Design critique in the AI era: from judging the output to analyzing the decisions

In her newsletter, MC Dean argues that design critique (the "crit") needs an overhaul when an AI model takes part in the creative process. A model can produce a dozen credible variations in very little time, so the final output says far less about the human imprint on the work. That is her starting thesis.

The author returns to a metaphor she used in an earlier piece: working with AI is like driving a Formula 1 car. The car is fast, the driver brings the "feel," and performance comes from the combination of the two. What has changed since then, she says, is that the car now has opinions: it suggests, it pushes back, and it produces things that look like decisions. That is why critique matters more, since it is the only space where the team looks at those decisions together.

She proposes that the crit work like a race "debrief." On an F1 team, nobody spends the meeting on the finishing position; instead, they open up the telemetry: where the driver braked, how much throttle was used, what the car was doing when it was corrected. The learning lies in the decisions. A debrief is not a performance review: driver, engineers and strategists sit on the same side of the table. Critique of AI-assisted work should be a space to learn together, set direction and calibrate what "good" means. Evaluating the designer, or how they used AI, belongs in performance conversations, in her view.

Instead of "why did you design it this way?", she suggests questions such as these:

- What logic, architecture and reasoning sit behind the work? - Where was there friction with the model or the tools, and what workarounds were found? - What did we learn about how the model behaves and how users behave? - Which user or business insights set the direction, where are the gaps, and was the data put to good use? - Where did you take the wheel and where did you let the car run? - How do we know it's right, and what was checked? - What's next, and how will the car be driven differently on the next lap?

Dean cites others who have already picked up on the shift. Eric Chung says traditional critiques were designed for a world where generating ideas was expensive, and that the designer's role is now moving from creating concepts to evaluating, refining and selecting them. According to the piece, Chung has changed how he runs the session: he starts by asking the designer to explain how AI fit into their workflow, what it gave them, what constraints they included, what context was missing, and which concepts they discarded and why. Itamar Medeiros, for his part, argues that AI's polished output pulls feedback toward surface-level details, and proposes roles such as "Prompt Reviewers" to keep the conversation focused on intent.

The author's second argument is that design no longer lives only on the lap but also in the "garage." In a race, much of the outcome is decided before the lights go out: wing angles, tyres, fuel, strategy. In the same way, the briefs, constraints, rubrics, brand checks and design-system rules handed to a model are the "setup." A crit that only reviews the lap misses the garage, so the setup should be critiqued too: what the model was told about the context, what it was told never to do, who agreed on those limits, and whether anyone can change them on the fly.

She also warns of the risk of not feeling the limit: when the car is this fast, a small lapse is costly, such as approving a plausible design at full speed or a confident answer that nobody checked. The problem, she says, is not speed but the gap between speed and the driver's judgment. She also cites Tanya Donska, who warns that if you delegate critique to an AI, it tends to tell you you're right. According to the piece, Donska describes how the same design received opposite opinions depending on how the prompt was framed, and how a feature launched with a 29% completion rate after a prior AI review had given it the green light. Dean concludes that the driver needs a good engineer to tell them the tyres are gone.

The author does not see any of this as bad news for designers. Sensitivity, taste, empathy and the ability to sense a limit before reaching it are qualities they have spent entire careers building, and they are exactly what that very fast car lacks. She proposes learning the new things as well: the setup, the telemetry, and handling a car with a mind of its own.

To pin down what makes a great "driver" in AI-assisted design, she lists six parallels:

- Feeling the limit: noticing when the output stops being reliable, such as a plausible layout that goes wrong or copy that starts to drift, before it causes a problem. - Giving engineers useful feedback: "understeer in turn four" is worth more than "I'm not convinced." In design, that means explaining what's wrong precisely enough to turn it into a constraint, a design-system rule or a code change. - Managing the tyres: saving full judgment for the important decisions instead of spending it on every variant. Attention is a limited resource. - Consistency lap after lap: holding the same standard on variant forty as on the first, which is hard. - Reading the conditions: models and tools change every few weeks, and those who keep adjusting will outpace those who learned a workflow in the spring and stopped there. - Knowing when to take the wheel: deciding when to let the car run, when to override its suggestions, and when to stop generating and decide.

None of these skills shows up in the finished design; they appear in the telemetry, which is why the debrief matters. The article closes by asking readers how they have seen this ritual evolve and what they think it needs now.

In the comments, a reader signing as Tash A notes that critiques go round in circles when the room lacks a shared vocabulary, and that "I'm not convinced by this" only becomes actionable if someone can name the problem, for example a response that streams with no recovery path. She mentions a UX-for-AI glossary she has created herself. This is a reader's recommendation, not a claim made in the article.

This is an opinion piece, based on the author's personal experience and the F1 metaphor. It presents no original data or studies, and the figures or positions it mentions come from the third parties cited.