AI transparency. What the AI does, what it refuses to do, and where it is weak.
Several jurisdictions now require a page like this. It is also the page a careful buyer reads first, so it is written to be useful rather than to be minimally compliant. The limitations section is not padding.
What the system does
Three separate jobs, deliberately kept separate.
| Job | What it sees | What it produces |
|---|---|---|
| Conducting | The conversation so far as text, and the problem | Spoken questions and follow-ups, phase transitions, hints when asked. It does not score anything while the interview is running. |
| Evaluating | The transcript and the rubric’s dimensions. For System Design interviews, also the final whiteboard as text: boxes and labels, connections, written requirements and napkin math. Not your name, email, school, employer, location, photograph, or any demographic field. Not the video. Not the audio. | Evidence quotes with timestamps, then a score per dimension against the anchors, with the reasoning attached. |
| Integrity assessment (planned, not built yet) | Planned: timing and editing patterns, the divergence between what was said and what was produced, and window and environment events the candidate consented to. Nothing of this is collected today. | Planned: a likelihood band with the specific evidence. It will route to a human and will never be an input to the score. |
The hard limits
These are enforced in code and in database permissions, not asserted in a policy document. They are permanent product decisions rather than a roadmap position.
The system never
- Rejects, filters, hides or reorders a candidate. There is no automatic adverse action anywhere in the product
- Performs facial recognition, emotion inference, gaze tracking or attention scoring
- Scores accent, fluency, speech rate, appearance or perceived confidence
- Lets the planned integrity estimate touch the score or the recommendation
- Sees a candidate’s name, or any demographic data, in the component that scores
- Trains any model, ours or a vendor’s, on candidate interviews. Contractual with Anthropic directly, and with the speech providers through Live’s LiveKit agreement
The system always
- Checks every quote against the transcript line it cites, and discards any that does not appear in that line. An evaluation left with too little checked evidence is reported as insufficient signal, with no overall score and no recommendation
- Computes the weighted total in code rather than asking a model for it
- Records which rubric version and which model version produced each report
- Will require, once integrity signals ship, corroboration from independent signal families before raising an integrity band above the lowest two
- Leaves the decision to a named person at the employer
On facial analysis specificallyReading emotion or attention from a face is the feature buyers ask for most often and the one we will not build. Its accuracy for this purpose is poor, its bias risk is the highest in the category, and it is the thing most likely to be regulated out of existence. Refusing it costs a little signal and removes a large liability.
Which models, and what they are given
Named, because "our proprietary AI" is not a transparency statement.
| Stage | Provider | Data sent | Retention at the vendor |
|---|---|---|---|
| Speech to text | Deepgram, through LiveKit Inference | Candidate and interviewer audio | We have no direct agreement with Deepgram, so we make no retention claim for it |
| Conversation | Anthropic, Claude Sonnet | The conversation so far as text, and the problem | Deleted within 30 days under Anthropic’s standard commercial terms, or kept up to 2 years if its safety systems flag it, or longer where the law requires. Not used for training |
| Text to speech | Cartesia, through LiveKit Inference | Interviewer text only | We have no direct agreement with Cartesia, so we make no retention claim for it |
| Evaluation | Anthropic, Claude Sonnet | The complete transcript as text, with no name or contact details sent alongside it, and the rubric’s dimensions. For System Design interviews, also the final whiteboard as text: its boxes and labels, the connections between them, the written requirements and the napkin math. Not the drawing, and not where anything was placed | Deleted within 30 days under Anthropic’s standard commercial terms, or kept up to 2 years if its safety systems flag it, or longer where the law requires. Not used for training |
Enterprise customers can supply their own model credentials, in which case inference runs against their account under their own vendor agreement and their own retention terms. The full sub-processor register is on the security page.
Where it is weak
Every one of these is a real limitation with a real consequence. A transparency page without this section is a marketing page.
Unusual but correct answers
A candidate who solves a problem in a way the anchors did not anticipate can be scored below what a strong human interviewer would give. This is the failure mode we watch most closely. Recording overrides so that the override rate can be reported to you is planned and not built yet.
Silent thinkers
The rubric can only score what is observable. A candidate who reasons brilliantly and says little will score lower on more than one dimension. The interviewer prompts for narration, but the effect does not disappear.
Domain edges
The problem bank is strongest in backend, distributed systems and data. Front end, embedded, machine learning and mobile are thinner, and the anchors were written with backend work in mind.
Integrity signals will be probabilistic
They are planned and not built yet. Every one of them will have an innocent explanation: a fast typist on a second monitor looks like an assisted candidate. This is why corroboration will be required, why nothing will be automatic, and why the language will be likelihood rather than cheating.
English only, for now
Interviews run in English. Candidates code in the language they choose, and the choice is never remarked on, but the conversation itself is not yet multilingual.
It is one signal
A 45-minute interview is a narrow window into a person. Live is designed to replace an inconsistent first round, not a hiring decision, and any use of it as the whole process is a misuse of it.
Human oversight, concretely
| Who decides | A named person at the employer. The tool cannot advance or reject anyone. |
| Reviewed reports | Planned, not built yet: marking a report reviewed as a tracked action with a person’s name on it, and a queue of unreviewed reports. Reports cannot be marked reviewed today. |
| Overrides | Planned, not built yet: a hiring manager disagrees with a score and records why, and the rate of overrides goes back to the customer and to us, per rubric version. Overrides are not stored today. |
| Thin evidence | An evaluation left with too little verified evidence returns insufficient signal instead of a recommendation. The model’s own confidence figure is shown beside each score and does not change the recommendation. |
| Contest route | Documented, and it reaches a person at the employer rather than a support queue here. |
This page is versioned with the product. Material changes to what is assessed are dated here and notified to customers, because a change to what is assessed changes what candidates were told before they consented.