Identifying at-risk students? A conversation, not a label

How to identify at-risk students? Not with a risk list. Cohort patterns by rubric criterion, and individual signals that go to a tutor — never onto a label.

A common answer to how to identify at-risk students is a risk score built from attendance, grades and activity data, with a list of names at the end. NUADU gives a different answer. It shows cohort patterns by rubric criterion, and where a signal appears about one learner, that signal goes to a tutor for a conversation. NUADU does not produce automated predictions or labels about named learners. For programme directors and the DPOs who review them, the one thing to remember: a signal is for a conversation; a label is a verdict.

How do you identify at-risk students?

Start one level up, with the cohort. Patterns by rubric criterion show where teaching is not landing for a group. Where a signal appears about an individual, it goes to a tutor with drill-down to the work that prompted it, and the tutor talks to the learner. Nobody receives a risk score about a named learner.

Five steps keep that mechanism honest, and NUADU's assessment analytics follow them:

  1. Aggregated by default. Programme- and cohort-level analytics are the default view, with patterns by rubric criterion, by standard and by item.
  2. A signal, not a score. Where a signal appears at individual level, it is framed for human conversation — tutor outreach, support resourcing — never as a prediction or a label.
  3. The work comes with it. Every figure can be traced to the rubric criterion, the item, the marker and the candidate's response, so the tutor opens a piece of work rather than a category.
  4. A person decides. Educators decide what to do with what the data shows. There is no automated routing, no automated labelling, no automated escalation of named learners.
  5. Resourcing stays human. Where it is switched on, the Cohort indicator assistant sends cohort indicators to programme leads and tutors; the resourcing call is theirs.

A list of names is quicker to produce. It is also the step at which a pattern becomes a verdict about a person before anyone has spoken to them. The teacher's side of the same rule — tracking student progress by rubric criterion — starts from the class pattern for the same reason.

Individual signals go to a tutor — never onto a label.

What goes wrong when predictive analytics labels students?

Three things, and NUADU's Report page names them. Predicted outcomes become labels on learners. Teaching narrows to what gets measured. And the labels create self-fulfilling prophecies that disproportionately affect learners who were already underserved. None of this needs bad intent — only a dashboard that renders an inference as a fact.

Predictive analytics in higher education is the use of historical student data and statistical models to estimate which current learners are likely to struggle, fail or leave. NUADU does not build it for named learners. Its cohort indicators serve programme resourcing and are not individual forecasts about named learners.

Most analytics dashboards confuse signals for verdicts. A learner misses two deadlines and a colour changes next to their name. A score dips and a category appears. Nobody decided anything, yet the next person to open the page reads a decision, and treats the learner accordingly.

A self-fulfilling prophecy in education is an expectation about a learner that changes how the learner is taught or treated, so that the expected result becomes more likely. That is how a prediction starts to fulfil itself: a learner read as likely to fail may be offered less stretch, judged against lower expectations or steered away from a demanding option, and the outcome then appears to confirm the forecast. The forecast was drawn from past students; the consequence lands on one present learner.

The cost for a programme director is practical: a label replaces a conversation that should have happened. The cost for the institution is slower to show — teaching shaped around the indicator rather than around the learning.

What are the ethical issues in learning analytics?

The recurring ones — the questions behind any ethical use of student data — are transparency, consent, data minimisation, bias, and the moment an inference about a named person is treated as a finding. That last one is where design decides the outcome, because it is also where data protection law starts asking questions.

Article 4(4) of the GDPR treats as profiling any form of automated processing that uses personal data to evaluate aspects of a person — analysing or predicting their behaviour or reliability among them. Article 22(1) then gives a person the right not to be subject to decisions made solely by automated means, profiling included, where those decisions have legal or similarly significant effects on them. A risk score about a named learner raises both questions, and a DPO will ask them before anything is switched on.

NUADU's design avoids the score rather than defending it: no automated predictions or labels about named learners, and no automated decisions about a named learner. Where AI contributes to the analytics, it surfaces patterns by rubric criterion, by standard and by item — it does not classify learners, and educators decide what to do with what the data shows. Equity breakdowns are configurable per institution, and bias review of any AI components contributing to the analytics is included in the documentation pack on request. The data residency and DPA page lists data residency options and the sub-processors, and explains how to request a signable DPA.

What does NUADU build instead of a risk list?

Two workflow assistants and one AI capability, each working above the individual. The Cohort indicator never creates a profile of an individual. The Programme reviewer identifies where intended learning outcomes are being met and where they aren't, for academic committees, and never ranks or profiles individuals. Item analysis and cohort indicators never label individual learners or predict their outcomes.

All three are opt-in, supervised by a named educator and logged end to end, and an institution can run any of them, or none. Which to switch on is a governance decision in its own right: responsible AI adoption in higher education walks through it, and the page on AI in assessment lists what the human decides and what each module never does.

On the learner's side there is the result, the feedback and — only if the learner opts in — a practice route drafted from approved content and an optional study companion the learner controls (see what an AI tutor should be allowed to see), bounded to their own recent work and approved content, and switchable off. Neither is a profile. The Enterprise page applies the same rule to people-managers: individual signals framed for development conversations — never as automated predictions.

→ Talk to us about your assessment workflow (30 min)