Polished answers are cheap. Lived experience still has texture.

For decades, the technical interview ran on a quiet assumption: a good answer was hard to produce, so producing one was a signal. Ask a candidate to explain how they would design a rate limiter, or walk through a tricky concurrency bug, and the quality of the answer told you something real, because shallow knowledge could not fake depth on the spot.

That assumption is gone. A polished, correct-looking answer to almost any generic technical question is now free and instant. This is not a moral panic about cheating. It is a simple observation about what has become cheap. When the thing you were measuring becomes abundant, your measurement stops measuring anything.

What stayed scarce

Here is the useful part. AI made one thing cheap, but it left something else expensive, and the expensive thing is exactly what you want to hire for. AI is excellent at the clean summary of how something is supposed to work. It is weak at the specific, messy trajectory of having actually done it.

Real work has texture that no summary contains. The things a practitioner actually remembers are specific:

• The approach that seemed right and was abandoned three days in.
• The constraint that arrived from a different team and forced a redesign.
• The trade-off that everyone disagreed about, and how it was settled.
• The thing that worked in staging and fell over in production, and the specific reason why.

This texture is not in any training set, because it belongs to one person's particular history. It either happened to you or it did not.

What this looks like in practice

Suppose a candidate's CV says: led the migration from a nightly batch ETL to a real-time streaming pipeline. The old interview asks: how does stream processing work? The answer is now free, and tells you nothing. The claim-anchored interview asks something different.

Walk me through the first production incident after the migration. What broke, how did you notice, and what did you change?

Then it perturbs the scenario: now assume downstream consumers need exactly-once delivery, but the source occasionally sends duplicates. What changes? A candidate who lived this reaches for idempotency, deduplication, watermarks, event time versus processing time, replay strategy, the specific monitoring they wished they had set up sooner. A candidate who did not lived it names some tools and stalls, because there is no remembered path to walk down.

The difference is immediate and it does not depend on catching anyone doing anything. The strong candidate is not penalised for using AI to prepare. The weak candidate is not flagged as a fraud. The interview simply asks for texture, and texture is either there or it is not.

Why anchoring matters

The anchor is the whole trick. A generic hard question can still be answered generically. A question tied to a specific claim the candidate made, in their own words, cannot, because the follow-ups demand details that only the real history contains. Each perturbation pushes one layer deeper into a path that a real practitioner remembers and a reciter has to invent on the spot.

This is why the move is not to make questions harder. Harder generic questions just have harder generic answers, equally free. The move is to make questions specific to the person in front of you, and to probe the reasoning rather than the recall. Depth that is anchored to a real claim is the one thing that stayed expensive when everything else got cheap.

The static-question interview is not broken because candidates became dishonest. It is broken because good answers stopped being scarce. Stop testing for the answer. Test for the path. The path is where the signal went.