IAEA 2026: The assessment problem each country chose to solve first

Reflections from the 51st Annual IAEA Conference, Toronto

Published: 10/7/2026
IAEA 2026: The assessment problem each country chose to solve first

After the first days of IAEA 2026, one thing became increasingly clear: the future of assessment will not arrive in the same form everywhere.

The conference brought together systems facing many common pressures: unequal access, weak public trust, high-stakes accountability, demands for better data, and now the rapid arrival of generative AI. But the most interesting presentations did not offer one universal solution. They showed different jurisdictions choosing different problems to solve first.

That matters because assessment technology can only be judged against the problem it is meant to solve.

“Technology must work for the assessment; assessment should not work for the technology.” Confidence Dikgole made this point during the IAEA panel Global Perspectives on Assessment Modernization: Balancing Technological Innovation with Public Trust.

She also raised a sharper concern. Technology can inadvertently become a way of testing privilege rather than the intended capability. If participation depends on a particular device, reliable bandwidth, a quiet space, digital familiarity, or access to better AI, then the assessment may measure those advantages as much as it measures learning.

The question is not whether a system is digital, uses AI, or gathers more data. The question is whether its approach produces better evidence for the learners and decisions it serves.

South Africa: access, equity, and evidence that reflects the whole learner

In Digital Assessment Beyond Platforms: Building Low-Bandwidth Systems for Holistic Learner Data at Scale, Crysty Swift and Dr Bridget Mthembu described a South African approach that begins with unequal infrastructure.

The challenge is not simply to digitise assessment. It is to create a system that can operate fairly where connectivity, electricity, devices, and digital familiarity cannot be assumed. Their work treats low-bandwidth delivery as a design principle, not a compromise.

It also seeks to broaden the evidence collected about learners, bringing together academic attainment, competencies, and learner profiles. The choice is not always between a conventional examination and a sophisticated online platform. There may be intermediate approaches that protect access while still making better use of assessment evidence.

The wider lesson is uncomfortable for systems that equate technical sophistication with progress. An assessment that depends on fast connectivity or a particular device may measure access to those things as much as it measures learning.

Kazakhstan: the admissions problem

Kazakhstan’s starting point is different. In Kazakhstan’s Journey in Reforming University Admissions for the 21st Century, Nurseit Baizhanov and Terran Brown set out a reform agenda focused on the limitations of a single, rigid university-entry examination.

Admissions tests are often expected to select students fairly, identify readiness for different courses, shape what happens in schools, and provide a credible national benchmark. A single test is rarely well suited to every one of those purposes.

The proposed modular approach separates common skills, such as critical thinking and data reasoning, from more programme-specific profile assessments. It treats admissions as more than an administrative gateway. It becomes a lever for curriculum and system reform.

The question for other jurisdictions is whether every candidate is asked to demonstrate the same thing because that is what matters, or because the system has found it administratively convenient.

Italy: data that helps people act

Italy’s contribution begins with a familiar frustration: national assessment can produce vast quantities of data without making much difference in classrooms.

In The Italian Approach to Accessible and Equitable Communication of Assessment Data, Dr Patrizia Falzetti and Michele Cardone described the work of INVALSI in returning evidence to schools and making assessment data more accessible to users beyond the national system.

The problem is not a lack of measurement. It is the gap between data production and data use.

Assessment systems often collect information from learners, schools, and teachers, then send it upwards into reporting structures. A score eventually returns, but the people closest to learning may receive little that helps them understand what to do next.

Italy’s example makes a more demanding proposition: assessment data should serve several audiences, including schools, researchers, policymakers, and the public. Assessment should give something back.

Malaysia: local validity before scaled AI

AI-supported scoring featured throughout the conference, but Malaysia’s work offered a corrective to the idea that an apparently capable global model is automatically suitable for national assessment.

In Promoting Fairness for Assessment of English Writing with Small Language Models, Dr Shafiza Mohamed and Paul Edelblut explored the use of locally tuned smaller language models for writing assessment in a multilingual context.

Their starting problem is not simply how to automate scoring. It is how to avoid importing assumptions about language, expression, and quality that may not fit the learners or the national rubric.

A model can be technically impressive and still be locally invalid. It may misread regional discourse patterns, multilingual expression, or the legitimate ways candidates communicate in a particular context.

The lesson is clear: assessment organisations cannot outsource validity. Local expertise, inspectability, fairness analysis, and alignment with the intended construct remain essential.

New South Wales: attributable individual performance

The New South Wales example confronts the question now troubling many systems most directly: what can a take-home product tell us about an individual student when generative AI can produce polished work?

In When Take-Home Assessment Loses Trust: Australian Evidence on AI and Supervised Digital Exams, Alex Dore and Monica O’Brien examined the pressure on take-home assessment in the age of AI and the case for supervised digital environments.

This is not an argument against screens. Nor is it an argument that every assessment must become a traditional examination.

It is an argument about attribution.

Where the claim concerns what an individual can do unaided, an unsupervised product may no longer provide sufficiently strong evidence. That does not invalidate AI-open work. In some contexts, using AI well may be part of the intended competence. But assessment systems need to be explicit about which claim they are making.

Are they assessing independent human performance, human-AI collaboration, or the ability to supervise, challenge, and correct AI output?

The strength of the evidence must match the strength of the claim.

Ghana: equitable access in high-stakes assessment

In Evaluating Examination Access Arrangements for Candidates with Visual Impairments in the West African Senior School Certificate Examination in Ghana, Wendy Enyonam Addy-Lamptey focused on a problem that is often treated too late: access for disabled candidates in high-stakes assessment.

The underlying issue is whether access arrangements depend on the capacity, resources, or confidence of an individual school. If they do, then access becomes uneven by design.

A national approach to accessible materials, assistive support, and examination arrangements treats inclusion as part of the assessment system itself. It is not an optional layer added after the core assessment has been designed.

Pakistan: fragmented evidence

Pakistan’s case begins with a challenge that many systems will recognise: data exists, but it is spread across organisations, jurisdictions, and reporting structures that do not automatically work together.

In Navigating the Landscape of Multiple Data Sources for Educational Reporting in LMICs: The Case of Pakistan, Saima Maheen examined the complexity of using evidence across provincial, federal, citizen-led, and specialist assessment systems.

The problem is not simply collecting more data. It is deciding what evidence different users need, helping them interpret it, and ensuring they have the authority and capacity to act.

Data abundance is not the same as understanding.

The question that remains unresolved

These presentations do not point to a single future of assessment. They show systems tackling different immediate problems: infrastructure, admissions, data use, local validity, attributable performance, equitable access, and governance.

That diversity is one of the strengths of IAEA. It prevents us from treating the preferred technology of one system as the answer for every other system.

But it also reveals a problem that remains unresolved.

Many innovations now seek to capture more of the assessment process: response times, revision patterns, confidence ratings, navigation behaviour, keystrokes, scaffolding, and interaction trails. In her keynote, Reimagining Assessment in the AI Era: From Assessment Triangle to Assessment Ecology, Dr Eunice Jang argued for a broader view of the evidence and context surrounding assessment.

The promise is compelling. We may be able to understand more than the final answer alone.

But richer telemetry does not automatically provide better evidence of individual cognition.

In an AI-rich environment, a pause may represent independent thinking or consultation with an AI assistant. A revision may show reflection or acceptance of AI advice. A detailed process record may document human-AI collaboration rather than the work of an individual mind.

The issue is attribution.

Dr Kathryn Richardson’s Reconstructing Validity in the Age of AI: A Stage-Differentiated Framework for K–12 Assessment provides an important distinction between intended AI use, construct representation, and construct contamination.

The field needs to go further, however, because many proposed solutions still assume the limitations of today’s technology.

Assessment cannot be designed on the assumption that AI will remain difficult to access, unable to interpret context, unable to see or hear what the candidate sees and hears, or unable to guide work in real time.

The relevant design challenge is not current AI capability. It is whether the assessment will still support its intended claim as AI becomes more multimodal, capable, cheap, and embedded in everyday tools.

This also changes what we mean by human oversight.

In From Prompt to Publication: Designing Workflows That Make AI Item Generation Work, I argued that “human in the loop” is too weak a standard. A human reviewer can look at plausible AI output and approve it without genuinely engaging with the underlying task.

We need effort in the loop.

That means designing processes in which the human expert must do meaningful independent work: solve an item before seeing the AI-generated answer, make a judgment before seeing the model’s recommendation, explain disagreement, or test whether the evidence really supports the construct.

Otherwise, we risk automating the judgment and asking people to rubber-stamp it.

The systems represented at IAEA have chosen different places to begin. There is much to learn from each.

But every one of them will eventually have to answer the same question:

What evidence do we need before we can responsibly make a claim about this person?

Presentation references

Global Perspectives on Assessment Modernization: Balancing Technological Innovation with Public Trust, Anand Karat, Confidence Dikgole, Dan Koenig, Dr Mary Pitoniak, Dr Naveed Yousuf, Dr Wayne Wesley, and Pascal Couture, IAEA 2026.

Digital Assessment Beyond Platforms: Building Low-Bandwidth Systems for Holistic Learner Data at Scale, Crysty Swift and Dr Bridget Mthembu, IAEA 2026.

Kazakhstan’s Journey in Reforming University Admissions for the 21st Century, Nurseit Baizhanov and Terran Brown, IAEA 2026.

The Italian Approach to Accessible and Equitable Communication of Assessment Data, Dr Patrizia Falzetti and Michele Cardone, IAEA 2026.

Promoting Fairness for Assessment of English Writing with Small Language Models, Dr Shafiza Mohamed and Paul Edelblut, IAEA 2026.

When Take-Home Assessment Loses Trust: Australian Evidence on AI and Supervised Digital Exams, Alex Dore and Monica O’Brien, IAEA 2026.

Evaluating Examination Access Arrangements for Candidates with Visual Impairments in the West African Senior School Certificate Examination in Ghana, Wendy Enyonam Addy-Lamptey, IAEA 2026.

Navigating the Landscape of Multiple Data Sources for Educational Reporting in LMICs: The Case of Pakistan, Saima Maheen, IAEA 2026.

Reimagining Assessment in the AI Era: From Assessment Triangle to Assessment Ecology, Dr Eunice Jang, IAEA 2026.

Reconstructing Validity in the Age of AI: A Stage-Differentiated Framework for K–12 Assessment, Dr Kathryn Richardson, IAEA 2026.

From Prompt to Publication: Designing Workflows That Make AI Item Generation Work, Tim Burnett, IAEA 2026.