed to take the phrase “brush calligraphy” quite literally. Every time the specification mentioned any “brush-drawn text,” an image appeared with letters that looked as if they were created with a wide brush. A human designer would understand “brush calligraphy” as “something that looks like handwritten text,” but ChatGPT seemed unable to interpret the phrase as anything other than the use of an actual brush.

The generated posters don’t even strictly use a “font.” (Often, letters that should be identical had minor variations.) The part of the chatbot responsible for generating images still uses a distribution model , so it can’t use a vector font per se. But the brush-style font is so common that its appearance is detected in any statistical model.

This is a peculiar feature of ChatGPT, and it’s entirely possible that it will change in the future, especially if OpenAI deliberately tries to eliminate this behavior from its chatbot. However, it’s useful to illustrate the underlying mechanism. The only way a chatbot can conceptualize design elements in a visual environment is by generating more specific text to feed to an image generator. And this text is just as prone to error and uniformity as any other AI model.

Every chatbot has its own “brand identity” (even if it’s just a coincidence).

I studied this issue, focusing on the types of posters generated by ChatGPT, as they are very common, but posters (or any AI-generated content) generated by other models may look different. Each model has its own characteristics. For example, many websites created with vibecode are built using Claude by Anthropic, and this can often be identified by small text, buttons with rounded corners, and a tiled structure.

To a certain extent, this is inevitable. While AI models are very different from humans and don’t make “choices” in the same way, design elements still have to come from somewhere, and whether humans or AI make these decisions in a vacuum. Humans will have preferences based on what they like or what they connect with emotionally, while AI models will have tendencies based on what their statistical model determines to be the most appropriate outcome.

This means that the AI ​​model’s “style” will naturally change over time. Even if the model improves or makes fewer errors, it’s impossible to change the fact that all ChatGPT images were created using ChatGPT. Everything you generate with any given model will have a lot in common with what everyone else using the same tool generates.

More…

Leave a Reply