You know “the look” by now. Every poster for a business or event has pictures of uncanny-looking people, buildings that don’t look right, and always that one font. But why? If generative AI is supposedly able to “think” and be “creative,” why do so many of its results come out looking exactly the same? The answer is complicated, and it has to do with what chatbots define as “thinking.”At this point, gone are the days of chatbots like ChatGPT, Claude, or Gemini simply generating text immediately in response to a prompt. Some very fast models do, but most employ what’s known as “thinking models”, which are a bit of a misnomer. The “thinking” really just consists of generating a set of sub-prompts before executing the main one. And it’s in that 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 that determine how a design looks.The hidden “design brief” behind every AI posterTo borrow a classic and implausible example, if you were to give an AI chatbot the instruction “Make a peanut butter and jelly sandwich,” it would start by generating step-by-step instructions on how to make the sandwich before trying to make one (if it could). Most of this happens invisibly in the background, and many proprietary chatbots won’t show you their in-process “thinking” prompts even if you ask.In the case of an AI poster, it might break down “make a poster” into smaller steps such as “create a design brief for a poster.” To demonstrate this while doing some testing for this piece, I added additional prompts for ChatGPT to explain the “design briefs” it generated, which included details like layout, colors, and typography.The problem is those instructions are still subject to the same flaws as any other AI-generated text. They can’t properly fact-check; they can’t walk into the real world and see the thing they’re describing; and they can hallucinate or fall back on stale formats that become easy to notice after a while.For example, while testing, I had ChatGPT generate a poster for a Lifehacker-themed nightclub. In every case, the underlying brief it generated outlined some basic details, but occasionally it added things I hadn’t asked for. This included additional text, iconography, and in one case it even added the “design brief” explanation I had requested as the bottom half of the poster. Those mistakes filter into the final product, when the brief is then fed into the image generator that ultimately makes the poster. In my case, ChatGPT showed me the underlying explanations for its reasoning (though even that could be different from what was actually sent to the image generator), but most people who use a chatbot for this task won’t see these factors before getting the final result. Credit: Eric Ravenscraft/Lifehacker “Choices” are still driven by statistical modelsWhile most chatbots won’t show you exactly what underlying “thinking” they use, most can be coaxed into describing the vague outlines of how they reached certain outcomes. And those descriptions provide some insight into why certain chatbots keep outputting similar results over and over. It doesn’t take long for some patterns to emerge. For example, any prompt to generate a poster for a nightclub almost invariably generated a brief that had specifications for a color palette including “electric purple,” “electric yellow,” and “electric pink.”The reason for this should be pretty obvious: When the statistical model underlying LLMs needs to come up with an example for “colors for a nightclub poster,” the most obvious answers are neon, high-energy colors. Unless you know something about the specific club’s vibe (and write that out and feed it to the chatbot), the AI will have to fill in the gaps with something. And that something will be based on the same statistics it uses to predict any word.This pattern of repetition was everywhere in all the design rationale ChatGPT gave me. Across multiple attempts to generate a Club Lifehacker poster, the generated design brief added an overarching “concept” that described the fictional club’s atmosphere as “high-energy” and “futuristic.” Every bot mentioned using images with the “silhouette” of either the DJ or a dancing crowd (often both). All the text had to be “bold” and “futuristic."That one font “choice” might be a misinterpretationIf there's one dead giveaway that a poster was made with AI, it's that one brush-stroke font that they all seem to use. Now, this is a bit more speculative, but I have a theory on where that specific design comes from. And it’s worth noting upfront that this seems to only apply to posters generated by ChatGPT specifically.One of the most common factors I found in my design brief testing was the use of phrases like “brush lettering” or “brush script.” Brush lettering is a design term that refers to using any number of methods like markers, paint brushes, or brush pens to write text. In digital design, that can take the form of fonts that are designed to emulate the look of fancy, handwritten text.In every test poster I created, however, ChatGPT seemed to take the phrase “brush lettering” rather literally. Every instance I encountered of a design brief that mentioned any kind of “brush” text generated an image with letters that looked as though they were created with a broad paint brush. A human designer would understand “brush lettering” to mean “something that looks handwritten," but ChatGPT didn’t seem capable of understanding the phrase as anything other than using an actual paint brush.The generated posters aren’t even using a “font,” strictly speaking. (Often, letters that should be identical had slight variations.) The image generator part of the chatbot is still a diffusion model, so it can’t employ a vector-based font, per se. But the brush-style font is so common that its look surfaces in any statistical model.This is an odd quirk of ChatGPT specifically, and it’s possible that it could change in the future. Especially if OpenAI intentionally tries to train this behavior out of its chatbot. However, it’s helpful to illustrate the underlying mechanism. The only way a chatbot can conceptualize design elements in a visual medium is to generate more specific text to feed to the image generator. And that text is just as prone to mistakes and blandness as any other AI model.Every chatbot has a “house style" (even if it's an accident)I’ve been examining this with a focus on the kinds of posters ChatGPT generates because they’re very common, but posters (or any AI-generated content) from other models might look different. Every model has its own tendencies. For example, a lot of vibecoded websites are made with Anthropic’s Claude, and you can often tell by the tiny text, buttons with rounded corners, and tile-based layouts.To a certain extent, this is inevitable. While AI models are very different from humans and do not make “choices” in the same way people do, design elements still have to come from somewhere, and whether human or AI, it’s impossible to make these decisions in a vacuum. Humans will have preferences based on the things they like or emotionally resonate with, and AI models will have tendencies based on what their statistical model determines is the most appropriate output.That means that an AI model’s “style” is naturally going to change over time. Even if a model improves or makes fewer mistakes, it’s impossible to ever change the fact that all ChatGPT images are made by ChatGPT. Whatever you generate with any particular model will have a lot in common with what everyone else who uses the same tool generates.