AI Tools Gave Designers Superpowers and Took Away Their Words

Andrej Karpathy called AI “a powerful alien tool with no manual.” He meant it about large language models, but the line lands hardest in design, where the discipline has always run on manuals. Grid systems. Type hierarchies. Brand guidelines. Pattern libraries. Design was never just about making things. It was about having a shared vocabulary for explaining why you made them.

That vocabulary doesn’t exist for AI-assisted design. Not yet.

The pre-convention era #

A widely circulated essay on UX Collective compared the current moment to designing for the web in 1999, before the hamburger menu, before card UI, before mobile-first thinking, before anyone agreed on what a “good” website even was. Designers had HTML. They did not have conventions. The analogy is precise. Today’s practitioners have generative AI, and they do not have a shared framework for evaluating what comes out of it.

The terminology that does exist lives in the wrong rooms. “Design tokens,” “drift signals,” “nearest established name.” These phrases circulate in engineering documentation and systems specs at companies like Sameness.co. They haven’t crossed into the studios, the critiques, the portfolio reviews. A designer using Midjourney to generate concepts and a designer hand-drawing wireframes cannot describe their respective processes in a common tongue. They’re working in the same field with incompatible grammars.

Output without judgment #

Platforms like Looka and Tailor Brands now generate complete brand identity systems from text prompts in minutes. Logo, color palette, typography, layout. A founder can have a full visual identity before they can explain what their brand stands for. The manual for using the tool exists. The manual for judging its output does not.

And judgment is the entire game. AI tools suggest color palettes, generate layouts, produce visual concepts from a sentence. They cannot evaluate whether any of those outputs align with a brand’s values, user research, or strategic positioning. That layer is human. It has always been human. But the speed of the tools is compressing it out of the workflow. When you can generate forty options in ten minutes, the pressure shifts from “what should we make?” to “which one looks best?” Those are different questions. The second one is worse.

The demos make it harder. AI tool demos show best-case outputs and skip the invisible labor. The retries, the prompt refinement, the adjustments that separate a lucky generation from a usable one. Beginners see the demo, try the tool, get mediocre results, and blame themselves. Nobody told them the demo was a highlight reel. Nobody could, because the shared understanding of what “normal” AI-assisted work looks like hasn’t solidified.

Convergence as the default #

The structural problem runs deeper than workflow. AI design tools train on enormous volumes of existing work, which means their outputs converge toward the average of what already exists. A Medium essay by S.M. Roqunuzzaman documented this across product interfaces. In the writer’s words, “Users do not remember your product. They do not feel anything about it. It is just another app that looks like every other app.” The tools are fast. The tools are capable. The tools produce sameness.

Research from Historica backs this up at a conceptual level. AI and machine learning algorithms trained on existing data reinforce existing preferences rather than challenge them. The tools structurally reproduce the canon. They don’t expand it. Brand differentiation, the thing design is supposedly for, requires the opposite of what the default output provides.

So the real skill is strategic thinking about what should be different and why. That skill is precisely what the tools cannot supply. And the tools’ ease of use is quietly devaluing it. When anyone can generate a “professional” logo in ninety seconds, the market stops paying for the part that actually matters, the reasoning behind the mark.

The missing discipline #

Design has survived every previous tool shift because it adapted its language. Desktop publishing brought “bleed” and “CMYK” into every studio. Web design gave us “above the fold” (borrowed, but adopted). Mobile design introduced “thumb zones.” Each era produced constraints, and from constraints came vocabulary, and from vocabulary came the ability to teach, critique, and improve.

AI tools arrived with no constraints anyone agrees on. The result is a generation of makers building at extraordinary speed with no way to explain their decisions to a client, a collaborator, or themselves. The power is real. The fluency isn’t. And the gap between those two things is where craft goes to disappear.

What fills that gap will define whether AI-assisted design becomes a discipline or stays a parlor trick. Somebody is going to write the manual. The question is whether designers write it or whether it gets written for them, by the platforms, in the language of engineering, optimized for engagement metrics. The conventions that will feel obvious in ten years are being invented now, in studios and Slack channels and one-off client presentations. Most of the people inventing them don’t know that’s what they’re doing.

Subscribe to Threads of Culture for weekly writing on design, branding, and the forces reshaping both.