Human in the loop AI in education: a named educator decides
Human in the loop AI in education, made checkable: eight AI capabilities and seven assistants, each naming who decides and what the AI never does.
Human in the loop AI in education should mean a named educator, not a slogan: on NUADU there is one at every decision point, and the AI prepares the evidence. The phrase usually describes a person who reviews an AI output before it takes effect, which is easy to promise and hard to check. So here is the check: NUADU's eight capabilities and seven workflow assistants, each with the role that decides and the one thing the AI never does, written down for a head of assessment or a compliance lead to test.
Human in the loop AI in education: what does it mean?
A named educator who decides, and a record that shows it. The site puts it in three short sentences: AI prepares the evidence. A named educator decides. Every step is logged. A loop without a name in it is a diagram; a loop with a name in it is a decision someone can be asked about later.
The name matters because AI in assessment rarely arrives as one feature. Most platforms market “AI everywhere.” Institutions need AI they can defend. NUADU's answer is to unbundle it into small, named, task-bounded modules. Each does a single thing, and each feeds evidence into a step that a named educator owns. That is what makes the human visible: if a module cannot say who decides, there is no loop to inspect, only an output.
None of them decide. None of them act on a named learner.
Where is the human in each of the eight capabilities?
In every one, by role: the educator who approves an item, the marker who decides a grade, the invigilator who reviews a signal, the educators who keep programme decisions. In each line below the human decision comes first and the AI's hard limit second.
- Content Scanner — drafts candidate items and a draft rubric from your materials. Human: approve, edit or reject every drafted item. Never: publish an item.
- AI-assisted marking, written — applies your rubric across written responses and shows where it is less confident. Human: the educator decides the grade. Never: override the marker's sign-off.
- AI-assisted marking, spoken — transcribes responses and applies your speaking rubric, CEFR-referenced where applicable. Human: replay, review the transcript, adjust scores, sign off. Never: finalise a speaking grade.
- Pronunciation feedback — phoneme-level feedback during practice. Human: teachers decide whether to act on it. Never: used in high-stakes scoring.
- Long-form review — structural signals on long-form writing at draft stage. Human: output goes to the supervising educator. Never: a substitute for thesis examination.
- Resource search — semantic search across approved learning content. Human: the educator decides what to assign. Never: profile the learner.
- Integrity signals — layered and proportionate to exam stakes. Human: a named invigilator reviews anomalies. Never: treat signals as findings.
- Item analysis and cohort indicators — item statistics and aggregated patterns. Human: programme decisions stay with educators. Never: label individual learners or predict their outcomes.
Four of the eight have live demos, among them the live demo of the Content Scanner and the live demo of written marking. Marking is where the loop is tested hardest, and AI-assisted marking with the marker signing off follows one script from draft score to signed result.
Who decides in the seven workflow assistants?
A named person at the end of each one. Workflow assistants prepare evidence for a human decision. They do not decide. They do not act on a learner. Every output is reviewable, every override is logged. Read each line as trigger, destination, decision.
- Marking assistant — exam submission; to the marker, borderline cases and high rubric-criterion variance first; the educator signs off, and no mark on a flagged case is released without the marker.
- Standards mapper — item creation or import; to the author or QA reviewer; a named human approves the proposed alignment to Bloom's, CEFR, EQF or your own framework.
- Feedback assembler — marks signed off; to the marker, who edits the rubric-anchored draft before release; no criteria are invented outside the rubric.
- Study-path drafter — result released; to the learner, opt-in; following it is the learner's choice, and steps are never mandatory.
- Optional study companion — learner opt-in; limited to the learner's recent work and approved content, and on only when the learner chooses.
- Cohort indicator — scheduled or on demand; to programme leads and tutors for a human resourcing call; never a profile of an individual.
- Programme reviewer — the review cycle; to academic committees for curriculum and programme decisions; it never ranks or profiles individuals.
The pattern is the point. Each line names who receives the output and who makes the call, and that is the shape a reviewer can verify.
How do you check a human-in-the-loop claim?
Ask for one sentence per module: what it does, which named role decides (educator, marker or invigilator), what it never does. NUADU's AI page sets the bar as AI you can describe in one sentence to a procurement reviewer. If a vendor cannot write the sentence, the human in that loop is a hope, not a named educator.
Then test three things the sentence implies. Approval: does every output that reaches a learner carry a named educator's sign-off, with each override logged? Limits: is there any grade, flag or credential the AI settles on its own? On NUADU there is none, and no automated decision is taken about a named learner. Rubric: do the modules apply your rubric, or a generic model's opinion? The last test is the line between an assessment module and a general chat assistant asked to grade essays.
The switch belongs to the check as well. Every module is opt-in, and each can be enabled per assessment, per programme or per institution, or left off entirely; responsible AI adoption in higher education, one module at a time, sets out the order. The paperwork belongs to it too: the documentation pack — methodology, validity evidence, bias review, retention, data flows, sub-processors — is available for institutional review, and the AI vendor due diligence checklist covers what to request alongside it. The AI in assessment page keeps the full list in one place.
Frequently asked questions
What are examples of human-in-the-loop AI?
In assessment, three are typical: an item-drafting module whose every candidate item a named educator approves, edits or rejects; marking assistance where the marker decides the grade and signs off; and integrity signals that a named invigilator reviews after the fact, with the institution's own process deciding.
What is a human-in-the-loop AI system?
One in which an AI output cannot take effect until a named educator, marker or invigilator has reviewed it. On NUADU that person approves, edits or overrides, and the record keeps their name beside the change.
Can AI make decisions?
Not on NUADU. A module can draft, propose, surface evidence and apply your rubric; the decision belongs to a named educator, and every output is logged and reviewable on request.