Education AI and Assessment Risk
Education providers use AI to explain course material, draft feedback, answer administrative questions, score responses and identify learners who may need support. These applications do not carry the same consequence. A weak tutoring suggestion can be corrected during learning, while an unsupported score or admissions recommendation can affect progression and opportunity.
Marketways evaluates each AI application against the educational decision it supports. We define what the system may do, what evidence would justify that role, which errors matter, who must review the result and how a learner can challenge a consequential decision.
Begin with the educational decision
An institution cannot evaluate “AI in education” as one use case. A course-information assistant retrieves approved facts. A tutor generates explanations and practice. An assessment system may classify, score or recommend action. Each application serves a different user and requires different evidence.
Marketways identifies the task, user, context and consequence before selecting an evaluation. We ask which decision follows from the output, how quickly an error can be detected and who bears the consequence. This definition determines the required accuracy, explanation, oversight and appeal.
The institution should also state what the AI system must not decide. For example, a workflow may prepare evidence for an academic review without determining misconduct. Clear boundaries prevent a convenient support tool from acquiring authority that the institution never evaluated.
Assistance and decision authority require different controls
An AI assistant can propose an explanation, draft a message or organise evidence while a teacher or administrator remains responsible for the final action. However, human involvement does not automatically make the process safe. The reviewer may receive too many cases, lack the source evidence or assume that the system is usually correct.
Marketways defines the reviewer’s task rather than merely adding a “human in the loop” label. The interface should show the relevant learner record, source material, uncertainty and reason for escalation. The reviewer needs enough time and authority to reject the recommendation and record the reason.
For higher-consequence decisions, management must also examine automation bias and inconsistent review. If reviewers accept the same unsupported recommendation under time pressure, nominal human approval does not provide meaningful protection.
Assessment validity begins with what the score is meant to represent
An assessment supports a claim about a learner’s knowledge or capability. Before using automated scoring, the institution must define that claim, the task that elicits the relevant evidence and the scoring rule that connects the response to the claim.
An AI system can agree with historical scores while measuring the wrong feature. It may reward writing style when the assessment intends to measure scientific reasoning, or penalise a valid response because it uses an unfamiliar expression. Agreement with prior markers is therefore insufficient if the prior process or task does not support the intended interpretation.
Marketways tests the complete validity argument. We examine whether the tasks represent the intended capability, whether the model uses relevant evidence and whether the resulting score supports its stated use. The acceptable evidence differs for formative feedback, course grading, admissions and professional certification.
Generated feedback must remain grounded in the curriculum and learner evidence
Generative AI can produce fluent explanations that appear suitable for a learner. However, fluency does not establish factual correctness, curriculum alignment or diagnostic relevance. A generic explanation may also conceal that the system misunderstood the learner’s error.
Marketways tests responses against approved course content, learning outcomes and instructional policy. Scenarios include incomplete questions, misconceptions, ambiguous wording and requests beyond the learner’s current level. We examine whether the system identifies uncertainty, retrieves the correct source and refers the learner to a teacher when the available evidence does not support an answer.
Personalisation must also have a defined basis. The system should not infer ability, disability or motivation from weak proxies and then narrow the learner’s opportunities. Each adaptation should connect to an observed learning need and remain reviewable.
Performance must be tested across relevant learners and subjects
An average score can conceal systematic error. Performance may differ by subject, language, response length, dialect, disability, educational background or type of reasoning. An institution that serves several learner groups must know whether the system remains dependable across the conditions in which it will operate.
Marketways builds representative test cases from the intended use. We compare accuracy, calibration, false-positive and false-negative rates, explanation quality and escalation across relevant groups and tasks. Where sample sizes are limited, we report the resulting uncertainty rather than treating the absence of a detected difference as proof of equal performance.
The institution must then decide which differences are acceptable for the stated use. A small scoring difference may be consequential near a pass threshold or when the output affects access to a programme. Statistical significance alone does not determine educational importance.
Academic integrity requires evidence, not a model accusation
AI has changed how learners can generate, revise and translate academic work. Institutions need to protect assessment integrity, but automated detection is not equivalent to proof of misconduct. A detector may classify legitimate work incorrectly, and a learner may have an authorised reason for using language or accessibility tools.
Marketways helps institutions define the evidence required before initiating a review. A detection signal may identify a case for examination, but the institution should consider the assessment design, drafts, sources, learner explanation and applicable policy. The AI system should not produce a final misconduct finding.
Assessment design may also need to change. If a task can be completed by reproducing generic information, stricter detection does not restore evidence of the intended learning. Institutions may need tasks that require situated reasoning, documented process, oral defence, practical demonstration or reflection on tool use.
Administrative AI must preserve authoritative records and service access
Education providers also use AI for enrolment questions, timetable support, document checks and service routing. These applications may appear less consequential than assessment, yet incorrect guidance can cause a learner to miss a deadline, submit the wrong evidence or lose access to support.
Marketways identifies the authoritative source for each answer and tests whether the workflow retrieves the current policy, programme and learner context. We include conflicting records, policy changes, missing information and cases that require staff judgement. The system should state when it cannot resolve the question and transfer the learner with the relevant context.
Service quality also depends on accessibility and completion. A quick response has limited value if the learner cannot understand it, act on it or reach a responsible person when the answer is wrong.
Test the complete workflow, not only the model response
A model can answer a test prompt correctly while the deployed workflow still fails. The system may retrieve an outdated policy, attach the wrong learner record, apply a rule to the wrong programme or lose an escalation. Marketways therefore tests the steps that surround the model.
We examine data retrieval, tool permissions, prompt and policy application, case state, handoff, notification and audit records. Multi-step scenarios test whether the workflow maintains the correct learner, programme and decision context as information changes.
The test must also include denied actions. A tutoring agent should not change a grade. An administrative assistant should not expose another learner’s record. An assessment workflow should not bypass a required academic review. These boundaries are part of educational performance, not an optional security test.
Monitoring must connect system behaviour with educational outcomes
Deployment changes the conditions under which the AI system operates. Course material, policies, learner behaviour and model versions can change. Institutions therefore need continuing evidence that the system remains suitable for its approved use.
Marketways defines monitoring measures from the decision and consequence. Measures may include grounded-answer rates, scoring disagreement, subgroup performance, teacher overrides, escalation completion, learner appeals, response correction and downstream progression. Response time and usage volume remain operational measures; they do not establish learning, validity or fairness.
Each material incident should update the evaluation set. This practice turns a production failure into a reproducible test and allows the institution to verify the correction before the system returns to use.
How Marketways evaluates education AI
Marketways begins with the educational purpose, learner population and decision authority. We define expected and prohibited behaviour, develop representative and adverse scenarios, test the complete workflow, validate automated evaluation where it is used, and connect findings with review, appeal and monitoring controls. The resulting evidence shows where the system can assist, where qualified staff must decide and which conditions require redesign or withdrawal.
The engagement can combine AI Evaluation and Assurance with AI Governance and Model Risk, Business Systems Design and Architecture and Statistics and Econometrics. Marketways provides technical and management advisory support; the education provider retains its academic, admissions and disciplinary authority.