Coursework authenticated on evidence of how it was made, not on a detector's guess. Teachers get a timeline and the right questions to ask, and a person decides every concern through a fair process.
At a glance
An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.
Every year the college's teachers sign declarations that thousands of pieces of coursework are the students' own. Generative AI makes that harder to stand behind. Teachers see work that doesn't sound like the student, references that don't exist, or a whole essay that appears in one evening. They want a quick, objective answer, and AI detectors promise one: paste in the work, get a percentage.
That percentage is not evidence on its own. Detectors estimate how predictable text is, and they produce both false positives and false negatives. A peer-reviewed study published in 2023 tested several detectors on essays written for an English proficiency test by non-native speakers, and on average they flagged more than 60% of those human-written essays as AI-generated. Students who write simply, who have learned formal templates, or who use assistive technology can look 'machine-like' too. Light editing of AI text often gets past the same tools. JCQ's AI guidance, effective from September 2026, says detection tools must only be part of a holistic judgement and must not be relied on in isolation.
The cost of getting this wrong falls on students. An allegation of malpractice is stressful, and a finding can cost a student a whole qualification. A process that leans on detector scores will make more accusations against some groups than others, and it won't stand up to the student's right to see and answer the evidence. The college still has to authenticate the work, so it needs a better approach, not just a better tool.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Share of coursework with retained drafts or version history and at least one recorded conversation with the student.
ExampleEvery piece, by the authentication deadline
Share of authenticity concerns settled through the college's own process before students sign the declaration, as JCQ allows.
ExampleMost concerns, reviewed each term
Share of concerns that, on review, were not upheld. A high share means something is flagging too readily.
ExampleTracked termly, with every case reviewed
Concerns raised and upheld, split by EAL, SEND and access arrangements.
ExampleNo group raised more often without evidence
Time for a teacher to prepare for and hold an authentication conversation.
ExampleAbout 10 minutes
Malpractice reports the awarding body accepts without asking for more.
ExampleEvery report
Cross-sector rules are on the UK regulation map.
Teachers must only accept work that is the student's own. Centres should keep drafts, plans and version history, supervise some work, and hold short discussions with students about their work. Detection tools can produce false positives and false negatives, their accuracy varies with factors including a student's English language competency, and they must only form part of a holistic judgement. The design follows that: process evidence first, any detector output as one input with its limits recorded.
Concerns found before the student signs the declaration are handled by the centre's own procedures. After the declaration, suspected AI misuse must be reported to the awarding body, and the student must be told of the allegation, shown the evidence and allowed to respond. The awarding body decides any sanction. Students must be given JCQ's Information for Candidates on AI before starting coursework from September 2026.
If the college lets teachers use AI to help mark, it cannot be the sole marker. A teacher reviews all of the work and remains responsible for the mark. This design doesn't mark at all.
Ofqual asked exam boards in March 2026 for more rigorous authentication of coursework, and published resources for schools and colleges the same month and an advice note for awarding organisations in April. Expect board requirements to tighten. Evidence of process is what boards and moderators can check.
Student work is personal data, and most students are under 18. Uploading it to any tool needs a lawful basis, a DPIA, UK processing and a contract that stops the provider training on it. A malpractice finding is a significant decision about a person, so it is made by people, never by a model's output.
Products must not use learners' work to train or fine-tune models without the copyright owner's consent, from a parent or guardian for under-18s, and should have a DPIA and clear privacy information. Staff never paste student work into tools the college hasn't approved.
A method that flags students with English as an additional language, or disabled students using assistive technology, more often without evidence risks indirect discrimination. Do an equality impact assessment, monitor concerns by group, and check access arrangements before any concern goes forward.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Intake: drafts, version history, filesVersion history comes from the learning platform or productivity suite the college already uses. Students are told this at the start, not at the allegation. | API Gateway, S3 | API Management, Blob Storage | Kong or your existing API gateway; Ceph or IONOS S3 Object Storage |
| Reading handwritten notesFor plans and notes made in class. Test on real handwriting, including students with access arrangements. | Amazon Textract | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Timelines and questions (drafting)Both clouds offer a choice of models, including Claude. On-prem and sovereign options run open-weight models. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Specs and rules (retrieval)Subject specifications, the college's task briefs, JCQ documents and the malpractice policy, current versions only. | Bedrock Knowledge Bases | Foundry IQ, built on Azure AI Search | OpenSearch, or PostgreSQL with pgvector |
| AI-writing detectionIf the college chooses to use a detector, its output is one input, stored with its known limits, and never the reason for a concern on its own. Ask any supplier for false-positive rates on writing by non-native English speakers and a contract that bars training on student work. | Not part of this design | Not part of this design | Not part of this design |
| WorkflowThe steps of the malpractice process are rules in the workflow: student told, evidence shown, response recorded. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| GuardrailsBlocks outputs that label work as AI-generated or a student as dishonest. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Access controlA teacher sees their own classes. Case files are visible only to the exams officer, the head of centre and named staff. | IAM Identity Center, Amazon Verified Permissions | Entra ID, app roles | Keycloak or your existing Active Directory; Open Policy Agent |
| Keys and data protectionCoursework is personal data about children. Retention follows the awarding body's rules and then ends. | KMS, Macie | Key Vault, Purview | OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog who opened which work and why. No bulk scanning of a whole cohort. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.
Stays on-prem
The student information system, the exams office records and any coursework kept on the college's own file servers stay where they are. Case files can stay on-prem too, since they are small and sensitive.
Uses cloud AI
The timeline, question drafting and retrieval over specifications run in a UK cloud region, or on open-weight models in the college's or a trust's own data centre. Only the piece of work a teacher has opened is sent.
How they connect
Through the API layer over the college's existing secure connection; a private link such as AWS Direct Connect or Azure ExpressRoute is only worth it for a large trust or group. Results come back to the teacher inside the learning platform.
If the link drops
Nothing about authentication depends on the AI. Teachers hold their conversations, supervise sessions and look at drafts as they always have; they just prepare without the timeline and suggested questions.
What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.
Authentication assistant
Where a person signs offThe teacher authenticates the work and decides whether there's a concern. Before the declaration, the college's process decides what happens next; after it, the exams officer reports to the awarding body, which decides any sanction. The assistant's access is enforced by the platform's identity and policy controls, not by instructions in a prompt.
Task design helper
Where a person signs offThe head of department approves every task before it is set, following the awarding body's rules for that subject.
The learning platform stays the record of the work and its versions. The assistant reads from it through the API layer and writes nothing back except a timeline and draft questions attached to that piece of work for the teacher.
The malpractice process is a workflow with required steps, not a model output. A case can't move to the exams officer until the student has been told, has seen the evidence and has had the chance to respond, and those steps are recorded.
There is no fast path. If the AI is down, nothing about authentication or malpractice changes. If it were removed entirely, the college's process would still be complete and fair.
Review the malpractice policy against JCQ's current guidance, map where drafts and version history live, and look at last year's concerns and outcomes by student group.
Gate · must be true to passPolicy updated, DPIA and equality impact assessment started, decision recorded on detectors, safeguarding and SEND leads involved.
Two subjects with substantial coursework. Teachers use timelines and suggested questions on every piece of work, not only suspected ones, and compare with how they worked before.
Gate · must be true to passTeachers find it useful, no automated verdicts anywhere, fairness monitoring in place, and head of centre sign-off.
All coursework subjects, with case files for the exams officer, monitoring, support and a rollback plan, ready well before authentication deadlines.
Gate · must be true to passTraining done, monitoring and rollback in place, and not launched in the final submission window.
Use the task design helper in more subjects, and share the approach across a trust or with a partner university, where the OfS's condition B4 and academic misconduct rules apply instead of JCQ's.
Gate · must be true to passEach subject or partner passes the same pilot gate.
| Activity | Head of centre | Exams officer | Heads of department | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | DPO & equality lead | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Malpractice and AI policy | Accountable | Responsible | Consulted | Informed | Informed | Informed | Consulted | Informed |
| Target architecture | Consulted | Consulted | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted |
| DPIA and equality impact assessment | Accountable | Consulted | Consulted | Consulted | Informed | Informed | Responsible | Consulted |
| Build and integration | Informed | Consulted | Informed | Accountable | Responsible | Responsible | Informed | Consulted |
| Pilot with teachers | Accountable | Consulted | Responsible | Consulted | Responsible | Responsible | Consulted | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Consulted | Responsible | Informed | Consulted | Consulted |
| Run, monitor and review fairness | Accountable | Responsible | Responsible | Consulted | Consulted | Responsible | Responsible | Consulted |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: the integration with your learning platform, the evidence timeline, the case workflow and the gates. I'd report to your head of centre, and I'd say plainly where AI can't answer the question, starting with whether a student wrote their own work.
Option B
If your partner or your own team already has an architect, I'd take on the learning platform integration, the drafting assistant and the case workflow under their lead, and give them a second senior opinion on the rest, especially data protection for children's work.
Goes wrong
A detector score becomes the evidence: '87% AI' goes in the allegation.
How the design avoids it
No score is ever the reason for a concern. The policy and the workflow require evidence of process and a conversation with the student.
Goes wrong
Students with English as an additional language are flagged far more often.
How the design avoids it
Don't rely on style-based signals alone, monitor concerns by group every term, and involve the EAL lead before a concern goes forward.
Goes wrong
Students first hear about version history and supervised drafts when they're accused.
How the design avoids it
Explain the approach at the start, with JCQ's Information for Candidates, and build checkpoints into every schedule.
Goes wrong
Staff paste student work into free online tools to 'check' it.
How the design avoids it
Approved tools only, under a DPIA and a contract that bars training on student work, with the rule in the staff AI policy.
Goes wrong
The tool turns into surveillance of every student.
How the design avoids it
It only runs on work a teacher has opened, it never ranks students, and every access is logged.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Update the policy and coursework schedules now, so the evidence of process exists by the time work is authenticated.
Watch each board's updates for your subjects, and keep authentication evidence in one structured place so new requirements are a report, not a project.
Compare your own concerns and outcomes with the national picture, by type of malpractice.
Keep records showing that people, not models, made each decision, and what evidence they used.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.
