Adaptive and psychometric AI uses machine learning and related AI methods in test routing, trait estimation, candidate selection, and measurement models. In assessment terms, it is less about productivity and more about whether AI-supported measurement still produces valid, fair, and explainable decisions about ability or competence. The evidence base shows clear research and market momentum, but much less independent assurance about high-stakes operational use. The main question for assessment teams is whether the model measures the intended construct and can be governed, explained, and challenged in live use.

Adaptive and psychometric AI refers to AI used inside measurement systems: changing what a candidate sees next, estimating ability from observed performance, or supporting decisions in practical and selection-based assessment. The core issue is whether AI changes the measurement in ways that preserve validity, reliability, fairness, and transparency. Examples already visible in the field include deep computerised adaptive testing, sector-specific pilot assessment platforms, AI-supported driving licence exams, and multistage adaptive testing that blends statistical design with newer AI and machine learning methods.

This is closer to assessment infrastructure than to general AI assistance. If AI changes item routing, scoring, observation, or interpretation, it can alter validity, reliability, bias, candidate experience, and appeal routes. It also raises a governance question: what counts as defensible evidence when AI contributes to an assessment decision?

- **Adaptive testing**: the test changes in response to a candidate’s answers. - **Trait estimation**: the model estimates ability or competence from observed performance. - **Psychometric validity**: whether the assessment measures what it claims to measure. - **Operational governance**: whether routing, scoring, monitoring, and exception handling can be explained and controlled. - **Candidate recourse**: whether a candidate can understand and challenge an AI-supported judgement.

The source set strongly suggests that AI in assessment is moving into the measurement layer rather than remaining a bolt-on tool. There is broad convergence that adaptive testing is an established assessment approach, and that AI can extend it into more complex routing, modelling, and decision support. The strongest evidence here comes from the academic literature and practical assessment tradition, while vendor and reporting material mainly shows market direction and early adoption. The shared practical view is that these systems need explicit validation, not just technical sophistication.

The open question is how much trust can be placed in AI-supported adaptive or psychometric systems in high-stakes settings. The evidence supports the method’s development, but it does not by itself establish routine validity, fairness, or explainability in live operations. Vendor material often presents AI-enabled assessment as deployment-ready, but that is better read as a market signal than independent validation. The unresolved issue is not whether AI can be used here, but what level of assurance is needed before it is acceptable for licensing, certification, or other consequential decisions.

- Weak construct validity if the model measures proxy behaviour rather than the intended competence. - Reliability problems if routing or scoring varies in ways that are difficult to explain. - Bias or accessibility issues if performance differs unfairly across candidate groups. - Over-reliance on vendor claims where independent evidence is limited. - Poor appeal routes if AI-supported decisions are opaque. - Operational risk where monitoring tools blur the line between observation, assistance, and decision-making.

1. Define what must be measured unaided and what may be adapted by the system. 2. Check whether the AI approach measures the intended construct, not just correlated behaviour. 3. Test performance across relevant candidate groups, including accessibility and fairness impacts. 4. Put routing, scoring, monitoring, and exception handling under clear human governance. 5. Require evidence on drift, calibration, and change control before live use. 6. Make candidate challenge routes understandable and workable. For practical tests, ask whether AI-supported observation can distinguish competence from noise, context, or sensor error.

| Option | What it means | Main strength | Main concern | |---|---|---|---| | Traditional fixed-form testing | Everyone gets the same test | Simple to explain and govern | Less efficient measurement for some candidates | | Adaptive testing without advanced AI | Item selection changes by response pattern | Established psychometric tradition | Still needs strong calibration and monitoring | | AI-supported adaptive or psychometric assessment | AI helps route, estimate, or interpret performance | More flexible measurement and operational efficiency | Harder to validate, explain, and challenge | The deeper assessment issue is not which option sounds most advanced, but which one best matches the intended construct, risk level, and governance capacity.

A licensing body wants to digitise a practical driving exam. If AI is used to support observation or scoring, the body needs to show that the system distinguishes safe driving competence from sensor noise, weather conditions, or route variation. A vendor demonstration alone would not answer that question; validation evidence and appeal handling would matter more. A further example is multistage adaptive testing in language assessment: the attraction is not just faster item selection, but the chance to combine psychometric control with AI and machine learning methods in a design that remains explainable and secure enough for live use.

- Scholarly article on deep computerised adaptive testing. - Cambridge Assessment Network blog on computer adaptive testing. - Vendor description of the PACE pilot assessment platform. - Reporting on AI oversight in Morocco driving licence exams. - Duolingo English Test blog on multistage adaptive testing and AI/ML.

Vendor material shows AI being packaged for selection, simulation, and sector-specific practical assessment. These offerings are useful market signals, but they do not by themselves settle the validity, fairness, or readiness questions for high-stakes use. Assessment teams should treat supplier claims as prompts for evidence requests rather than proof.

### What is adaptive and psychometric AI in assessment? It is AI used in test adaptation, trait estimation, candidate selection, and measurement models, where the system helps determine which items are shown and how performance is interpreted. ### Why does AI-assisted adaptive testing matter in exams or certification? Because it can affect validity, reliability, fairness, candidate experience, and the transparency of decisions in high-stakes assessment. ### Is AI-enabled adaptive testing already being used? The source set suggests active development, vendor packaging, and emerging public-sector use, but evidence quality varies and does not by itself prove that each use is psychometrically sound. ### What should assessment teams ask suppliers? Ask for evidence on model validity, reliability, bias across candidate groups, monitoring of drift, explainability, and candidate challenge routes.

Tim Burnett (Admin)

Test Community Network. "Adaptive and psychometric AI." TCN AI & Assessment Wiki. Last reviewed 2026-05-03. https://www.testcommunity.network/wiki/adaptive-and-psychometric-ai.html

- Psychometrika article on deep computerised adaptive testing. - Cambridge Assessment Network blog on computer adaptive testing. - PACE pilot assessment platform description. - Report on Morocco driving licence AI oversight. - Duolingo English Test blog on multistage adaptive testing.