Ask whether AI can write a test question and the answer is plainly yes. Hand a model a curriculum point and a target difficulty and it returns a clean, professional-looking multiple-choice item in seconds. The question worth arguing about is whether that item belongs anywhere near a high-stakes exam.
In high-stakes assessment, an item has to survive expert review, fairness analysis, and field testing—and, if a candidate disputes a result, legal scrutiny. Very few generated items have ever faced any of that.
Why Assessment Programs Are Turning to AI
Good test items are slow and expensive to produce. A single operational item for a credentialing exam can cost into the four figures once you account for authoring, review, field testing, and psychometric calibration, though costs vary by item type and review process. Banks also wear out. Once candidates start passing questions around, those questions stop measuring what they were built to, so banks need constant replenishment. Mathematics has been the most common entry point—stems can be precisely specified, wrong-answer patterns are predictable, and the subject lends itself to automated generation more readily than constructed-response or clinical reasoning items.
AI helps most at the front of that pipeline. A model can draft a structurally complete multiple-choice question in seconds. For practice banks and formative quizzes that’s genuinely useful, which is why many item writers have quietly folded ChatGPT into their workflow ahead of any formal policy, often alongside the digital assessment software they already use.
Trouble starts when that informal, unstructured use drifts toward high-stakes work and nobody stops to ask whether the items are good enough.
When Polished Items Are Still Wrong
A badly written human item usually gives itself away—clumsy phrasing, a nonsensical distractor, a stem that contradicts itself. AI gets the surface right and the substance wrong. The options look balanced, the rationale reads with confidence, and a reviewer moving at the pace most programs can afford waves it through.
Patrick Plichart, co-founder of Open Assessment Technologies (OAT), showed one at a webinar he ran with the Luxembourg Institute of Science and Technology (LIST). He’d asked a model to build an item for young children on reading an analog clock and ordering daily events. It looked finished—until you looked closer. The clock faces were ambiguous, you could answer without reading a clock at all, and the pictures showed a girl doing the laundry while a boy rode a bike. That last detail isn’t harmless. Items like it have drawn formal complaints from teaching unions in France, forcing tests to be pulled and rewritten. The model wasn’t malicious—it reached for the most statistically obvious picture of who does the laundry, straight from its training data. The lesson is not just that the item contained several flaws. It is that those flaws were easy to miss because the item still looked publishable.
It shows up in the research too. A blinded study in Medical Teacher (2025) set GPT-4 against human authors on clinically complex MCQs, scored by a panel blind to which was which. The AI items held their own on structure. But the model also produced fabricated references—citations to sources that don’t exist, a classic sign of LLM hallucination that surfaced only on close inspection.
Bias is harder to catch, because volume hides it. Researchers at Carnegie Mellon and the University of Hong Kong ran a bias-detection method over AI-generated educational material in 2025 and found gender, racial, and national stereotypes inside otherwise fluent text. A single item writer might let one through occasionally; a model reproduces the same stereotype across hundreds of items, and no review team will catch every instance by hand.
Why Trust Depends on Validation, Not Generation
Not all the research is bad. Some of it is surprisingly strong. A 2025 large-scale field study ran AI-generated exams across 91 university classes with close to 1,700 students; by item response theory the AI questions were slightly easier than the expert-written comparisons but discriminated between students a little better. A preregistered, blinded study in npj Digital Medicine reached the same verdict on medical MCQs, with neither students nor physicians able to tell AI and human items apart.
Those findings are encouraging, but they should not be treated as proof that any AI-generated item is ready for operational use. Both studies used carefully engineered pipelines—cycles of generation, critique, revision, and expert screening—not a single prompt. If an assessment program uses a simpler process, the headline finding does not carry over.
There is also a more serious evidence gap around fairness. Published evidence on differential item functioning (DIF) in AI-generated items remains thin. DIF analysis checks whether an item behaves differently for equally able students from different groups, and the AERA/APA/NCME. Standards make subgroup fairness evidence part of the validity argument for operational use. Maeda (2025) showed that DIF can be partly predicted from item text using transformer models, which could make fairness testing more affordable at scale—but for now, the field has spent much of the past two years debating quality while fairness evidence has not kept pace.
When evaluating an item-generation tool, one question cuts through the marketing: do the quality numbers come from raw generator output, or from a full review-and-revision pipeline? The answer changes what every other figure means.
Separate the Generator From the Judge
Ask a model to mark its own homework and it tends to miss exactly what it got wrong—the blind spot that produced the error also conceals it. Self-review by the generating model buys little.
A more credible pattern splits the work. One model writes the item; independent models judge it against explicit criteria—is the key correct, are the distractors plausible, is there bias, does the difficulty match the spec? When they disagree, a vote settles it. Plichart argues for this split, and there’s supporting evidence that agreement across several independent models beats trusting any one of them.
OAT and LIST have built this into TAO Studio, where every item is scored against eleven quality traits by independent model judges, with arbitration between them, before a human reviewer sees it. The same framework can also be useful for existing expert-written item banks, where it may surface problems that were missed because those banks were never screened this systematically. It redraws the line. What matters isn’t whether a human or a model wrote the item—it’s whether anyone validated it properly. Plenty of human-written items sit on the wrong side of that divide.
Because TAO is built on open standards—QTI for the items, LTI for the integrations—that quality layer isn’t welded to one supplier. A program can adopt multi-model checking without rebuilding its whole assessment stack around a single vendor.
Provenance, Sign-Off, and What Happens to Rejects
A validated item still needs a paper trail. If a candidate challenges a score, a program must show where the item came from, what checks it passed, what a human changed, and who signed it off.
Governance is catching up. In 2024 the Educational Testing Service (ETS) set out five principles for responsible AI in assessment, anchored in the AERA/APA/NCME Standards. At its core, a person—not the model—is responsible for any item that reaches a scored exam, however it was drafted. In December 2025 the 1EdTech Consortium published guidance that puts high-stakes assessment in its top risk tier and recommends carrying provenance metadata through standards like QTI, so an item’s origin travels with it between systems.
Keeping a human in the loop at every stage is the direction OAT and LIST are building toward. Next, they plan to close the feedback loop—edits, accept-and-reject decisions, and live item statistics fed back into the generation and auditing models. Polymorphic testing, where items are dynamically varied to limit exposure while keeping measurement consistent, adds another reason why quality controls at the item level need to be rigorous.
One useful concept from OAT and LIST is the negative item bank. When a generated item fails review, many teams discard it, which means the failure teaches little. Keeping rejected items instead—tagged by what went wrong, whether a wrong key, a gender stereotype, weak distractors, or difficulty drift—turns them into quality-improvement data. Assessment teams can see where the generator keeps failing, tighten validators, and catch similar failures earlier in the workflow.
OAT’s wider partnership with LIST—the €4.2 million project funded mainly by OAT—is the vehicle for wiring that field data back in.
Trust Comes From the Workflow, Not the Generator
So, can AI item generators be trusted in high-stakes assessment? On the strength of generation alone, no—and a fluent item with four tidy options doesn’t earn a different answer just because a model can produce ten thousand more overnight.
What earns trust is everything around the generator: independent checks, a named person accountable for what ships, field data on how items behaved, and a record that reconstructs an item’s journey from draft to delivery. If a nursing board faces a legal challenge, it needs to trace how the disputed item reached the candidate’s screen—that traceability is what this kind of pipeline makes possible. Programs getting this right can show their working for every item they keep, and in an open-source, standards-based ecosystem, that working is portable.
Explore TAO’s AI Quality Framework
TAO’s work with LIST brings multi-model validation, trait-based quality scoring, and standards-based interoperability into a single assessment workflow. To see what quality-first AI item generation looks like in practice, schedule a demo.
