Project: RAISE-UP (Responsible AI Strategy for Education in Universities)

Role: Coordinator (RTU/RBS Leads)

Duration: 24 Months

INTRODUCTION

AI changes what assessment evidence can prove

THE SHIFT

Generative AI can plan, draft, revise, explain and evaluate work at low cost and increasing quality.

A polished submission may represent the student's capability, the AI system's capability, or a combined human-AI performance.

WHY IT MATTERS

Validity asks whether the evidence supports the intended learning claim.

Integrity asks whether assistance, authorship and disclosure match legitimate expectations.

Trustworthiness asks whether the evidence, grading and procedures deserve confidence.

TAKEAWAY

The key question is not simply "Was AI used?" but "What capability does this evidence justify?"

Sources: Dawson et al. (2024) · Kofinas et al. (2025) · Luo (2025)

THESIS 01 OF 16

Validity comes before the question of cheating

PROBLEM

AI can contaminate the construct: a grade may reflect external capability that the assessment did not intend to measure.

Task vulnerability varies with the learning outcome, task demand and current AI capability; simply increasing control can introduce new validity problems.

SOLUTION

Define the exact capability the assessment must certify and specify whether it must be demonstrated independently or with AI support.

Judge whether the available evidence supports that claim, rather than treating the mere presence of AI as the primary issue.

TAKEAWAY

Assessment reform should begin with defensible inference: what can this performance legitimately tell us about the student?

Sources: Dawson et al. (2024) · Bannister et al. (2025) · Kofinas et al. (2025) · Rai et al. (2025)

THESIS 02 OF 16

Integrity now depends on contextual authorship and responsibility

PROBLEM

GenAI can replace expected intellectual work without copying an identifiable source, so plagiarism rules alone are insufficient.

Students may not understand declarations consistently and may avoid disclosure when they fear that honesty will be treated as self-incrimination.

SOLUTION

Define which decisions, judgements and contributions must remain the student's own for each task.

Use concrete examples of permitted assistance, prohibited substitution, attribution and realistic disclosure.

TAKEAWAY

Integrity is a shared practice: students remain responsible, while educators must make legitimate assistance intelligible.

Sources: Perkins (2023) · Luo (2024) · Gruenhagen et al. (2024) · Gonsalves (2025)

THESIS 03 OF 16

AI detectors cannot restore trust on their own

PROBLEM

Detector performance varies across tools, models, texts and simple adversarial changes. False positives and false negatives remain consequential.

Even comparatively strong aggregate accuracy is not reliable enough to prove individual misconduct, and error may be distributed unequally.

SOLUTION

Treat detection only as a possible signal for careful review, never as proof of authorship, intention or learning.

Require corroborating evidence, trained human judgement, student response and proportionate decision-making.

TAKEAWAY

A classifier can outperform chance and still be unsuitable for a high-stakes judgement about an individual student.

Sources: Elkhatat et al. (2023) · Perkins et al. (2024b) · Paustian & Slinger (2024) · Kofinas et al. (2025)

THESIS 04 OF 16

Trust depends on reciprocal transparency and fair procedure

PROBLEM

Students may be asked to disclose AI use or share process records while educator AI use, grading interpretation and suspicion remain hidden.

Teachers may apply unstated values such as diligence, honesty or effort beyond the published rubric.

SOLUTION

Explain what information is collected, how declarations are used, how concerns affect grading and how decisions can be challenged.

Make values relevant to AI-supported work explicit in criteria and moderation, not invisible moral penalties.

TAKEAWAY

Trust is relational: one-way visibility produces surveillance, while two-way accountability supports legitimate assessment.

Sources: Gonsalves (2025) · Kizilcec et al. (2024) · Luo (2025) · Luo & Dawson (2025)

THESIS 05 OF 16

Fluent production is no longer reliable evidence of learning

PROBLEM

Students can submit acceptable surface responses without sufficient understanding, making product quality easier to separate from capability.

Process logs are not automatic proof: they can be manufactured, become intrusive and create workload.

SOLUTION

Assess evaluative judgement: the ability to recognise limits, verify claims, adapt suggestions and justify decisions.

Collect proportionate evidence of reasoning, revision, error and dialogue where the assessment stakes require it.

TAKEAWAY

The evidential focus must move from production alone toward judgement and learning, without turning assessment into total surveillance.

Sources: Bearman et al. (2024) · Walton et al. (2025) · Winstone et al. (2026)

THESIS 06 OF 16

AI permission must be aligned with outcomes, criteria and conditions

PROBLEM

Labels such as prohibited, restricted or permitted can conceal very different uses and do not protect validity by themselves.

If AI is allowed but the rubric rewards only a polished final product, the assessment still cannot distinguish the capability that matters.

SOLUTION

Start with the learning outcome, then define permitted tools and uses, prohibited substitution, disclosure, verification and grading.

Use a continuum: independent performance, limited support or accountable AI-integrated performance, depending on the claim.

TAKEAWAY

The defensible level of AI use follows from the learning claim, not from a blanket institutional preference.

Sources: Furze et al. (2024) · Perkins et al. (2024a, 2025a, 2025b) · Corbin et al. (2025)

THESIS 07 OF 16

Redesign the intellectual work, not just the topic

PROBLEM

A conventional essay or generic solution can often be reproduced by changing prompts. Higher-order wording alone does not ensure higher-order performance.

Tasks remain vulnerable when context, evidence and criteria allow generic AI output to pass without scrutiny.

SOLUTION

Ask students to critique, verify, create, adapt and justify in relation to disciplinary standards.

Assess how students compare alternatives, reject weak outputs, validate sources or calculations and explain their final decisions.

TAKEAWAY

AI can generate options; the grade should depend on the student's accountable evaluation and transformation of them.

Sources: Bearman et al. (2024) · Bannister et al. (2025) · Peters & Angelov (2025) · El-Soussi & Yousef (2026)

THESIS 08 OF 16

Layered evidence makes the learning trajectory visible

PROBLEM

A single final artefact provides weak attribution when AI can shape planning, drafting and revision.

Requiring every prompt and version can create privacy concerns, strategic performance and unsustainable workload.

SOLUTION

Use selected stages such as a proposal, annotated sources, draft extract, feedback response, revision rationale or critical-decision note.

Combine portfolios and reflection with concrete examples of what the student accepted, changed or rejected.

TAKEAWAY

The goal is triangulation, not total capture: collect only the process evidence that strengthens the intended inference.

Sources: Hsiao et al. (2023) · Foung et al. (2024) · Portuguez-Castro & Castillo-Martínez (2026) · Winstone et al. (2026)

THESIS 09 OF 16

Dialogue adds evidence, but it must be moderated and accessible

PROBLEM

Written products may not show whether students can explain, transfer or defend their reasoning.

Vivas and presentations can introduce anxiety, language, disability, cultural and assessor effects, as well as substantial workload.

SOLUTION

Add a short, targeted defence, demonstration or synchronous question sequence linked to an existing product.

Design questions, moderation, scoring and reasonable adjustments so dialogue complements rather than replaces other evidence.

TAKEAWAY

Dialogic assessment is strongest as a focused second source of evidence, not a universal substitute for written work.

Sources: Head & Willis (2024) · Kickbusch et al. (2025) · Dollinger & Nieminen (2026) · Perkins (2026)

THESIS 10 OF 16

Authenticity requires accountable AI use, not just realism

PROBLEM

A realistic workplace task may be easy to complete with GenAI, and markers may still be unable to identify AI-influenced work.

Authenticity alone therefore does not protect integrity or demonstrate foundational knowledge.

SOLUTION

Use situated data, local constraints, evolving scenarios, disciplinary tools and stakeholder consequences.

Require students to verify AI output, explain professional responsibilities, defend decisions and identify what remains their own judgement.

TAKEAWAY

Authentic assessment should certify responsible performance under realistic conditions, balanced with independent evidence where needed.

Sources: Head & Willis (2024) · Kofinas et al. (2025) · Kickbusch et al. (2025) · Peters & Angelov (2025)

THESIS 11 OF 16

Redesign at programme level and keep educators accountable

PROBLEM

If every module adds its own viva, log or controlled test, workload and duplication grow while capability gaps may remain.

Educator-facing AI can assist with tasks, rubrics, feedback or provisional evaluation, but it cannot own the final judgement.

SOLUTION

Map where programmes assure independent knowledge, teach AI literacy, integrate professional AI use and triangulate high-stakes decisions.

Use AI for design or feedback only under educator calibration, review, privacy safeguards and human responsibility.

TAKEAWAY

Coordinated human-led augmentation is more sustainable and valid than isolated task-by-task reactions.

Sources: Xia et al. (2024) · Kadel et al. (2026) · Nicola-Richmond et al. (2026) · Ahangama (2026) · Portuguez-Castro & Castillo-Martínez (2026)

THESIS 12 OF 16

Responsible AI assessment rests on knowledge and judgement

PROBLEM

Students cannot critically evaluate fluent output in a field they have not yet learned.

A declaration form or revised rubric cannot substitute for instruction in verification, bias, attribution, privacy and appropriate non-use.

SOLUTION

Sequence foundational knowledge, AI literacy and evaluative judgement across the programme.

Provide modelling, formative practice and feedback before applying high-stakes expectations of responsible AI use.

TAKEAWAY

AI literacy is not prompt technique; it is disciplined judgement about when, why and how to use or reject AI output.

Sources: Dawson et al. (2024) · Bearman et al. (2024) · Corbin et al. (2025) · Walton et al. (2025) · Perkins et al. (2025a)

THESIS 13 OF 16

Transparency must work in both directions

PROBLEM

General policy leaves students uncertain when rules vary across modules, while extensive process disclosure can become one-sided surveillance.

Hidden moral judgements about honesty, diligence or effort weaken grading consistency and validity.

SOLUTION

State task-level permissions, prohibited substitution, disclosure, error responsibility and the role of AI in grading.

Explain educator AI use, data processing, concern procedures and how declarations affect decisions.

TAKEAWAY

Ethical assessment requires proportionate, reciprocal transparency, not maximum disclosure from students alone.

Sources: Perkins et al. (2024a, 2025a) · Gonsalves (2025) · Luo (2025) · Luo & Dawson (2025)

THESIS 14 OF 16

Suspicion must trigger due process, not automatic punishment

PROBLEM

Detector scores and marker intuition are probabilistic and can produce false accusations, missed cases and biased grading.

A serious integrity decision is illegitimate when screening and judgement are collapsed into one step.

SOLUTION

Set a clear review threshold, preserve relevant evidence and let the student explain using drafts or related performance.

Use trained decision-makers, a visible standard of proof, an appeal route and monitoring for disproportionate impacts.

TAKEAWAY

Accountable human judgement supported by corroborating evidence protects students and strengthens genuine integrity findings.

Sources: Elkhatat et al. (2023) · Perkins et al. (2024b) · Paustian & Slinger (2024) · Kofinas et al. (2025)

THESIS 15 OF 16

Equity, privacy and student agency are validity conditions

PROBLEM

Students differ in access to paid models, devices, connectivity, accessible interfaces, language support and prior experience.

AI-resistant formats can also disadvantage disabled, multilingual, anxious or time-constrained students; commercial tools create privacy and intellectual-property risks.

SOLUTION

Provide approved tools or equivalent routes, inclusive adjustments, data-minimisation guidance and clear external-processing information.

Preserve a legitimate non-AI option when AI use is not itself a learning outcome.

TAKEAWAY

Neither compulsory AI use nor blanket resistance is equitable; students need equivalent ways to demonstrate the same outcome.

Sources: Perkins et al. (2024b) · Jin et al. (2025) · Dollinger & Nieminen (2026) · Lindkvist et al. (2026)

THESIS 16 OF 16

Institutional capacity determines whether redesign is sustainable

PROBLEM

Individual educators cannot absorb programme mapping, redesign, moderation, privacy, integrity procedure and accessibility work without time and support.

Static policy cannot keep pace with changing tools, professional practices and student behaviour.

SOLUTION

Invest in staff capability, collaborative programme work, defined roles, design support, resources and stakeholder participation.

Use stable principles with disciplinary discretion, pilots, moderation, monitoring, scheduled review and transparent reasons for change.

TAKEAWAY

Adaptive governance means accountable local judgement within stable principles, not permanent inconsistency.

Sources: Awadallah Alkouk & Khlaif (2024) · Jin et al. (2025) · Corbin et al. (2026) · Nicola-Richmond et al. (2026) · Zeivots et al. (2026)

SUMMARY · FIVE DESIGN COMMITMENTS

Confidence comes from a coherent assessment system

01

Define the capability to be certified and the conditions under which it must be demonstrated.

02

Match AI permissions, task demands and criteria to that learning claim.

03

Triangulate meaningful product evidence with selected process, dialogue or performance evidence when stakes require it.

04

Make disclosure, grading, educator practice and integrity procedures reciprocal, transparent and fair.

05

Coordinate assessment across programmes and continually evaluate learning, equity, trust and workload.

TAKEAWAY

AI does not remove the need for assessment. It makes the logic, evidence and accountability of good assessment impossible to ignore.

Sources: integrated synthesis of all 16 theses in the supplied literature review

PUTTING IT INTO PRACTICE

What this means for you

01

A lecturer: redesign one course. Define the capability being certified, set the AI-permission level, adapt the rubric.

02

A programme: map where independent knowledge is assured, where AI literacy is taught, and where high-stakes decisions are triangulated.

03

An institution: write your own guidelines rather than adopting someone else’s, and trial them in a controlled sandbox first.

04

A student: clear rules, a legitimate non-AI route, due process, and transparency that runs both ways.

TAKEAWAY

Any institution can start making these decisions now. All RAISE-UP outputs are published openly at raiseupai.eu.

Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Education and Culture Executive Agency (EACEA). Neither the European Union nor EACEA can be held responsible for them.

Share your use case

We are surveying teaching staff, students, management, and policymakers

We are collecting real use cases

We are developing three tools and a sandbox starter kit

WHAT WE ASK OF YOU

Describe your generative AI use case

https://raiseupai.eu/take-part/

Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Education and Culture Executive Agency (EACEA). Neither the European Union nor EACEA can be held responsible for them.