The challenge is not simply whether to trust AI, but how to integrate it responsibly into the assessment process. This talk, given at the 51st annual conference of the International Association for Educational Assessment, whose theme was Trust, Transparency and Technology in Educational Assessment, follows that question in five parts.
1. The assessment challenge
Digital assessment is a long sequence of steps: didactic and pedagogical design, item writing, psychometric considerations, statistical validation, technical implementation. Only part of it is technical. Assessments must also fit the way institutions are organised, and be accepted by learners, teachers and society.
Delegation can go more or less far:
- Human only: humans design, administer and score.
- AI-assisted: AI supports specific tasks.
- AI-led, human-supervised: humans validate AI decisions.
- Fully delegated: AI conducts the entire assessment.
2. The AI landscape
Capabilities move fast. In our field, the markers follow one another: automatic scoring of closed responses (2022), reliable scoring of short written responses (2024), interpretation of graphical and handwritten responses (2025), agents simulating entire test sessions (2026).
The frontier is also jagged. A model can perform above expert level on graduate science questions and still fail at tasks most people find trivial. The closer we get, the higher the bar is set.
3. The Big Switch
Where do assessment platforms stand, and what comes next? The question at each stage is simple: who is driving?
| Stage | What happens | Who drives |
|---|---|---|
| 1. Platform without AI | Humans design, deliver and score | The platform is the tool |
| 2. AI inside the platform | AI assists specific operations | Humans still drive |
| 3. AI uses the platform | The agent operates the platform’s functions | Humans commission and validate |
| 4. No platform needed | AI generates the tool for the need at hand | Humans state the need |
Stage 2 is already here: retrieving and generating items, quality control of content, estimating the equivalence of test forms, scoring open-ended and graphical responses, feedback and reports, translation and accessibility adaptation.
Stage 3 turns the AI into the platform’s user. A human commissions the work, for example “a 40-minute diagnostic test on fractions, grade 6, two parallel forms”. The agent drives authoring, the item bank, publishing, scoring and reports through the platform’s own functions. The human reviews, adjusts and signs off. The condition is that the platform exposes its functions as tools, with permissions and traceability.
Stage 4 is the switch: the platform stops being the product and becomes something the AI can improvise. It would mean oral interaction, fluid administration and handwriting read directly, with cultural bias explicitly tested. None of this exists as a product today. Every line is an open engineering and validation problem, and it demands a very high level of trust.
4. Keeping control
Why keep control
Models drift, and none of the drifts looks like a bug. A wrong score looks like a score.
| Drift | What the model does | Risk when marking |
|---|---|---|
| Hallucination | Produces false information with confidence | Flags an error that does not exist |
| Poor calibration | States a conclusion more firmly than the evidence allows | The student cannot tell a certain error from an uncertain judgement |
| Stochasticity | Gives the same work different analyses | Reproducibility and fairness break down |
| Context failure | Misses the teaching situation | Correct in general, wrong for this level or objective |
| Sycophancy | Tries too hard to agree | Validates a weak answer because the student insists |
| Framing sensitivity | Depends on how the prompt is written | Two slightly different prompts, two different marks |
| Position bias | Is influenced by order and presentation | Some parts weigh more than they should |
| Confirmation bias | Lets an early reading shape what follows | Once read as weak, every ambiguity is read that way |
| Rubric drift | Moves away from the criteria | Marks general quality, not what the teacher asked |
| Specification gaming | Optimises the instruction rather than the goal | Feedback that fits the rubric without helping learning |
| Overcorrection | Always finds something to improve | The student is overloaded |
| Style normalisation | Prefers the most conventional wording | Penalises originality |
| Bias and stereotyping | Treats some groups or styles differently | Linguistic, social or cultural discrimination |
| Automation bias | Here the bias is ours: we trust the machine | The teacher stops checking |
One drift is slow. A student writes, the AI corrects, the student rewrites, and learns what the AI prefers. Each correction looks reasonable. Over a year, the system puts steady normative pressure on how students write, and on what counts as good writing.
Every safeguard has a price. Scoring twice, checking the rubric and auditing a sample all consume tokens. Prices per token keep falling, yet spending is hard to predict: reasoning tokens are invisible but billed, and agentic workloads use far more tokens than a chatbot.
Six priorities to settle before delegating anything
- Reliability: make the result dependable, and prove it. Not a demo: proof, on your own population. Reliability comes from the rubric and the reference texts, not from the model’s memory. Short responses already reach high reliability; longer written responses still need substantial validation.
- Feasibility: mix deterministic and probabilistic systems. Never replace an exact rule with a probability. Hybrid architectures extend existing systems rather than replace them.
- Controllability: keep humans in command. Trust should follow the stakes. In formative assessment, AI can act as a tutor. In summative and certifying assessment, it acts as an assistant under human supervision. Three safeguards make it acceptable: transparency about how AI contributed, a right to contest with a second human marking, and human authority to settle disputes, not one AI arbitrating another.
- Infrastructure: decide where the models run. Cloud APIs buy speed and immediate access to the most capable models, with variable bills and content leaving your infrastructure. Local open-source models buy predictable costs and control of data, and require in-house expertise. High-stakes content should not go through public models.
- Agentic design: specialised agents, one task at a time. WASPL illustrates this with item evaluation before use, scoring of open responses against the rubric by a language model, reading of visual responses by a vision-language model, and a composite score that combines deterministic rules with AI judgements, each traceable.
- Orchestration: coordinate the agents into one process. Retrieval, generation, quality control, equivalence, scoring and reporting agents become a system only when a higher-level layer coordinates and supervises them.
One example of agentic design: testing a rubric before students see the item. The item writer drafts the item and its rubric; AI simulates a wide range of student responses; the rubric is tested on them; instructions and rubric are revised before any real student takes the test.
5. Conclusion: optimise what exists, anticipate what comes
Today, with the models we already have:
- use AI where it is already reliable: content quality control, short-answer scoring, feedback drafts;
- keep deterministic scoring where it works;
- right-size the model for each task;
- trade speed for quality in batch work;
- expose your platform’s functions as tools;
- audit continuously, with a human-marked sample every session, on every population.
Tomorrow is mostly organisational work, and it can start now:
- write specifications that machines can read;
- invest in item-bank metadata;
- prepare a sovereign infrastructure;
- define the human rules: who validates, who arbitrates, how a score can be contested.
Other professions have already faced this shift: radiologists, translators, software developers, legal professionals. Their fields reorganised around hybrid workflows. Digital assessment will likely follow a similar path. Trust, after all, must be earned.
The Advisory page shows how these priorities structure our work with institutions.