AI has made it incredibly easy to create assessments for your students, teachers, and school. Here are five reasons why you shouldn’t. Yet.
1. Measures are only as good as they are valid.
Have you ever received a test score and felt it was not an accurate reflection of what you know? Assessment scores are not inherently correct. Validity—the evidence that the scores from an assessment instrument represent the construct we claim to measure—has to be established through instrument development, pilot testing, and statistical evaluation.
When a large language model generates a “growth survey” in seconds, what it produces may or may not represent actual student growth. Assessment-based decisions about student placement, learning interventions, and funding allocation rest on the assumption that the instrument behind them is sound. If an AI-generated artifact only looks like a valid assessment, your educational decisions will be based on incorrect data.
Don’t take an AI tool’s word for its own competency at scoring student responses. Pull a sample of scored responses and have a human score them independently of the AI’s scores. If the two don’t agree closely, that means the tool isn’t ready to score on your behalf.
Before you trust a score, check that the assessment was built for what you’re actually using it for. Ask yourself: Does it align with what you taught? Is the wording appropriate for your students’ age and reading level? Would your students engage with it as designed? An AI tool can generate items that look right on-screen and still fail all three. And if you have only the scores from an AI tool and no supporting evidence that they measure what you think they do, reconsider using it in your classroom until you have that evidence.
2. AI systems inherit bias from the data that trained them.
In measurement science, a central concern in developing a good instrument is whether it measures what it is intended to measure across different subgroups. Large language models absorb the demographic and linguistic patterns embedded in their training data. When that system is prompted to build an assessment, the output does not treat every student the same. Any AI platform can, and currently does, systematically misjudge scoring multilingual learners, students with disabilities, and students whose response patterns diverge from the majority in the training data.
How do you know if your AI assessment tool is fair? Questions to ask include: What was this validated for and on whom? How was efficacy of the tool evaluated and was efficacy reported separately for student subgroups? Does the vendor state the limitations of their product and clarify what it should and should not be used for? Are the data from the system handled securely and responsibly?
3. AI-generated charts can show “growth” even when the data underneath don’t.
Legitimate growth measurement requires a stable metric across time points, comparable administration conditions, and an explicit model of change. AI tools can produce visually compelling growth trajectories from data that meet none of those conditions, conflating correlation, noise, or scaling artifacts with student gains. The same failure shows up in something as simple as a bar graph. A taller bar looks like a meaningful difference, but height says nothing about whether that difference exceeds what chance or measurement error would produce on its own and whether that difference would be large enough to have any practical meaning in the school context.
Reliability of an assessment means that if we were to conduct an assessment again, the inferences we make from it wouldn’t change in a meaningful way. It is important to ask whether a vendor has evaluated the tool for reliability; if they haven’t, you won’t know whether the scores or inferences you are making from them are anything more than random noise.
4. AI-generated interpretation of assessment data introduces a new, largely invisible source of measurement error.
It isn’t just the instrument producing assessment data that needs to be evaluated; the interpretations of the data must be validated. When a system summarizes assessment results or builds a dashboard from school data, it is making inferential judgments about which patterns are meaningful, which differences that cross a threshold are worth flagging, and what a score change implies about student learning.
What’s more, when practitioners routinely hand off interpretation to a system, they diminish their own data-interpretation skills. Brookings’ Global Task Force on AI in Education, which conducted a year of consultation with educators and families across 50 countries, found that offloading cognitive work to AI systems that simply supply answers erodes the underlying thinking process. The same dynamic applies when educators offload interpretive judgment to a system that is, by its very design, frequently confident and occasionally wrong. If it’s unacceptable for our students to offload their own thinking to AI when they’re being assessed, why is it acceptable for us to offload the thinking behind interpreting their results?
5. You don’t need AI to build an assessment.
The Buros Center for Testing, a research organization based in the University of Nebraska–Lincoln, publishes the Mental Measurements Yearbook, an independent, expert-reviewed guide for evaluating commercial assessments before adoption. State assessment offices also typically publish which locally selected, nationally recognized assessments meet the peer-review requirements of the Every Student Succeeds Act, along with the content standards and technical documentation to check for compatibility. Massachusetts and Texas are useful examples.
Right now, the Standards for Educational and Psychological Testing—the field’s authoritative guidance for over 60 years that are jointly issued by the National Council on Measurement in Education, the American Educational Research Association, and the American Psychological Association—is currently under revision specifically to address AI-integrated assessment. Schools building measurement infrastructure on today’s ad hoc AI use are building on a foundation the field itself is actively reconstructing.
Remember that you are likely a better expert to develop an assessment on the fly when needed in your classroom or school than AI is. AI will say it can do it and will be very confident about that, but we have more confidence in your expertise in the area that you teach than we have in AI.
