The AI behind OEZ
Anyone can point a chatbot at a role-play card. Building an AI that behaves like a trained OET interlocutor — and marks like a trained OET assessor — turned out to be a very different problem.
Three problems, in ascending order of difficulty
When we started building OEZ, we assumed the hard part would be the conversation. We were wrong three times.
1. A patient, not an assistant
Large language models are trained to be helpful — and a helpful patient is a useless one. A real OET interlocutor answers briefly, waits to be asked, drops half-cues (“…it's just, with the flight coming up…”) and only surfaces their real concern if the candidate earns it. Getting a model to withhold, to be ordinary, evasive, worried and human, took a long series of behavioural constraints and a lot of recorded sessions that told us exactly where the illusion broke. The current AI patient is built on everything those sessions taught us — including the things we only learned by watching it fail.
2. Marking that matches the descriptors
The second problem was the one that consumed most of the effort: an unconstrained model is a generous marker. It rewards fluent-sounding answers, forgives what a trained assessor would not, and drifts from session to session. OEZ's scoring engine is criterion-referenced against the official OET descriptors — every band must be justified with evidence quoted from the candidate's own words, and the engine is calibrated so that the same performance earns the same marks tomorrow. How exactly we constrain it is our own recipe, refined through more trial and error than we'd like to admit; what we can say is that every one of the nine speaking criteria and six writing criteria went through its own cycle of testing against performances where we knew what the mark should be, and the engine wasn't released until it agreed.
3. Feedback that teaches, not describes
“Work on your fluency” is a horoscope, not feedback. The last problem was making every report answer three questions: what exactly happened, what would the better version have sounded like, and what do I drill tomorrow? That's why OEZ feedback quotes the moment you missed, writes the sentence that would have earned the band, and ends with a drill — a report format that went through as many revisions as the scoring itself.
The measurement layer
Language models are poor judges of sound — so we don't ask ours to guess. OEZ measures your delivery from the audio itself: words per minute against the natural conversational range, pausing inside your own turns, fillers per hundred words, false starts, lexical range. The linguistic criteria are graded on measured signals, not vibes — and in writing, your letter is checked fact by fact against the case notes, the way a marker with the notes open actually works.
What we deliberately don't do
- No invented encouragement. If a performance is a C+, the report says C+ and shows why — inflated practice scores are a cruel way to prepare someone for exam day.
- No recycled official material. Every case is original, written in the authentic OET format. You get the exam's shape, never its copyrighted content.
- No black-box grades. Every band arrives with the evidence that produced it. If you disagree with a mark, you can see exactly what it was based on.
Still improving, every week
Every session teaches us something — a phrasing the patient should have resisted, an edge case the marker should weigh differently. The engine you practise with this month is measurably sharper than last month's, and the gap between an OEZ report and what a trained assessor would say keeps narrowing. That is the whole project: not an AI that chats about the OET, but one that gives it.
The best way to evaluate it is a session of your own: start free — a full role-play or letter with the complete report, no card needed.
Practise this on OEZ — start freeFull speaking role-plays with an AI patient, timed writing tasks, and marking modelled on OET’s published criteria. No card needed.