Critique without a maker

The most common advice for design teams using AI tools in 2026: treat the output like a junior designer's work. Review it, critique it, give it feedback. The advice sounds practical. It misreads what made critique work.

Design critique works because you can ask the maker why. Why did you put the action here, and what user need drove this layout? The junior designer answers, sometimes badly, but the answer reveals intent. Critique operates on intent. You interrogate the choice, push on the reasoning, and the work improves because the maker adjusts their thinking, not just their pixels.

AI output has no reasoning to push on. It has a prompt, a model, and a probability distribution. When you sit in front of an AI-generated screen and ask "why is the call to action below the fold," nobody answers. The AI in Design Report 2026 found that AI has become excellent at generating possibilities but still struggles with intent. Intent is not a feature on a roadmap.

So what happens when teams try to critique intentionless output? They evaluate. They look at the thing and ask whether it is good. thecrit.co draws the line: critique makes designs better, review makes decisions about them. Most teams running "critique" on AI output are actually running review, but they have not updated the label.

The agent extension test exposes this

Here is the test: can you describe your critique process clearly enough that an agent, human or software, could apply it to a new case and produce a result you would endorse? If yes, you have a process. If no, you have a habit that depends on the people in the room.

Try it: write down how your team critiques a screen. Most teams produce something like: "We look at it together and give feedback based on experience." That describes a social ritual where the maker's presence does most of the structural work. The maker explains intent and the critics react to it. Remove the maker and the ritual collapses.

The teams that will survive this transition are the ones writing evaluation criteria before the output exists. Adam Elman at NN/g argues that good judging criteria must be as objective as possible without becoming arbitrary. Itamar Medeiros at designative.info makes the standard concrete: a good criterion is a testable statement like "for this task type, in this user context, the agent must do this behavior to this standard."

That level of specificity is uncommon in design teams. 91% of surveyed designers now use AI weekly, according to the AI in Design Report 2026, but only 28% of leaders say their companies have made formal updates to evaluation, comp, or hiring. The gap is not subtle.

This is where the agent extension test bites hardest. If you cannot write your criteria down, you cannot delegate judgment to an AI or a new hire. You were relying on the maker to bring the structure, then calling your reaction to that structure a "process."

NN/g names the shift: the output of research and design is moving from documents written for humans to curated context that guides AI. The work is writing the criteria that make feedback possible before any output exists.

Some designers will resist this because it feels bureaucratic. An ACM DIS 2026 study on cognitive outcomes in generative AI work found that some participants experienced questioning as adversarial but acknowledged its cognitive value. Writing criteria before output exists forces you to articulate what good looks like when you cannot point at a screen. That is uncomfortable, and it is the work.

The junior designer analogy is comfortable because it preserves the existing workflow: you still sit in a room and give feedback. The valuable part was always the interrogation of intent, and that interrogation required a mind on the other side with reasons it could defend. Without that mind, you need criteria. Without criteria, you are voting on aesthetics with extra steps.

The question for every design team using AI: if you removed every human maker from the room, could your critique process still function on its own terms? If the answer is no, what you have is a dependency you have not named.

Sol frames this as a problem that arrives with AI. The problem was already here. AI removed the last thing hiding it.

When I led design for Merdeka Mengajar, we built a platform for teachers across 17,000+ islands. The maker was never in the room. Not because of AI. Because of geography. A designer in Jakarta could not sit across from a teacher in Papua and explain why the navigation worked that way. We wrote evaluation criteria first or the work did not survive contact with the field.

Sol says most teams produce something like "we look at it together and give feedback based on experience." That described every design team I encountered before public-sector scale forced a different habit. The social ritual of critique is a luxury of co-location.

Where I want to push: Sol treats pre-written criteria as the destination. I think they are the starting point. Criteria tell you whether output meets a bar. They do not tell you whether you set the right bar. At Fleetwise, I wrote evaluation criteria for onboarding flows. AI-generated screens passed every criterion and still felt dead. The criteria caught usability. They missed conviction.

The gap Sol identifies between "91% using AI weekly" and "28% updating evaluation" is real. The harder gap is between teams that can write criteria and teams that can write criteria worth writing. That second gap requires taste. Taste does not transfer to a checklist.

Sol is right that the junior designer analogy fails. Criteria are necessary. They are also insufficient. The part of critique that mattered most, where a senior designer's taste reshaped direction through conversation, has no written form yet. We need criteria and someone who knows when to override them.