Protecting evidence of learning, human judgement and staff capacity.

Human Agency in AI-Era Assessment

Most of the people I talk to in universities right now are holding two things at once: the legitimate worry that AI is changing what assessment can evidence, and the day-to-day work of marking, moderating and supporting students.

The instinct in that state is to retreat into surveillance or rush into redesign. Both make life harder.

In brief

Human agency, in this piece, means the meaningful capacity of students and staff to think, choose, judge and explain within the assessment process.

The argument: AI-era assessment quality depends on protecting human agency across the whole assessment workflow, not only on academic-integrity policy or task redesign.

Generative AI hasn't only made older assessment questions harder to ignore. It has also created a genuinely new challenge. At the population scale, AI-assisted student work is now functionally indistinguishable from independent work in many text-based tasks. Universities will not solve this by detection alone, by redesign alone, or by policy alone. They need to act across the whole assessment workflow, in a way that protects what human judgement and student agency contribute to academic decisions.

The response doesn't need to be a panicked redesign or retreat into suspicion, surveillance or blanket restriction. A careful evaluation of assessment design practices, feedback processes, and workflow pain points can help identify where human agency is needed. Then, new practices and processes can be developed to ensure that this effort is intentional, visible, supported and sustainable.

Book a scoping conversation

Read the AI-Era Assessment Guide

AI is not only an academic-integrity issue

Academic integrity is essential. Universities need fair processes, clear expectations and careful guidance about AI use, but if the conversation stays only at the level of detection, misconduct or acceptable-use policy, it can miss the deeper assessment questions.

AI asks universities to review:

how students demonstrate understanding;
where judgement, interpretation and decision-making appear in the assessment;
what kinds of evidence are strong enough to support academic standards;
how staff can mark, moderate and give feedback without unsustainable workload;
where technology can reduce avoidable friction;
where human effort should remain central;
how students are guided to use, question or avoid AI appropriately;
how trust is maintained between students, staff, departments and institutional systems.

AI-era assessment reviews need to connect design, workflow, policy, platform use, student agency and staff adoption. As with any connected system, changing one part in isolation can lead to unintended consequences elsewhere.

Human judgement needs support, not romanticising

Human judgement isn't automatically fair, consistent or sustainable. Existing processes (e.g. criteria, moderation, calibration, transparent evidence, student communication, and escalation routes) already mitigate the risk of human error.

The case for protecting human judgement against full AI substitution is not that human judgement is reliable on its own. It is that human judgement is improvable in ways that matter for academic standards: through deliberate criteria, through moderation and calibration, through external examining, through transparent evidence trails. AI judgement is improvable too, but along different axes, and not all the axes that matter for academic standards transfer.

This is the working basis for treating human judgement as something to be protected and supported, not replaced. Lodge, Howard, Bearman and Dawson set out a fuller version of the argument in Assessment Reform for the Age of Artificial Intelligence (TEQSA, 2023). The piece here draws on that work and on Carless's work on feedback literacy, alongside the practical experience of universities reviewing assessment under workload pressure.

AI can reduce some of these processes. Not all of them.

Understanding what already happens, deciding what needs to remain human, and then supporting staff and students to make these judgements ensures that human agency isn't replaced by AI where it is needed.

“Should students use AI?”

This is still a contentious question, but it's no longer enough. They already are.

To move the conversation from anxiety into deliberate action, more useful questions include:

What is this assessment trying to evidence?
- Is it assessing knowledge, analysis, judgement, process, performance, creativity, professional reasoning, disciplinary method, communication or something else?
Where does the student need to make decisions?
- Can those decisions be seen, explained, reflected on or discussed?
What kind of AI use is legitimate?
- Can students use AI for brainstorming, planning, editing, feedback, simulation or critique? Where would that support become inappropriate?
What should remain human?
- Where do students need to demonstrate personal judgement, ethical reasoning, disciplinary interpretation, care, dialogue or accountability?
What does the workflow need to support?
- How will submission, marking, moderation, feedback, grade handling, exceptions, declarations or evidence trails actually work?
What would make this sustainable for staff?
- A stronger assessment design can still fail if it creates unmanageable marking, unclear moderation or extra manual work.

Human agency means students are not only producing outputs

In an AI-era assessment context, human agency means students have meaningful opportunities to think, choose, judge, explain, revise, respond and take responsibility for their work.

That doesn't mean every assessment must become highly complex or fully authentic. Some existing essays, exams, problem sets, performances, presentations, portfolios and practical tasks may still be appropriate.

Assessments need to provide students with a fair and meaningful way to demonstrate learning.

Human agency may become more visible through:

explanation of choices made during the work;
staged drafts or checkpoints;
reflective commentary;
feedback response;
oral discussion or viva-style elements;
process evidence;
applied or localised tasks;
discipline-specific judgement;
professional or ethical reasoning;
collaborative decision-making;
appropriate use, critique or non-use of AI.

A practical caveat: students themselves are now using AI to generate the reflective commentaries that were designed to evidence agency. This is not a reason to abandon reflective approaches. It is a reason to design them so that the reflection happens in the room (oral discussion, in-class viva, conversation with a marker), or so that the reflection refers to specific decisions only the student could have made: this version, this choice, this revision.

None of these approaches is a universal answer, and each brings workload, equity, accessibility, moderation and workflow implications, but with careful planning, they can make student thinking and engagement in assessment visible.

Student agency includes evaluative judgement

In the AI era, students need practice judging the quality of sources, arguments, feedback, AI outputs, AI-supported processes and their own developing work.

Assessment should therefore create opportunities for students to explain choices, compare alternatives, check evidence, revise in response to feedback and decide when AI use helps or weakens their work.

The concept of evaluative judgement in the AI era is developed in Bearman, Tai, Dawson, Boud and Ajjawi's 2024 paper Developing evaluative judgement for a time of generative artificial intelligence (Assessment & Evaluation in Higher Education, 49(6)), which is a useful starting point for assessment leads.

Evidence of learning needs to be designed, not assumed

Before AI, many assessments relied heavily on the final submitted product. In some contexts, that may still be enough, but in others, it may no longer provide strong enough evidence on its own.

A useful review asks:

What evidence does this task currently produce?
What evidence is missing?
What evidence would be stronger, fairer or more educationally useful?
Who will interpret that evidence?
How will staff be supported to make consistent judgements?
How will students understand what is expected?
What will happen when the evidence is ambiguous?

An AI-writing indicator, similarity report, process log, draft history, oral explanation or reflective statement can all provide different kinds of information. None of them removes the need for academic judgement.

The risk is that by attempting to reduce inappropriate use of AI, institutions end up with more data than they can interpret fairly, or introduce tools without clear processes around them.

Strong evidence of learning is technical, educational, operational and ethical all at once.

Staff workload is part of assessment quality

Staff capacity is not abstract. In many UK and US institutions, marking burdens in assessment-heavy semesters are already past sustainable, before any AI-era redesign is considered. An AI-era redesign that ignores this is not a redesign; it's a reallocation of the same problem.

AI-era assessment conversations often focus on student behaviour. Staff capacity needs equal attention.

A redesigned assessment that protects integrity but doubles the marking burden will not be sustainable. A tool that promises efficiency but creates new checking, exception handling, or communication work may not reduce the workload in practice. A policy that looks clear centrally may still leave staff unsure what to do in day-to-day marking, moderation or student guidance.

This is why assessment change needs workflow thinking.

Universities need to understand:

where staff time is currently being lost;
which processes rely on local memory or workarounds;
where Moodle, Canvas or another platform is supporting the process well;
where the platform is being blamed for unclear process;
what guidance staff need;
what decisions should be standardised;
where academic judgement should remain flexible;
what support is needed before any new tool or approach is rolled out.

Staff can't exercise careful judgement if they are surrounded by unclear, fragmented or overloaded policies and processes. Protecting human agency includes protecting staff capacity.

Where AI should reduce friction

AI and assessment technologies can be useful where they reduce avoidable friction.

That may include:

summarising themes from feedback or evaluation data;
supporting staff with first-draft guidance materials;
helping students practise with formative questions;
improving accessibility of instructions;
supporting workflow mapping or documentation;
helping teams compare policy options;
reducing repetitive administrative drafting;
supporting structured feedback processes;
helping staff identify patterns that need human review.

In these cases, AI can help to reduce the workload.

Where can AI most effectively improve the educational process without weakening judgement, trust, fairness or accountability in your institution?

Where human effort needs to remain

Some forms of effort should not be removed simply because they are difficult.

Assessment still needs human judgement.

Feedback still needs educational purpose.

Moderation still needs careful interpretation.

Students still need to develop their own understanding, voice and responsibility, while staff still need space to make contextual decisions.

Human effort needs to remain where the work involves:

academic standards;
disciplinary judgement;
ethical reasoning;
interpretation of ambiguous evidence;
student support;
feedback that helps learning;
moderation and fairness;
decisions with consequences for students;
care, dialogue and accountability.

Some existing processes may be unnecessarily heavy, inconsistent or unclear, and where this is true they don't need to stay.

Separate wasted friction from useful effort.

Wasted friction should be reduced.
Useful effort should be designed for, supported and made sustainable.

Equity, access and the unintended cost of surveillance

AI-era assessment responses have equity consequences that don't always surface in policy discussions:

Detection tools have well-documented higher false-positive rates for non-native English writing. Surveillance-led responses disproportionately affect international students and students with English as an additional language.
For neurodivergent students and students with some disabilities, AI scaffolding now functions in practice as a form of accommodation. Blanket restriction can withdraw an accommodation a student depended on.
Students with significant caring or working commitments outside university increasingly rely on AI for the kinds of organisational scaffolding that better-resourced peers receive from parents, tutors or peer networks.

A response framework that protects human agency without addressing these dimensions reproduces existing inequities under a new label. Assessment design needs to recognise the differences in how students approach AI, while protecting the standards that make the qualification meaningful.

Trust is built through clarity, not certainty

AI has introduced more ambiguity into assessment.

Universities don't need perfect certainty before they act, but clearer processes, better questions and more careful implementation.

Trust is strengthened when students understand what is expected, staff understand how to respond, and policies translate into practice.

That means universities may need to review:

assessment briefs;
AI-use guidance;
declaration processes;
feedback workflows;
marking and moderation arrangements;
student communication;
platform settings;
evidence trails;
escalation routes;
staff development;
pilot design and evaluation.

Trust doesn't come from a single policy, tool, or detection method, but from aligning assessment purpose, evidence, workflow, and human judgement.

Trust also has a regulatory dimension. The UK's Office for Students Condition B4 requires that assessment is effective, rigorous and produces credible awards. The QAA's papers on assessment in the generative AI era translate the regulatory ask into practical guidance. The workflows that support assessment quality also support the evidence trails that regulators and accreditors require.

A practical review lens

The five lenses below are the framework I use in scoping diagnostics with universities. They draw on the assessment-for-learning tradition (Boud), feedback literacy (Carless), and the evaluative judgement work of Bearman, Lodge and colleagues. The framing here is my own; the underlying ideas are rooted in research.

When reviewing assessment in the AI era, five connected areas usually need attention together.

Lens	Question	This includes:
Capacity	Is the human effort required sustainable for staff and students?	Staff workload, training, support, local workarounds, digital processes, pilot planning and implementation capacity.
Trust	Are expectations, evidence and decisions transparent, fair and contestable?	Academic standards, transparency, consistency, student communication, human oversight, policy-to-practice alignment and shared confidence in how evidence is interpreted.
Workflow	Can the process support the intended learning, feedback, moderation and records?	Submission, marking, moderation, feedback release, grade handling, exceptions, platform use, handoffs and staff guidance.
Agency	Where do students think, choose, question, revise, explain and take responsibility?	Opportunities for students to make decisions, justify their approach, respond to feedback, reflect on process, use tools critically, explain their reasoning and show how their understanding has developed.
Evidence	What learning, judgement or capability does the assessment actually show?	Assessment design, student judgement, process visibility, feedback response, and the role of any AI-use declaration or supporting evidence.

Looking across these areas helps universities avoid treating AI as a narrow policy problem or a simple tooling decision. It also helps identify the right next step: assessment redesign, workflow redesign, Moodle or platform support, staff guidance, a pilot, or a broader diagnostic.

Make the right work easier to do

AI-era assessments require redesign, feedback and workflows that protect learning, judgement, trust and staff capacity.

Before taking action, a useful first step is to understand what is happening now: where the evidence is strong, where the workflow is fragile, where staff are carrying hidden and unnecessary workloads, and where students need clearer guidance.

From there, the next step becomes easier to see.

About this piece, and how I can help

Naomi Rowan is the founder of Gratitude Worldwide Ltd, a fully-remote, UK-based consultancy supporting universities on AI-era assessment, feedback, Moodle and platform workflow, and staff adoption. Engagements include work with the London School of Economics (assessment workflows and Moodle-based processes) and Nord Anglia Education.

The consulting work most often takes one of these forms:

assessing whether current assessment tasks still provide strong evidence of learning;
identifying where student judgement and process could be made more visible;
reviewing marking, moderation and feedback workflows;
mapping where staff workload is being lost to avoidable friction;
clarifying where AI or assessment technology may help;
supporting pilot design and evaluation;
developing staff guidance, adoption materials or briefing papers;
helping teams move from broad AI concern into practical next steps.

Most work begins with a scoping conversation, a focused senior-leader briefing, or an Assessment & AI Workflow Diagnostic.