Teaching a model the language of recovery.

People often imagine an AI system as something you program with rules. SmartRTW doesn't work that way. The model underneath it learns the way a careful reader does, it reads return-to-work material closely, predicts what comes next, checks itself against the source, and corrects every word it gets wrong. The animation above is that loop in miniature.

Prediction is the lesson

The technique is called supervised learning, and the version we lean on most is deceptively simple: given everything in a passage so far, predict the next word. Each token in a document becomes its own tiny exam question and because the correct answer is already sitting there in the text, every guess can be marked the instant it's made. Right answers reinforce what the model already believes; wrong ones become a correction.

That's what makes it supervised: the label comes free with the writing. No one has to hand-tag millions of examples. Run that predict-check-correct cycle across a carefully chosen library of return-to-work content and, word by word, the model internalises how the language of recovery actually fits together; the phrasing of a graded return, the rhythm of a supportive check-in, the structure of a well-formed workplace adjustment.

What we teach it from

A model is only ever as good as what it reads. Ours doesn't learn from the open internet, it learns from a deliberately curated corpus of return-to-work-from-mental-injury material, assembled and vetted with our clinical partners:

From a filing cabinet to a teacher

Raw documents aren't training data. Before anything reaches the model, our psychologists and AI researchers work side by side to turn a filing cabinet of material into something a model can learn from safely. That means de-identifying every record so no individual can be recognised, structuring long documents into clean, self-contained passages, and weighting authoritative, clinically-endorsed sources more heavily than incidental ones.

The goal isn't volume for its own sake. A smaller, expertly curated corpus that reflects genuine best practice will always beat a larger one full of noise, especially in a domain this sensitive, where the wrong tone can do real harm.

“We're not teaching the model to give medical advice. We're teaching it the language and logic of good return-to-work practice, so it can better support the people who do.”

Why this matters for mental-injury recovery

Return-to-work from psychological injury has its own vocabulary, its own pacing and its own emotional weight. A general-purpose model, trained on the internet at large, tends to miss that nuance. It can sound clinical when it should sound warm, or generic when it should be specific. By grounding the model in vetted return-to-work content, its responses stay relevant, evidence-aligned and humane, rather than improvised.

It's the same principle the rest of SmartRTW is built on: an AI that listens between the lines is only possible if it has read the right things first.

The guardrails

Learning from sensitive material carries real responsibility, so the training pipeline is wrapped in safeguards from the start:

• De-identification first. Personal information is removed before content ever enters the corpus.
• Clinicians in the loop. Psychologists curate, review and sign off on what the model learns from.
• No automated clinical decisions. The model supports people and surfaces best practice; it never diagnoses or decides.
• Bias detection. Outputs are checked for the kinds of bias that can creep in from historical data.
• Privacy by design. Data handling is built to the standards expected of a WorkSafe Victoria initiative.

Where this is heading

We're currently in Phase 1: Feasibility & Data Exploration, the groundwork that decides what the model can and should learn. The corpus and methods described here feed directly into Phase 2: AI Engine & Prototyping, where the predict-check-correct loop becomes a working companion for returning workers and the managers who guide them.

The animation at the top of this page loops endlessly, fixing the same sentence again and again. That's the point: learning to support people well isn't a one-off, it's a practice of predicting, checking and correcting, over and over, until the language of recovery becomes second nature.