Production is now cheaper than UX evaluation

A shift happened in how design work flows through your team, and it happened faster than the job descriptions caught up. AI made building cheap. It did not make judgment cheap. So now the work shows up already built, and someone on your team has to figure out what to save. Let me catch you up.

The deep cut

• Production got cheap, judgment did not. NNG's line that "production has become cheaper than UX evaluation" reset the whole sequence.
• Automate the chore, keep the lesson it taught. Transcript tagging left the junior queue, and so did learning to hear "it's fine" mean "I gave up."
• Grade the catch, not the polish. Microsoft and CMU found high confidence in the tool predicts less critical thinking.

The repair shop at the end of the line

The old order was understand, define, build, evaluate. Each step had friction, and that friction was a checkpoint. Someone had to stop and ask if the thing made sense before going further. AI erased the friction and kept none of the checkpoints. Now teams build, ship, then evaluate when users complain.

One designer at Kiyo AI walked through a normal week: rewriting AI copy that sounds like every other SaaS product, reverse-engineering a dashboard shipped two weeks ago that now generates support tickets. A v0 dashboard organized by what was logical to display, not what users came to find. A three-hour audit would have caught it. Six weeks of tickets did not.

Plausible is not usable. A prototype that demos well is not a product that works. AI makes very plausible artifacts, which is exactly why teams mistake them for finished ones.

The training ground you are about to pave over

Here is the part that hits your hiring plan. The cheap tasks you are handing to AI, transcript tagging, first-pass wireframes, competitive audits, redline sheets, were never worth much as output. They were worth a lot as practice. Tagging transcripts is how a researcher learns that "it's fine" means someone gave up. Marking redlines is how a new designer learns which spacing the system decided and which is open.

The numbers back the worry. Erik Brynjolfsson and team found employment for 22-to-25-year-olds in AI-exposed jobs sitting 19% below trend, with no gap for experienced workers. The adjustment runs through hiring, not layoffs. AI substitutes for codified knowledge and complements the tacit kind, the kind built by doing. A lot of juniors are not doing much doing right now.

What the new work actually is

The craft moving forward is evaluation. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers on 936 tasks and found two things pulling apart. Confidence in the tool predicts less critical thinking. Confidence in yourself predicts more. A junior has the least self-confidence and the most exposure to polished output. Left alone, they defer, and deference looks like competence until it doesn't.

So make the critique the deliverable. Not "summarize these eight interviews," but "here is the synthesis the tool made, find the two claims the transcripts don't support, bring the timestamps." One audit of a Claude Code page found it overrode heading classes with inline sizes, misused a table-row token for active nav, and invented a font size, all while looking system-compliant. Familiar tokens disarm normal review. That is the exact skill you now need to train on purpose.

Skip the deliverable, ship the answer

While your team cleans up plausible messes, the smart move on the product side is to stop shipping tools nobody asked to operate. Luke Wroblewski built an AI newsroom, presented it as a news site, then watched a "compile me a personalized report" feature become how people actually used it. Outcome first, tool second.

Disclosure is its own tradeoff. NNG's research roundup found no single reaction to telling users AI helped, but a clear "disclosure paradox": people say they want it disclosed, then rate the content lower when it is. They also trust a company caught hiding AI use least of all. If it can surface later, say it first.

Three questions for your team

• Which tasks left the junior queue in the last 18 months, and what did each one teach the person doing it? List both columns before you automate the next one.
• Who owns the evaluation set for our biggest AI feature: the 30 real inputs, the accepted output, the failure categories, and the call to ship or hold?
• In our next review, are we grading the polish of AI output or the catch, the two claims the transcripts don't support?