The interview

The interviewer is a language model. It is told what to gather and it writes each question after reading your previous answer. No human reads your answers as they arrive, and nobody is listening live.

We put that sentence at the top rather than in a footnote. A person who discovers it later feels handled, and the whole product depends on them not feeling that. The disclosure appears in the product at the same moment, before anybody says anything.

A list of questions gets you the answer a person has already prepared. Ask somebody to rate a friend on reliability from one to seven and you learn what they think reliability means, which is a fact about them.

A conversation can ask the second question. When somebody says a friend is always there for people, the useful move is to ask when that last happened. The first answer is a summary. The second one has a Tuesday in it, and a Tuesday is something we can check against four other people.

The cost of this is real and we should say it. A conversation is slower to run, harder to test, and impossible to score with a lookup table. It also cannot be compared across people the way a scale can. We took that trade because a scale measures agreement with a word and we want the episode underneath it.

Not the interviewer. That decision is made by code, from the answers that actually exist, and the model is allowed to propose an ending it does not get to take.

This matters more than it sounds. Language models are trained to be agreeable. An agreeable interviewer facing somebody giving short answers will offer to wrap up early, because ending the conversation resolves the discomfort in it. That produces a finished interview with nothing usable inside it, and the failure is invisible: the transcript looks like a real conversation.

So completion is computed. Each thing the conversation has to gather material about needs a set number of answers carrying an actual episode. An answer under twelve words cannot carry one. If a proposed ending arrives before that bar is met, the engine refuses it and asks again.

Every question can be answered by speaking or by writing, and the two count exactly the same. That is not a courtesy. Somebody who cannot use a microphone, or is on a train, or hates the sound of their own voice, has to get the identical product. Text is the path that guarantees it.

When you speak, the audio goes to a transcription service, becomes text, and is deleted within 24 hours. We keep the text, because the text is where the patterns are. We never build a voiceprint, never analyse how you sound, and never use your voice to work out who you are.

The engine that runs the conversation cannot tell which answers were spoken. It receives text either way, and that is deliberate. An engine that knew would be one change away from reasoning about how somebody sounded, which we have promised never to do. The privacy page is the longer version.

About fifteen minutes for your own interview. Roughly ten for somebody answering about you, because they are describing one person rather than themselves.

There is no timer on any answer and no penalty for a long pause. If a question needs you to sit and think, sit and think.