AI 透明度

• The character. Generates the other side of the conversation in real time. It is fictional, represents no real person, and makes no judgement about the participant.
• The assessor. Produces a score against a rubric, and written feedback. This is the one that makes a decision about a person, and the rest of this page is about it.
• The performance director. Chooses the avatar’s expression and gesture so the character does not stare blankly. It reads the character’s own line, not the participant, and it is designed so that a failure produces a neutral face rather than blocking the conversation.

What is decided

A score against the rubric attached to that simulation, and written feedback explaining it.

What it is given

The transcript of that one session, and that rubric. Nothing else. It does not read the participant's other sessions, their profile, their team, their previous scores, or anything the employing organisation holds about them outside the session being assessed. It has no notion of who the person is.

Where it runs

A large language model hosted in Microsoft Azure's Australia East region, on a regional rather than a global deployment, so scoring is performed in Australia. Every provider that supports the platform, with the regions it processes and stores data in, is published on our Subprocessors page.

What it is for, and what it is not

It is a practice aid: a consistent, repeatable read on how a conversation went, available immediately and at any scale. It is not a professional, clinical, psychological or employment assessment, and it is not a measure of anybody's competence at their job. Like any AI system it is not infallible — which is precisely why every result is reviewable, overridable and open to challenge, and why every result the product displays carries that statement, including the ones that went well.

A score should not be used on its own for hiring, promotion, performance management, discipline or termination. Customer organisations decide how results are used, and the controls below exist so that decision always has a person in it.

Under the EU AI Act's transparency rules (Article 50, in force since 2 August 2026) a person must be told when they are interacting with an AI system, which is what the disclosure described below does. Where the platform sits against the Act's high-risk employment category, and why its intended purpose is practice rather than performance evaluation, is set out on our GDPR and the EU AI Act page.

A score with no provenance cannot be explained, defended or contested — so each one is stored alongside the things that produced it:

• Which model deployment and API version generated it.
• Which version of the scoring instructions was used — identified by a content hash, so an edit cannot reuse a version number.
• Which version of the rubric was applied, by the same method.
• Whether the score came from the AI assessor, from a fallback when the assessor was unavailable, or was not generated at all because the organisation had switched scoring off.
• When it happened, and for which session.

Because the model and version are recorded against every result, a change to the underlying model is visible in the record rather than invisible in a deployment: results produced before and after a change are distinguishable, and any result can be traced back to exactly what produced it.

An administrator can override or void a score

With a reason. The override supersedes the AI score everywhere the result is shown, and both the original and the override are kept, so the change is visible rather than silent.

The person assessed can challenge it

From their own result, stating why they think it is wrong. A challenge goes to the administrators who can act on it. The two acts are deliberately separate: contesting a result must never be the same action as changing one, or a participant could simply mark their own work.

An open challenge protects the evidence

While a challenge is open, that session and its transcript are held back from scheduled deletion. Otherwise a retention clock could destroy the evidence for a dispute while the dispute was still running — on a schedule, with nobody deciding to.

Before the conversation starts — not afterwards, and not in a policy nobody opens — every participant is shown this:

• The character you're about to speak with is generated by artificial intelligence. It is not a real person.
• It can be wrong, and it may say things that don't fit the scenario. Treat it as practice, not as advice.
• Your conversation is recorded and assessed automatically, and your feedback score is generated by AI.
• A human can review any score you're given.

Alongside it they are shown a collection notice covering what is collected, why, how long it is kept and who receives it. The version of each notice is recorded against the session, so the question “what was this person actually told?” has an answer that is a record rather than a recollection. The notice text lives in version control precisely so a version identifier cannot be rewritten under a participant who already saw it.

There is no consent tick-box. Under the Australian Privacy Principles this kind of processing is generally handled by notice rather than express consent, a facilitator is usually in the room, and a click on every run costs more completions than it adds in protection. What is recorded is therefore disclosure — this person was shown this version, at this time, and proceeded.

The assessed conversation can be held entirely in text. A participant who cannot or would rather not speak to the animated character types instead, at the same prompt, and is scored on the same rubric by the same model — so the way someone takes part does not change what is assessed. Our Accessibility statement sets out the rest.

AI features are controlled per organisation, and an administrator is told what turning each one off actually costs rather than being handed a switch with no consequence attached. Where an organisation belongs to more than one arrangement, the strictest setting wins.

Any participant can download every automated decision held about them, immediately and without asking, from their account settings — with the model, prompt version, rubric version and source for each. A departing customer's administrator can export the same for the whole organisation. See your privacy rights.

To discuss AI governance requirements, an impact assessment, or how scoring would apply to your own rubrics, contact privacy@evolvesimulations.com.