AI made the PM take-home a weak signal. The test is moving into the room.
Only 26% of candidates trust AI to judge them fairly, and hiring teams trust a polished PRD even less. The fix showing up in 2026 isn't better detection. It's watching candidates work live.
Gartner surveyed 2,918 job candidates in early 2025. Only 26% trust AI to evaluate them fairly. An earlier Gartner survey found 39% of candidates already use AI somewhere in the application process, mostly for résumés and cover letters. And in a third survey of 3,000 candidates, 6% admitted to interview fraud outright, either posing as someone else or having someone else interview for them.
So trust broke on both sides of the table at once. Candidates don't believe the screening is fair. Hiring teams don't believe what they're reading is real. Product management got hit hard here. A polished PRD, a clean case study, a confident answer to "walk me through a time you prioritized." All of it is now free to generate. None of it ever required product judgment to produce. It just used to correlate with it.
Product People, a firm that places interim product managers, said it out loud in April. Their candidate pool "exploded" while the signal-to-noise ratio "plummeted." Their fix was to cut a three-part take-home down to one focused piece and move the rest of the evaluation into live simulations. Gartner's own advice to recruiters points the same way: anti-cheating safeguards, including in-person interviews.
I recognize this from the other side of the QA table. Running QA on LLM training data, a rubric only worked as long as passing it required the skill it was meant to measure. Once a cheap shortcut could clear the rubric without the skill, the rubric was measuring the shortcut. An unsupervised take-home is a rubric with no shortcut protection left. It never really had any. AI just made the gap impossible to ignore.
The fix showing up in 2026 loops doesn't try to detect the shortcut. It removes the room for one. Aakash Gupta's 2026 interview guide says PMs are now being asked to build a working prototype in 45 minutes with tools like Cursor, Bolt or Lovable, live, with an interviewer watching. Which companies run this round, and how often, is still mostly secondhand, so I wouldn't treat it as standard yet. But the direction is clear.
That's a different test than "can you write a PRD." It's closer to "can you direct a model the way you'd direct a report." That's the exact motion behind RolesRadar, the job-search agent I built on n8n and Gemini. You don't get a good result from a vague ask. You break the task down, spell out the constraints, judge what comes back, and iterate when it's wrong. That loop is hard to fake live, because faking it live is the same work as doing it.
That said, the live round won't stay ungameable forever. Someone will build a tool that whispers the next prompt mid-interview, the same arms race that killed the take-home. The part worth keeping isn't the format. It's the principle. A test only works as a signal as long as passing it requires the thing you're trying to measure.
It's the same discipline evals brought to shipping AI features. Define what you're testing for before you write the test. If the artifact can be produced without the skill, the artifact was never the requirement.
The skill was. AI just took away the place it was hiding.