You cognize “the look” by now. Every poster for a endeavor or event has pictures of uncanny-looking people, buildings that don’t appearance right, and always that one font. But why? If generative AI is supposedly capable to “think” and be “creative,” why do so many of its results arrive out looking exactly the same? The answer is complicated, and it has to do alongside what chatbots define as “thinking.”
At this point, gone are the days of chatbots akin ChatGPT, Claude, or Gemini merely generating content immediately in reply to a prompt. Some extremely accelerated models do, but most employ what’s known as “thinking models”, which are a bit of a misnomer. The “thinking” really fair consists of generating a set of sub-prompts before executing the chief one. And it’s in that hidden, generated content that most of the “choices” (to the degree that we can call the statistical output of an AI example a “choice”) happen that decide how a scheme looks.
The hidden “design brief” rearward all AI poster
To get a traditional and implausible example, if you were to provision an AI chatbot the education “Make a peanut margarine and jelly sandwich,” it would commencement by generating step-by-step instructions on how to create the sandwich before trying to create one (if it could). Most of this happens invisibly in the background, and many proprietary chatbots won’t display you their in-process “thinking” prompts equal if you ask.
In the case of an AI poster, it power interrupt downward “make a poster” into smaller steps specified as “create a scheme concise for a poster.” To display this during doing several evaluation for this piece, I added additional prompts for ChatGPT to explain the “design briefs” it generated, which included particulars akin layout, colors, and typography.
The issue is those instructions are motionless topic to the identical flaws as any another AI-generated text. They can’t correctly fact-check; they can’t stroll into the genuine earth and see the item they’re describing; and they can hallucinate or autumn rear on old formats that rotate into uncomplicated to notice following a while.
For example, during testing, I had ChatGPT create a poster for a Lifehacker-themed nightclub. In all case, the underlying concise it generated outlined several essential details, but occasionally it added things I hadn’t asked for. This included additional text, iconography, and in one case it equal added the “design brief” clarification I had requested as the base fractional of the poster.
Those mistakes display into the final product, whenever the concise is afterward fed into the depiction generator that ultimately makes the poster. In my case, ChatGPT showed me the underlying explanations for its reasoning (though equal that could be distinct from what was really sent to the depiction generator), but most group who use a chatbot for this project won’t see these factors before getting the final result.
“Choices” are motionless driven by statistical models
While most chatbots won’t display 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 provision several understanding into why certain chatbots keep outputting akin results complete and over. It doesn’t obtain lengthy for several patterns to emerge. For example, any immediate to create a poster for a nightclub nearly invariably generated a concise that had details for a color palette including “electric purple,” “electric yellow,” and “electric pink.”
The logic for this have to be beautiful obvious: When the statistical example underlying LLMs needs to arrive up alongside an example for “colors for a nightclub poster,” the most apparent answers are neon, high-energy colors. Unless you cognize item concerning the particular club’s vibe (and compose that out and nourish it to the chatbot), the AI volition have to inhabit in the gaps alongside something. And that item volition be according to the identical data it uses to foretell any word.
This form of repeat was everywhere in all the scheme rationale ChatGPT gave me. Across multiple attempts to create a Club Lifehacker poster, the generated scheme concise added an overarching “concept” that described the fictional club’s ambiance as “high-energy” and “futuristic.” Every bot mentioned using images alongside the “silhouette” of either the DJ or a dancing assemblage (often both). All the content had to be “bold” and “futuristic."
That one font “choice” power be a misinterpretation
If there's one deceased giveaway that a poster was made alongside AI, it's that one brush-stroke font that they all appear to use. Now, this is a bit additional speculative, but I have a theory on anywhere that particular scheme comes from. And it’s value noting upfront that this seems to lone use to posters generated by ChatGPT specifically.
One of the most average factors I established in my scheme concise evaluation was the use of phrases akin “brush lettering” or “brush script.” Brush lettering is a scheme term that refers to using any figure of methods akin markers, coat brushes, or broom pens to compose text. In digital design, that can obtain the form of fonts that are designed to emulate the appearance of fancy, handwritten text.
In all test poster I created, however, ChatGPT seemed to obtain the expression “brush lettering” fairly literally. Every case I encountered of a scheme concise that mentioned any benevolent of “brush” content generated an depiction alongside alphabet that looked as although they were created alongside a broad coat brush. A individual decorator would comprehend “brush lettering” to average “something that looks handwritten," but ChatGPT didn’t appear capable of understanding the expression as item another than using an genuine coat brush.
The generated posters aren’t equal using a “font,” strictly speaking. (Often, alphabet that have to be identical had minor variations.) The depiction generator part of the chatbot is motionless a diffusion model, so it can’t employ a vector-based font, per se. But the brush-style font is so average that its appearance surfaces in any statistical model.
This is an odd quirk of ChatGPT specifically, and it’s imaginable that it could alter in the future. Especially if OpenAI intentionally tries to train this behavior out of its chatbot. However, it’s helpful to exemplify the underlying mechanism. The lone way a chatbot can conceptualize scheme elements in a ocular average is to create additional particular content to nourish to the depiction generator. And that content is fair as prone to mistakes and blandness as any another AI model.
Every chatbot has a “house style" (even if it's an accident)
I’ve been examining this alongside a concentration on the kinds of posters ChatGPT generates since they’re extremely common, but posters (or any AI-generated content) from another models power appearance different. Every example has its own tendencies. For example, a lot of vibecoded websites are made alongside Anthropic’s Claude, and you can frequently inform by the small text, buttons alongside rounded corners, and tile-based layouts.
To a certain extent, this is inevitable. While AI models are extremely distinct from humans and do not create “choices” in the identical way group do, scheme elements motionless have to arrive from somewhere, and whether individual or AI, it’s unattainable to create these decisions in a vacuum. Humans volition have preferences according to the things they akin or emotionally resonate with, and AI models volition have tendencies according to what their statistical example determines is the most suitable output.
That method that an AI model’s “style” is naturally going to alter complete time. Even if a example improves or makes small mistakes, it’s unattainable to always alter the fact that all ChatGPT images are made by ChatGPT. Whatever you create alongside any particular example volition have a lot in average alongside what everyone alternatively who uses the identical tool generates.