An interview that rewards memory is testing the skill AI is making cheaper.
The harder question is whether the candidate notices when the machine is wrong.
I have run interviews across different teams, countries and levels, and the weakest signal has often been polished recall without operational judgment. A candidate can know the pattern and still miss the broken assumption.
Coinbase has published what happened when that stopped being hypothetical. AI-generated code went from 5.7% of everything merged in Q1 2025 to crossing 50% for the first time in Q4. Human review stayed at roughly 100%. So they rebuilt the interview loop: let candidates use AI, then watch how they frame the task, inspect the repository, challenge the output, and recover when the model takes the wrong path.
That is much closer to modern engineering work than reproducing an algorithm from memory. Foundations still matter, but they matter as tools for detecting error — not as a performance ritual.
The interview should resemble the job: direct the system, inspect the evidence, own the final decision.