When Numbers Start Making Decisions · Season Three, “The People the Average Does Not See” · Article 9
1. A risk score can be both a gateway and a label
When a person claims income support, the Job Seeker Snapshot collects information about work history, education, health, housing, transport and other circumstances. Some answers contribute to the Job Seeker Classification Instrument, or JSCI. The score estimates relative risk of long-term unemployment and helps determine servicing eligibility under Workforce Australia.
The Department of Employment and Workplace Relations describes the Snapshot and JSCI as core parts of the Job Seeker Assessment Framework. It also says the information can inform service choice and should be updated when circumstances change. DEWR: Job Seeker Assessment Framework
The model has a benevolent purpose: people facing greater barriers should receive more intensive or appropriate assistance. Yet the same prediction can channel a person into more provider oversight, more assessment and less control over how they seek work. Sensitive disclosures can increase support while also expanding administrative knowledge and obligations.
The issue is not simply whether the model predicts accurately. It is whether a forecast about a group is used to enlarge a person’s capabilities or to narrow their autonomy.
2. JSCI predicts relative difficulty; it does not discover motivation
The JSCI uses associations between characteristics and employment outcomes. Earlier official evaluation explains that a logistic regression model estimates relative weights for factors linked to long-term unemployment. The score quantifies expected labour-market disadvantage and has historically helped determine a service stream. DEWR: Online JSCI Trial Evaluation Report
A statistical association is not an explanation of one person’s future. Age, health, location or time out of work can correlate with longer unemployment across a population without proving that an individual lacks effort or capacity. Labour demand, discrimination, caring duties, available jobs and provider quality also influence outcomes.
This distinction matters because administrative systems can moralise prediction. A high score may be heard as “hard to employ”; a low score as “needs no help”. Neither follows. The score estimates service difficulty from recorded factors. It does not measure character, willingness or the complete opportunity structure surrounding the person.
3. More information can create a disclosure dilemma
To obtain support, a job seeker may need to disclose disability, mental health, unstable housing, family circumstances, criminal history or transport barriers. Accurate disclosure can increase the JSCI score and prompt assessment or provider assistance. It can also expose highly sensitive information to a system connected to income support and compliance.
People may reasonably fear stigma, loss of privacy or greater control. If they withhold information, the model can underestimate need. If the questionnaire is completed quickly, with unclear wording or without trust, the resulting score may appear precise while the inputs are incomplete.
Consent in this setting is constrained. A person applying for payment cannot simply walk away from the process like an ordinary customer. Institutions therefore carry a higher duty to explain why each category is collected, who can access it, how it affects service and how it can be corrected.
Data minimisation should remain active even where collection is lawful. A system should not retain sensitive detail merely because it may be predictive. The value of the variable must be weighed against privacy, chilling effects and possible misuse.
4. Support and control must be separated where possible
Targeting greater resources to people likely to face barriers is defensible. Tailored assistance, professional assessment, training, workplace adaptation and a stable relationship with a capable provider can expand practical freedom.
But intensity can also mean more appointments, activities and monitoring. If a high predicted risk automatically produces stricter obligations, the model turns disadvantage into greater administrative burden. A person is penalised for characteristics that were supposed to justify support.
Program design should distinguish the supportive and coercive effects of classification. Additional services should be offered with meaningful choice wherever possible. Compliance consequences should require evidence about actual obligations and conduct, not follow from a predicted difficulty finding work.
The parliamentary inquiry into Workforce Australia criticised the gateway to services and examined problems with assessment and streaming. House Select Committee: Gateway to services That scrutiny reflects a central principle: an allocation tool should be judged by the experience and outcomes it creates, not only its statistical performance.
5. Circumstances change faster than a durable label
Employment risk is dynamic. A person can lose housing, develop a health condition, obtain reliable transport, complete training or move to a stronger labour market. The Department says the Snapshot should be updated when circumstances change and can be updated by the individual, a provider or Services Australia.
Update rights are essential, but they need practical access. A person must know which information is held, which parts can be changed and whether a correction affects servicing. Automatic updates, such as age, should not imply that other circumstances are current.
Versioning matters at system level. Models are re-estimated as labour markets and employment programs change. A factor associated with outcomes in one period may behave differently after an economic shock or policy reform. If the score remains unexplained, neither participant nor frontline worker can tell whether it is stale.
A classification should therefore carry a date, model version and review trigger. It should expire or be reassessed rather than becoming an enduring administrative identity.
6. Accuracy must include downstream outcomes
A model can predict long-term unemployment accurately and still fail as a service tool. If high-score participants are referred to low-quality or unsuitable support, the classification has not achieved its public purpose. Worse, the prediction can become self-fulfilling if a label lowers expectations or diverts a person from opportunities.
Evaluation should compare more than predictive accuracy. Did people receive useful services sooner? Did employment become more stable? Were participant choices respected? Did sensitive groups experience more sanctions or less control? Which incorrect classifications were reviewed and reversed?
The government’s employment services reform material indicates that the system itself remains under review and redesign. DEWR: Employment services reform That makes transparent outcome evaluation more important, not less.
Distributional audit should examine false negatives—people who needed support but received little—and false positives—people channelled into intensive service they did not need. The harms are different and should not be averaged into one accuracy figure.
7. A practical framework for predictive service allocation
When a risk model assigns public services, ask:
- What outcome is predicted, over what period and for which population?
- Which answers are sensitive, and why is each needed?
- Can the person understand, access and correct the relevant data?
- Does a higher score add support, obligations or both?
- Can the person choose among suitable service routes?
- Are compliance decisions kept separate from predicted disadvantage?
- When must the assessment be updated or allowed to expire?
- Are employment, autonomy, privacy and distributional outcomes audited?
These questions treat the score as part of a relationship between citizen and institution, not a neutral object floating outside it.
Conclusion: predicted difficulty should create support, not destiny
The JSCI addresses a real allocation problem. Uniform employment services can underserve people facing complex barriers and waste resources on people who need little assistance. A structured assessment can make hidden needs visible.
My judgement is that the classification is legitimate only when greater predicted disadvantage reliably opens proportionate, useful support; when sensitive data are explained and correctable; and when the score does not automatically justify greater coercion. People must be able to update circumstances and challenge the service consequence, while outcomes feed back into model reform.
A model may estimate who is at risk of long-term unemployment. It must not turn that estimate into the reason a person loses autonomy or remains unemployed.
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.