← Resources

Kirkpatrick's 4 Levels — and Where the Model Breaks Down

5 June 2026 · 9 min read · SessionData

For seventy years, almost every conversation about evaluating training has started in the same place: Donald Kirkpatrick's four levels. Reaction, learning, behaviour, results. It is the most widely taught model in the field, and for good reason — it gave L&D a shared language at a time when there wasn't one.

But "most widely used" is not the same as "best used." In practice the model is applied in a way its own structure quietly encourages: everyone measures the bottom rung, a few measure the second, and almost nobody reaches the top. This guide looks honestly at where the Kirkpatrick model holds up, where it breaks down, and what to reach for instead.

The four levels, briefly

The logic is sound: each level is more meaningful than the last. The trouble is what happens when real teams, with real time constraints, try to use it.

Where the model breaks down

The academic literature on Kirkpatrick is unusually consistent. Three criticisms come up again and again.

1. It collapses to the bottom two levels. Because Level 1 is trivial to collect and Level 4 is hard, evaluation in practice gathers at the shallow end. Reviews of how the model is actually used find most evaluators stop at reaction and learning — exactly the levels that say least about whether the training worked. The model doesn't force this, but its shape rewards it.

2. It assumes a causal chain it never proves. The four levels imply a ladder: happy learners learn more, learning changes behaviour, behaviour drives results. Decades of evidence show those links are weak. Satisfaction correlates poorly with learning; learning correlates poorly with transfer. Treating the levels as a causal sequence — "we got high reaction scores, so results will follow" — is one of the most common and most expensive mistakes in L&D.

3. It is silent on cost. Kirkpatrick measures outcomes but never weighs them against investment. It tells you whether something happened, not whether it was worth doing. That gap is exactly why a fifth level had to be bolted on.

The alternatives worth knowing

Several models exist precisely to patch these weaknesses:

• Phillips ROI Methodology adds a fifth level — return on investment — converting Level 4 results into monetary value and comparing them to cost. It answers the "was it worth it?" question Kirkpatrick ducks, though it inherits the hard problem of attributing money to learning.
• Kaufman's Five Levels widen the lens to include societal and organisational impact ("mega" and "micro" levels), arguing that training should be judged by its value to the whole system, not just the learner.
• Brinkerhoff's Success Case Method flips the approach: instead of averaging everyone, it studies the most and least successful cases to understand why training transfers — or doesn't.
• Thalheimer's Learning-Transfer Evaluation Model (LTEM) is the sharpest modern critique. Its eight tiers replace "reaction" with measures of whether learners can actually do the thing later, deliberately demoting satisfaction in favour of demonstrated capability and transfer.

The common thread across all four: push past satisfaction, and take behaviour change seriously as the thing that matters most.

The real lesson: effectiveness is a stack, not a number

Step back and the models agree more than they disagree. Reaction is cheap and weak. Behaviour is expensive and strong. A good evaluation doesn't pick one level — it measures across the whole stack, calibrated to the kind of learning.

That is the principle behind the way we think about it at SessionData. Rather than collapse everything into one figure, we measure across the whole stack and calibrate what counts as credible to the kind of learning — so a session full of delighted learners who change nothing scores poorly, which is exactly right. (For a fuller treatment, see our guide to measuring training effectiveness.)

Be honest about what you can evidence

Here is the part most models skip. Whichever framework you adopt, you will be able to evidence the lower levels well and the upper levels only partially. The disciplined response is not to fake the rest — it is to state plainly what your data supports, with its confidence, and what it does not.

A fabricated Level 4 result, or an ROI figure that can't survive a CFO's scrutiny, does more damage to L&D's standing than an honest "we have strong reaction data and early, moderate-confidence signal on behaviour." Calibration beats false precision every time.

Where this leaves you

Kirkpatrick isn't wrong — it's incomplete, and it's usually used at its weakest. Keep the shared language. Drop the assumption that the levels chain automatically. Measure the whole stack — calibrated to the kind of learning — with honesty about your certainty at each level.

That honest, calibrated view of effectiveness is what we call Return on Learning — and it's what SessionData is built to capture.