Marshall McLuhan argued in Understanding Media that every new medium changes not only how information is distributed, but also how it is created. Search engines rewarded discoverability, social media rewarded engagement, and LLMs reward interpretability. Whether we realize it or not, we are beginning to design with this new audience in mind. The problem begins when the signal replaces the thing it was meant to represent.
We have seen this before. Social platforms were designed to help people discover relevant content, but engagement became the dominant signal because it was measurable. Platforms became increasingly effective at predicting what would keep people watching, but not necessarily what would create meaningful connection.
The system optimized for the proxy, and the human experience became secondary.
We are entering a similar transition with AI. As organizations begin creating content that is easier for language models to retrieve, summarize and cite, are they still designing primarily for the person who needs the information, or for the system standing between them?
The distinction matters.
A healthcare article can be perfectly structured for LLMs citations and still fail if a patient cannot understand it. Optimizing for machine understanding is not the problem. The problem begins when machine understanding becomes the definition of success.
The same pattern is emerging in recruitment. Applicant tracking systems and AI screening tools help companies manage growing volumes of applications. As candidates optimize resumes and portfolios for algorithms, the hiring process risks becoming a conversation between machines. Candidates begin writing for the system evaluating them instead of the people they hope to reach. The recruiter becomes the second reader, the portfolio becomes a signal to decode, and the person behind it becomes harder to see.
What the screenshot cannot show
The same transition is happening on the creation side. Generative AI has dramatically changed the economics of creation. Producing interfaces, writing documentation and exploring hundreds of design directions has never been easier. A convincing interface can now be generated in seconds. It can look polished, complete and ready to ship. But a screenshot hides almost everything that makes a product work.
Move that same interface into a real design workflow and the gaps quickly appear. Layouts break once a consistent grid is introduced, components do not behave like a system, interaction states are missing, typography does not scale across breakpoints, and accessibility was never considered. The handoff becomes a collection of pixels instead of a product engineers can actually build.
Get Aurélie Radom’s stories in your inbox
Join Medium for free to get updates from this writer.
The pixels are there. The product is not.
Good design has never been defined by what appears in a screenshot. It lives in the decisions the screenshot cannot show: the spacing systems that create hierarchy, the component architecture that allows products to evolve, the interaction patterns that reduce cognitive load, the information architecture that makes complexity understandable and the accessibility decisions that ensure everyone can use the experience.
These decisions rarely appear in a static image, yet they determine whether a product survives its first release. They transform an attractive interface into something that can evolve over years instead of weeks.
This distinction extends far beyond digital products. An architect is not judged by a rendering alone. A building has to stand, adapt and support the people living inside it. Likewise, designers are not judged by beautiful screens alone, but by whether those screens continue working as products evolve, scale, and adapt to changing requirements and growing complexity.
AI dramatically lowers the cost of producing the visible layer. That makes the invisible layer exponentially more valuable. Ironically, the easier interfaces become to generate, the more important design becomes. Because design was never the pixels, it was everything behind them.
Deciding what remains human
This brings us back to the question at the heart of design. Not whether AI can create, but whether designers know what should remain human. Generation has become abundant, while discernment has become scarce.
Language models benefit from consistency, explicit structure and semantic clarity. Humans respond to personality, ambiguity, surprise and emotion.
Neither is wrong.
The best experiences will need to satisfy both. Documentation should be structured enough for an AI assistant to interpret accurately while remaining understandable for the engineer reading it. Healthcare content should be easy for language models to retrieve without sacrificing the empathy, clarity and nuance patients need. Design systems should be machine-readable without becoming so rigid that they erase creativity.
Google’s People + AI Guidebook describes AI as a collaborator that should augment human capabilities rather than replace human judgment. The goal is not to remove people from the process, but to design systems where people remain meaningfully in control. Machine-centered design does not replace human-centered design, it expands our responsibility.
For decades, designers focused on creating experiences people could understand, trust and enjoy. Today, we must also consider how those experiences are interpreted by the systems standing between us and the people we design for. Optimizing for machines is becoming part of the job, designing for machines is not.
Every technological shift introduces a new proxy. Every proxy creates new incentives. Every incentive eventually shapes what gets built.
Designers do not control those incentives, but they decide whether products ultimately serve the metric or the human behind it. The future of design will not belong to those who optimize everything. It will belong to those who know what should never be optimized. AI did not change who we create for, it changed who reads first.
Our responsibility is making sure those are never confused.