AI Transparency Statement

The provider-by-provider detail, including retention, is on the Subprocessors page.

• Participants are told before the interview begins that they are speaking to an AI interviewer. It is stated on the screen they read before they start, and it is recorded on the interview record which version of that text they were shown.
• The interviewer’s voice is synthetic. It is not a recording of a person and does not imitate any identifiable individual.
• Where the animated interviewer is used, the face is generated. It is not a real person, not a likeness of one, and the Interview Privacy Notice says so in terms.
• A customer is contractually prohibited from representing to a participant that they are speaking to a human being — clause 7.4 of the Subscription Agreement.

• There is no scoring model, and no field on any record that holds a score about a person.
• The artefacts describe work — projects, risks, dependencies, what a successor needs to know. They are not an assessment of the person who described it.
• Customers are contractually prohibited from using the output to make or materially inform a decision with legal or similarly significant effects — dismissal, discipline, promotion, pay, or the content of a reference.
• Declining an interview is recorded as a decline and nothing else. No reason is requested and none is stored.

The system does not infer emotional state, and no emotion, sentiment or stress signal is produced, stored or reported. There is exactly one exception, it exists for safety, and it is described here in full rather than left to be discovered.

The crisis safeguard

An interview about leaving a job can occasionally touch something serious. If a participant says something indicating self-harm, suicide or intent to harm another person, the system stops interviewing. It sets aside the topic, responds supportively, and surfaces the support contacts the employer has configured. The person outranks the interview.

Two mechanisms sit behind that, and they are different in kind:

• A deterministic phrase recogniser. A fixed list of explicit statements — literally, the words a person uses when they say this. It is pattern matching, not inference: it does not read tone, does not read voice, does not model mood, and cannot conclude anything about someone who has not said one of those things outright. It is tuned deliberately to fire rather than to be precise, because the cost of missing this is not symmetrical with the cost of an awkward moment.
• The model’s own response. The language model is instructed to respond to acute distress with support rather than with the next question, and it marks such a reply internally so the system can react.

What is recorded, and who sees it

When this happens, the interview record stores that it happened — a tier and a timestamp — and never what was said. The flag is visible to the manager and HR only. It is not shown to a viewer, never reaches the successor’s handover guide, and is not added to the searchable archive.

Article 5(1)(f) of the EU AI Act prohibits AI systems that infer the emotions of a person in the workplace, subject to an exception for medical or safety reasons. Our position is that a fixed phrase recogniser is not an inference system at all, and that a safeguard against self-harm falls squarely within the safety exception. That position is one we have reasoned about rather than assumed, and we will say so plainly to any customer or regulator who asks.

• Customer content is never used to train a model. Not ours, not a provider’s. Interview transcripts, summaries and handover documents are not training data, are not fine-tuning data, and are not evaluation data.
• The models used are general-purpose models supplied by the providers named in section 1. We do not train our own.
• The interview behaviour is shaped by prompts and by deterministic rules in our own code, not by learning from customer conversations.
• Improvements come from reading our own aggregate operational metrics and from testing against material we generate ourselves, not from mining customer interviews.

A language model summarising a conversation can be wrong. It can miss something, misread what someone meant, or state something with more confidence than the transcript supports. That is a property of the technology, and we would rather say it here than bury it in a warranty disclaimer.

What we do about it:

• Artefacts are generated from the transcript, and the transcript is kept, so anything in a summary can be checked against what was actually said.
• Where the system produces a recap of an earlier part of a conversation, that recap is checked mechanically against the transcript before it is spoken, and a deterministic fallback is used if it does not hold up. Grounding is a check, not a hope.
• An interview too thin to support a meaningful artefact is marked as such and produces an honest notice, rather than a confident-sounding document built from very little.
• Each generated artefact is produced independently, so one failure does not silently degrade the rest.
• Every interview record stores which version of the prompt produced it, so an output can always be traced to the instructions that generated it.

• A human decides that an interview happens, who is interviewed and what topics matter. The system never initiates one.
• A human decides who may read the result. Access is granted by a person, not inferred by a model.
• The customer is responsible for reviewing output before acting on it or distributing it further — clause 17.3 of the Subscription Agreement.
• Participants receive their own summary, which is the most direct form of oversight available: the person who was interviewed can see what was written.

Our current view is that OakHive is not a high-risk AI system under Annex III of the EU AI Act. Annex III point 4 covers systems used for recruitment, for decisions about promotion or termination, for allocating work, and for monitoring or evaluating performance and behaviour. OakHive does none of those: it captures what someone knows about their work and produces documents about that work, and the contract prohibits using it to inform decisions about the person.

It is subject to the transparency duties in Article 50, because it interacts with people and generates synthetic audio and, optionally, a synthetic face. Those duties are met as described in section 2.

Questions about how the AI in OakHive works, challenges to anything stated here, and requests for the detail a data protection impact assessment needs: privacy@oakhive.ai.

If you are a participant and you think a summary about your work is wrong, you can ask for it to be corrected. The route is through your employer, who can action it directly — see section 7 of the Interview Privacy Notice.