AI Made Design Faster. It Didn’t Make Review Optional.

Let me catch you up. AI stopped being a demo you watch and started being a coworker you supervise. Designers are wiring it into 3D tools, design systems, stock libraries, and classrooms. The fun part is real. The catch is that the same speed that makes a walkable house in a week also ships confident wrong answers to real people. Here is where we are, and what to bring to your next review.

The deep cut

• AI writes the draft, judgment writes the product. Xinran Ma typed 179 prompts to build a house, then spent most of the week revising, not generating.
• You own the interface, not the model. Patrick Neeman calls the wrong answer on its way to a person the second collision, and says designers own it completely.
• A tool is a means, so name the end first. Julie Zhuo's whole method starts with "It would make me happy if," not with picking Claude or Codex.

The week-long build is real, and so is the tradeoff

The speed is not hype. Xinran Ma built a walkable 3D house in seven days by talking to Claude Code through Blender, never touching the software by hand. Her kids showed up mid-project, asked for a second floor and a bouncy house, and became clients. The build cost her a flat subscription fee. The same work at per-use prices would have run $1,570.

The catch is where her time actually went. The first pass was rough. "It didn't take long to get to an initial result," she writes, "but it took much longer to use my judgment and experience to revise it." She traded precision and control for speed. Your team will make that same trade. Be honest that generating is cheap and judging is the job.

The sync that isn't a sync

If you run a design system, know what the AI hookup does and does not do. John Rodrigues points out that Figma's connection to coding agents "just gives an agent faster access to whatever state your design system is already in, drifted or not." A connection is not an audit. When Figma and code drift apart, the agent builds on the drift.

That drift is not free. Rodrigues ties it straight to your costs: more drift means lower-quality output, more re-prompting, and higher token bills. A faster pipe to a bad source pours the mess out faster. Before you brag about your AI-ready design system, check whether anything is actually keeping the two sides honest.

The wrong answer nobody padded

Here is the part to sit with. Patrick Neeman borrows Ralph Nader's frame from car safety. The first collision is the model being wrong. The second is what your interface does with that wrong answer before it hits a person. Confident prose, no visible uncertainty, a citation nobody can click, the answer pre-filled into the field the user was about to send. That second collision is yours.

The numbers make it concrete. The 2026 AI Index measured hallucination rates across 26 models from 22% to 94%. Even the best case, wrong one time in five, is a car with no seat belts. Neeman's fix is a crash test for the interface: seed prompts where the model is wrong, run real users through, and count how many wrong answers reach a real action. Call it the wrong-answer survival rate.

His warning is the industry blames the user. A bad citation gets called "prompting." A user acting on a wrong answer gets called "AI literacy." Telling people to prompt better is driver education with a worse safety record. Your team can build the guardrail now instead of waiting for the model to get honest.

Human still gets the last word

Two smaller moves point the same direction. Luke Wroblewski improved his AI by feeding it 30 years of his own alt tags. The model had titled one image "Green block layout comparison," which nobody would ever search. His human-written alt tag, "Why Off Canvas Layouts?", made it findable. His rule: AI begets more AI, but it "should still defer to us humans for the last word."

Google's classroom work runs the same guardrail. Its generative UI for teachers can spin up a custom physics simulation on request, but nothing enters the public library until a teacher validates and approves it. And the value shows up on the human side, not the machine side. One teacher said the win was finally being able to differentiate instruction, since off-the-shelf simulations mean "you get what you get." The buyers of your favorite categories are voting the same way: Unsplash and Pexels both ban AI-generated uploads, and the little imperfections of real photography now read as a feature.

Three questions for your team

• What keeps our Figma and code from drifting, and is it an audit or just a faster connection to whatever mess we already have?
• What is our wrong-answer survival rate on the AI features we ship, and who is running that test before launch?
• Where does a human get the last word in our AI workflow, and is that a real gate like Google's teacher approval or just a line in the deck?