Search for the benefits of assessment and you will find the same list everywhere. It measures outcomes. It provides feedback. It encourages reflection. It supports personalisation. Most articles present the list as settled.
Most of it holds up. One item on it does not, and the research on that one has been sitting in plain sight since 1996.
This post covers what the evidence actually supports about assessing adult learners, which assessment type suits which purpose, and what changes when the assessment is a regulatory requirement rather than a design choice.
The Claim That Holds Up: Testing Changes What Learners Retain
The most useful finding in assessment research is that the act of being tested changes what the learner retains. Testing is not only a measuring instrument. It is a learning intervention.
Roediger and Karpicke demonstrated this in 2006. Learners who practiced retrieving material retained 61 percent of a passage after a week. Learners who spent the same time rereading it retained 40 percent.
One detail in that study matters more than the headline. On an immediate test, rereading wins. The retrieval advantage only appears on delayed tests. So a quiz that looks less effective when you measure right after the session is often the one producing better retention a week later, which is the point at which anybody cares.
The practical implication sits awkwardly against how most courses work. If assessment appears only at the end, you have used it purely as a measuring device and thrown away most of its instructional value. Spaced retrieval through a programme does more for retention than the same minutes spent presenting content again.
The Claim That Can Go Either Way: Feedback
Every list says assessment provides feedback and treats that as self-evidently good. The research does not agree with itself on this point, and the disagreement is more interesting than either side of it.
Kluger and DeNisi published the largest and most careful analysis in 1996, examining 607 effect sizes across 23,663 observations. They found feedback improved performance on average, at an effect size of 0.41. They also found that over one third of feedback interventions made performance worse.
That finding was surprising enough that they built a theory around explaining it. The pattern they identified is that feedback directed at the person, rather than at the task, consistently produced negative effects. Telling someone their answer was wrong and how to correct it helps. Telling someone they are struggling with this topic often does not.
Other analyses report considerably larger positive effects, in the range of 0.66 to 0.94. Those studies are not wrong, and I am not going to pretend the literature is settled when it is not. What Kluger and DeNisi contribute is the question nobody else asked: not how much feedback helps on average, but when it backfires. An average of 0.41 that conceals a third of cases going backwards is a very different design brief from a clean 0.8.
For anyone designing assessment, the practical rule that falls out of this is narrow and worth following. Feedback should describe the work, not the worker. It should tell the learner what to do differently, not how they are doing.
The Claim I Am Not Going to Cite
There is a number I could have used in this post and deliberately have not.
Black and Wiliam’s 1998 work on formative assessment is the most cited research in the field, and the effect sizes attached to it, usually given as 0.4 to 0.7, appear in almost every article on the topic. Those figures have not held up well. Randy Bennett’s critical review described the way the conclusions have been characterized as “the educational equivalent of urban legend.” Later meta-analyses of formative assessment have landed nearer 0.20.
Formative assessment is still worth doing, and the evidence for it is real. The specific numbers everyone quotes are not dependable enough to build a business case on, so I have left them out rather than pass them along.
I mention this because the same caution applies to most effect sizes that circulate in L&D. If a figure appears in every blog post on a subject and never with its confidence interval, it is usually doing rhetorical work rather than evidentiary work.
The Seven Types, and What Each One Is For
The type you choose determines which claim you can defend afterwards.
Formative assessment runs during learning to guide it. Quizzes, polls, discussion, short written reflection. Use it to adjust while there is still time to adjust.
Summative assessment runs at the end to judge the result. Final assessments, projects, certification. Use it when you need a defensible verdict.
Diagnostic assessment runs before learning to establish what people already know. Pre-tests, skills audits, structured interviews. Use it to avoid teaching what the audience has already mastered, which is the most common preventable waste in corporate training.
Self-assessment asks learners to judge their own progress. Useful for reflection and ownership. Weak as evidence, because self-report and performance correlate poorly, and that gap widens among the least skilled.
Criterion-referenced assessment measures performance against a fixed standard. Skills tests, performance checklists, portfolios. This is the right instrument when the question is whether someone meets the bar.
Norm-referenced assessment ranks a learner against peers. Useful for selection and calibration. Rarely the right choice in workplace learning, where the question is usually whether someone can do the job, not whether they can do it better than a colleague.
Ipsative assessment compares a learner to their own past performance. Strong for development and motivation. Not evidence of competence against any external standard.
Two of these produce evidence a stakeholder will accept, and they are criterion-referenced and summative. The rest serve the design and the learner. They do not prove capability to anyone outside the room.
When Assessment Is a Regulatory Deliverable
For a large part of the enterprise audience, the assessment question is not a design question. A regulator has already answered it.
In most states, earning insurance continuing education credit through self-study requires a passing examination score, typically 70 percent. Washington’s Office of the Insurance Commissioner is explicit: the student must complete the course before reaching the examination, and a score of 70 percent or greater is required to pass. That regulation also prescribes minimum question counts and limits how far a retake may repeat the first attempt.
This is worth reading carefully, because the common assumption is that compliance training only ever requires attendance. For insurance CE, that assumption is wrong. The regulator already demands criterion-referenced Level 2 evidence and sets the threshold.
FINRA sits differently, and lumping the two together is a mistake. Rule 1240(b)(2)(A) requires member firms to maintain a continuing and current education program, evaluate training needs at least annually, and develop a written training plan that accounts for regulatory developments. The obligation there is documented completion against a plan rather than a passing score.
The practical consequence for L&D is the same in both cases. Compliance systems already generate, store and audit this assessment data, and almost nobody surfaces it as evidence of learning. The reporting gap is not a measurement gap.
Choosing the Assessment
Four questions, in order, settle most of it.
What will people do differently? If you cannot describe the behaviour, no assessment type will rescue the design. Start there.
What claim do you need to defend afterwards? Coverage, capability or improvement. Coverage means completion. Capability means criterion-referenced. Improvement means a baseline and a repeat measure, which you have to plan before delivery rather than after.
Is there a regulatory floor? Check before designing. In several contexts the instrument and the pass mark are prescribed, and modifying a state-approved curriculum can require separate filing.
Where does retrieval fit? Given the testing effect, assessment belongs throughout the programme rather than only at the end. This is a design decision, not a reporting one.
What This Is Actually For
Assessment earns its place when it does two things at once: helps the learner retain and apply the material, and produces evidence someone outside L&D will accept.
Most assessment in corporate learning does neither. It arrives at the end, it measures recall, and it produces a completion record that answers a question nobody asked.
Fixing the instrument only gets you so far, though. The prior question is what evidence the program was supposed to produce in the first place, and that gets decided long before anyone writes a quiz. I have written separately about why most evaluation stops at Level 2, and what changes when you plan the measurement backward from the business result.
If you are working through what defensible assessment looks like across a program, the Work With Me page is where that conversation starts. The AI Readiness Checklist covers measurement readiness if a self-assessment is the better starting point.
Editor’s Note: This post was originally published in September 2023 and has been completely revamped and updated for accuracy and comprehensiveness. in September 2026.
