An item bank is a structured store of test items along with the metadata required to assemble valid test forms from them: item statistics, calibration parameters, content tags, format type, exposure history, and provenance. A bank is not just a list of questions; it is a managed asset whose value comes from what is recorded about each item.
The discipline emerged from psychometrics in the 1970s alongside Item Response Theory (IRT) and computer-adaptive testing, both of which depend on stable, sample-independent item parameters that only a properly maintained bank can provide. Cambridge English, ETS, the IELTS partnership, and most national high-stakes testing programmes now run on item banks.
Operational item banks tag each item with most or all of the following:
Without these fields a bank is a corpus of questions, not a testing asset.
Three properties depend on a bank: parallel-form equivalence, longitudinal score comparability, and adaptive testing.
Parallel forms must cover the same construct at the same difficulty. This is achievable only when the bank carries enough calibrated items per cell of the test specification to assemble each form to spec. Longitudinal comparability — saying that a Band 7 in 2019 means the same as a Band 7 in 2026 — is what calibration data plus secure anchor items deliver. Adaptive testing is impossible without a deeply calibrated bank because the algorithm needs known item parameters to choose the next item conditioned on the candidate's running ability estimate.
The bottleneck in AI-assisted item development is not generation; it is calibration and tagging. A generator can produce thousands of plausible items overnight, but each item entering the bank needs construct tags, format tags, and either a pre-test calibration or a confidence-weighted prior on its parameters. Pipelines that skip these steps produce a corpus, not a bank, and will not support the test forms the developer wants to assemble.