Did you know? Two experiences can have different total durations and average quality, yet a person may remember the one with a better ending more favourably. The peak-end rule describes a recurring pattern in retrospective judgement: some experiences are remembered disproportionately through their emotionally salient peak and ending rather than through an equal-weight average of every moment. It is a research finding about particular kinds of memory and evaluation, not a universal equation that can predict everybody’s preferences.
The practical lesson for a learner, service designer or leader is to separate what happened moment by moment, how long it lasted, and what someone remembers afterward. Measure these outcomes separately instead of assuming a glowing final impression proves the entire experience was good. Use clear endings to support real understanding and humane closure, not to disguise unsafe or unfair treatment earlier in an experience.
This article explains the psychological mechanism, evaluates the original cold-water experiments, works through invented numerical examples, identifies ethical failure modes and shows how the idea transfers to school learning and organisational decisions. The examples are illustrations, not reports of real students or customer experiments.
Contents
- Experienced moments and remembered episodes are different measurements
- The cold-water study and the meaning of its result
- Salience, endings and the reconstruction of memory
- Worked example: a better ending is not a better average
- Duration neglect is not permission to increase harm
- How to measure momentary, remembered and chosen outcomes
- Lessons, examinations and the danger of a pleasing final question
- Service journeys: measure delays and errors, not just the closing smile
- Remembered experience is not a dependable forecast of the next episode
- Ethical endings are not techniques for hiding harm
- Eight interpretation errors and a repair method for each
- Transfer exercise: a two-day learning event with different endings
- What the research does not prove across all people and situations
- Frequently asked questions
- Practical reader routes and the original research
Experienced moments and remembered episodes are different measurements
A person’s experience unfolds through many moments. One period may be comfortable, another frustrating and a final period reassuring. Afterward the person produces a global judgement such as ‘That lesson was useful’ or ‘That service was dreadful.’ The global report is not a continuous recording of every second. Memory is selective, reconstructive and influenced by features that stand out. People can be sincere in both their momentary reactions and their later summaries while producing different evaluations.
Research on remembered utility distinguishes moment-to-moment experienced wellbeing from retrospective appraisal of an episode. For a person deciding whether to repeat an experience, the remembered evaluation may matter greatly even if the episode’s overall duration or moment-by-moment ratings tell a different story. This does not mean the remembered judgement is automatically false; it means it answers a different psychological question.
The peak-end pattern identifies two potential anchors. One is a particularly intense or emotionally salient point; the other is how the episode ends. For an unpleasant episode, the peak may mean its worst moment, while for a positive event a particularly good moment may carry more weight. Researchers must specify which intensity dimension they measure, otherwise a story about ‘the peak’ is ambiguous.
Duration neglect is a related tendency for the length of an episode to receive less weight in some global retrospective evaluations than an equal-time average would imply. This does not make duration unimportant for welfare, planning or safety. A longer painful procedure still contains more minutes of exposure. The human experience during that time remains consequential even if a later preference task does not weight every minute equally.
Use three separate questions: how did the person feel during the episode, what did they remember afterward, and which future option did they choose? These need not yield the same ranking. A responsible institution does not let one favourable recollection replace evidence about what took place across the entire session.
The cold-water study and the meaning of its result
The original paper When More Pain Is Preferred to Less: Adding a Better End, published in Psychological Science in November 1993 by Kahneman, Fredrickson, Schreiber and Redelmeier, tested retrospective choices after controlled aversive experiences. Participants experienced a shorter trial with a hand immersed at about 14 degrees Celsius for sixty seconds and a longer trial that added thirty seconds during which the water warmed gradually to about 15 degrees Celsius.
The extra time was still unpleasant for most participants, but less aversive than the initial portion. A significant majority chose to repeat the longer trial even though it added more exposure. The result is informative because the longer experience was not made shorter or free of unpleasant sensation. The better ending apparently changed remembered evaluation enough to influence choice for many studied participants.
This finding should be described carefully. It was a particular experiment with particular participants, temperatures, trial procedures and choice outcomes. It does not prove that people universally prefer more pain or that every unpleasant event is judged primarily by a peak-end calculation. The paper supports a specific contrast between duration and remembered evaluation under the studied conditions.
The publisher’s primary article record and Association for Psychological Science record allow a reader to inspect the original methods and conclusions. A teaching guide should point to this original evidence rather than attributing unlimited power to a slogan heard in a business presentation.
It is also possible that a participant’s choice reflects the particular way options were represented or remembered. Replications and boundary studies matter. A school lesson, a medical procedure and an online checkout differ enormously in stakes and meaning. The general mechanism is a useful hypothesis to examine, not a licence to treat those situations as experimentally interchangeable.
Salience, endings and the reconstruction of memory
An ongoing experience provides a stream of signals, but a memory report compresses that stream into a narrative. Moments that stand out, changes near a boundary and the event’s final state may be especially available when someone later answers an overall evaluation question. A global score formed in this way need not equal the arithmetic mean of every momentary rating.
A clear ending can help people interpret what happened. In a class, a final independent example may connect confusing intermediate work to a coherent method. In a service interaction, a final explanation may give a person a clear next action. These mechanisms involve meaning and closure as well as emotional intensity. It would be a mistake to infer that changing only the last thirty seconds always changes long-term wellbeing.
Peak and end are not the only determinants of memory. Expectations, perceived fairness, attention, personal goals, prior experience, narrative interpretation and opportunities to revise a judgement can all matter. People may also remember the beginning, the frequency of difficulties and distinctive interruptions. A broad psychological claim requires evidence about these alternatives rather than a confident two-factor story.
Temporal measurement is essential. A rating collected during an activity is not the same as one collected immediately after or a month later. The questions may ask about different features: momentary comfort, recollected satisfaction, willingness to repeat, total burden or learning retained. Each can reasonably change across time. An honest evaluator preserves those labels.
The broader lesson for intelligence is that human retrospective judgement is a compressed representation of experience. Compression is useful, because nobody can store every moment with equal fidelity, but it may introduce predictable distortion. Better decisions require knowing whether a compressed memory or a full time-based record is the relevant evidence.
Worked example: a better ending is not a better average
Consider two explicitly fictional twelve-point and ten-point satisfaction sequences. The scores range from zero to ten, with larger numbers representing better momentary experience. Episode A lasts ten equal one-minute intervals. The first nine receive score five and the final interval receives score one. Its time-weighted average is (9×5+1×1)/10=46/10=4.6.
Episode B lasts twelve equal one-minute intervals. The first nine score four and the last three score five. Its average is (9×4+3×5)/12=(36+15)/12=51/12=4.25. Under the equal-duration momentary rating average, episode B is worse than A by 4.6−4.25=0.35 points, and it also lasts two minutes longer. These are constructed values and cannot prove how real people will remember them.
A simple illustrative peak-end calculation for positive satisfaction might average the maximum score and the end score. Episode A’s maximum is five and its end is one, giving (5+1)/2=3. Episode B’s maximum is five and its end is five, giving (5+5)/2=5. The illustrative heuristic therefore prefers B even though B has the lower experienced average.
That pair of calculations demonstrates a possible ranking reversal, not a universal quantitative theory of memory. Real peak-end effects are not defined by a rule requiring every person’s recollection to equal the arithmetic mean of two numbers. The psychological process, scales, timing and situation can produce other weights or other influences.
A careful analyst would display all the values: Episode A averaged 4.6 across ten minutes and has a toy peak-end score of three; Episode B averaged 4.25 across twelve minutes and has a toy peak-end score of five. The exact denominators and operational definitions prevent a marketing slogan from hiding the distinction between experienced quality and remembered impression.
Duration neglect is not permission to increase harm
The original experiment can tempt an oversimplified lesson: if an event is unpleasant, add a milder ending so that people remember it better. That interpretation misses the ethical question. People still experienced additional discomfort in the longer trial. A healthcare provider, educator or employer should not add unnecessary suffering merely to improve a memory score or a recommendation metric.
Duration is relevant when exposure has direct physical, financial or opportunity costs. An extra ten minutes of waiting can make someone miss transport, work or childcare. A treatment’s total discomfort may matter clinically. A school lesson’s extra length may create fatigue. An institution cannot justify these costs by pointing to a favourable final survey response.
Peak-end effects are most useful as warnings about measurement. A survey at the exit may overrepresent the final interaction compared with earlier delays or problems. A question about willingness to return may differ from an audit of harm, accessibility or service completion. Use multiple measures rather than choosing the one that flatters the organisation.
Sometimes a better ending is genuinely valuable without extra cost. A teacher can reserve a clear recap and a changed-context question; a service desk can explain what will happen next after resolving an issue; a team can close a difficult meeting with a precise decision record. These changes improve information and respect, rather than manipulate memory through needless extensions.
The aim is better actual experience and more accurate evidence. If users report good memories while objective records show long delays or unresolved errors, both facts belong in the evaluation. Neither should be discarded to produce a single reassuring headline.
How to measure momentary, remembered and chosen outcomes
A study of experience needs a declared unit of observation. Is the unit one person, one interaction, one school session or one minute? Does a person submit multiple ratings? Are some activities longer than others? A time-weighted mean of momentary scores answers a different question from a survey collected once after the event. The denominator is central to interpretation.
Suppose a fictional service collects four equally spaced in-session scores from each of fifty independent visits. That yields up to 200 momentary observations, but not 200 independent visitors. Analyses that pretend every rating belongs to a different person may understate uncertainty. Preserve visit identifiers and consider within-visit dependence. A later recall score is another variable attached to those same visits.
Ask separate questions, with neutral wording. A momentary prompt may ask ‘How difficult is this step right now?’ A post-session question may ask ‘How useful was the overall session?’ A later choice question may ask which method the learner would prefer to use again. The questions need not correlate perfectly because they assess different constructs.
Nonresponse can distort the result. People with severe problems may abandon the process before its closing survey, making a glowing ending score especially selective. Those who remain may differ from those who left. Record the invitation count, actual response count, abandonment and the timing of collection. Do not call a survey representative merely because many answers were received.
An institution should also track objective outcomes: task completion, error correction, actual wait time, accessibility barriers or independently tested learning. A pleasing remembered end cannot substitute for proof that the underlying job was accomplished.
Lessons, examinations and the danger of a pleasing final question
A teacher may finish a challenging lesson with a clean demonstration that feels rewarding. Students leave with the impression that the topic is now clear. That remembered impression has value: coherence can support motivation and an accurate summary can organise earlier work. But a successful demonstration by the teacher is not proof that learners can reproduce the method unaided on a changed question.
Make learning evidence independent of mood. Near the end of a lesson, invite students to work a short fresh example and explain the step on which the method depends. The task should not merely copy the last worked answer. Keep the result and its marking criteria separate from an exit survey about confidence. A student can feel positive yet misunderstand the core relation; another may feel frustrated after productive work and still have learned a great deal.
Suppose twenty-four fictional pupils complete an exit check. Eighteen solve the transfer problem and six do not. The correct-response fraction is 18/24=75%. If twenty-one pupils also report a positive final impression, that share is 21/24=87.5%. The two figures do not contradict one another because they measure different outcomes. We cannot derive which particular positive pupils failed the problem without individual-level joint data.
A parent may remember that the child came home smiling after a tutoring session. That is a useful observation about the session’s end, not a substitute for inspecting corrections, prerequisite understanding and later recall. A tutor should value respectful closure while still reporting unresolved gaps clearly. It would be manipulative to hide a serious learning problem merely to preserve a good final feeling.
For a broader framework, the verified learning that transfers into life article explains why independent application matters. The point is not to engineer a beautiful ending in place of learning, but to make a good ending the truthful conclusion of worthwhile learning.
Service journeys: measure delays and errors, not just the closing smile
Customer journeys contain moments with unequal stakes. A person may wait a long time, struggle with an inaccessible form and finally speak with a helpful employee who resolves the immediate issue. The last interaction may dominate what they recall in a later satisfaction survey. An organisation can legitimately improve that final explanation, but must not treat a high survey score as evidence that the earlier barriers were acceptable.
Track operational outcomes alongside recalled impressions. These might include total waiting minutes, abandoned applications, unresolved complaints, accessibility failures, repeat contacts and independent confirmation that the person’s requested action was completed. The same session can receive a favourable memory rating while still violating a measurable service requirement.
Consider a fictional appointment service with 100 completed visits and 30 abandoned visits. If only completed visitors are offered a closing survey, a 90% positive score among eighty respondents means 72 positive responses out of eighty, not 90% of all 130 people who started the process. The sample coverage is 80/130≈61.54%, and 50 people are not represented in those responses. Without evidence about their experience, generalising is unjustified.
Better service design begins with fixing the long wait and the inaccessible form. At the end, a concise summary of what was done, what remains pending and where to find help provides genuinely useful closure. Such an improvement may also support favourable memory, but its purpose is truthful orientation and the person’s wellbeing.
This is especially important in health, education, public services and any process where a person’s rights or safety can be affected. The remembered story is one kind of evidence; the documented journey and its consequences are others. They should be analysed together, not used to cancel one another.
Remembered experience is not a dependable forecast of the next episode
A person’s memory of an earlier event often informs what they expect from the next one. If the remembered ending was excellent, they may confidently choose to repeat the activity despite significant earlier effort or discomfort. This can be a reasonable preference for some people, but it does not establish that the next event will have identical moments, risks or outcomes.
The psychological literature on affective forecasting studies how people anticipate future emotional states. That field is adjacent to, but not identical with, the peak-end pattern: forecasting concerns imagined future feelings, while the peak-end rule concerns evaluation of remembered episodes. The published affective forecasting guide examines that distinction in greater depth.
Suppose a student fondly recalls a revision workshop because the last twenty minutes included a satisfying solution. Before signing up for a new workshop, the student should still ask about content, prerequisites, total duration and whether there will be independent practice. A favourable memory is valuable personal information, but it is not a syllabus or a guarantee of future learning.
Managers should be equally cautious with customer return intentions. A high likelihood-to-return rating may reflect recent relief rather than the cost and quality of the whole journey. Compare expressed intentions with later observed choices where doing so is ethically and legally appropriate. Retain denominators and explain who had an opportunity to return.
The practical repair is to separate ‘I remember liking it’, ‘I expect to like it again’, and ‘I would choose it after considering all consequences’. Different evidence belongs to each statement. Intelligent decision-making becomes more accurate when that difference is made explicit.
Ethical endings are not techniques for hiding harm
Knowledge of cognitive biases can be misused. An organisation could attempt to schedule an enjoyable final moment while leaving excessive delay, unwanted pressure or unsafe treatment unchanged. Even if a closing survey improves, that would not make the preceding harm acceptable. An ethical programme aims to improve the experience itself and the accuracy of evaluation, rather than exploiting predictable features of memory.
Use the peak-end effect defensively. Ask whether your evaluation method neglects long periods of difficulty, whether people who left early are absent from the sample, and whether a new policy improves objective outcomes as well as final impressions. If the two move in opposite directions, investigate before claiming success.
In a classroom, do not manufacture fear or public embarrassment to make the later relief more memorable. A respectful final explanation should help learners understand and leave with clear next actions. In a clinical setting, unnecessary pain and exposure have independent significance regardless of retrospective preference. In an employment setting, last-minute praise cannot erase systemic problems with working conditions.
The original cold-water experiment was conducted under a controlled research protocol; it is not a general instruction to add unpleasant moments to real service processes. The research illustrates a discrepancy in retrospective choice under a specific design. It should inform careful measurement and ethical design, not be promoted as a universal playbook for persuasion.
Ethical transparency also requires giving users an opportunity to report poor parts of an experience without being pushed toward a single global rating. Separate prompts about waiting, comprehension, respect and unresolved problems can support repair. Always avoid exposing individual private complaints unnecessarily, especially when the person depends on the institution.
Eight interpretation errors and a repair method for each
Error one is reporting a peak-end score as the average experienced quality. Repair it by labelling the toy retrospective heuristic separately from the actual time-weighted mean. Error two is forgetting unequal durations: a twenty-minute rating should not receive the same time weight as a one-minute rating unless that is the declared measure. Error three is assuming every person remembers through the exact arithmetic average of two moments. The empirical effect is not a deterministic law.
Error four is mixing memory gathered immediately after an event with memory gathered weeks later without distinguishing the elapsed time. Repair by recording collection date and analysing changing interpretations. Error five is inferring that a good ending proves a successful task. Repair by collecting direct evidence of service completion or independent learning.
Error six is excluding people who abandoned the process. Their absence can make surviving participants seem unusually satisfied. Repair with a denominator that includes all eligible starts and with honest nonresponse analysis. Error seven is turning a research result in controlled pain trials into a universal forecast for education, employment and customer services. Repair by respecting domain boundaries and testing the relevant hypothesis in the actual setting.
Error eight is using a final survey score to rationalise avoidable harm. Repair by measuring real costs and outcomes, upholding ethical standards and giving people effective recourse. No bias effect authorises a designer to lower objective welfare for the sake of a better retrospective impression.
Across these errors, the shared discipline is simple but demanding: define what happened, what was remembered, who responded, and what decision followed. Keep the observations separate before trying to explain their relation. If evidence is incomplete, uncertainty is preferable to an attractive but unsupported story.
Transfer exercise: a two-day learning event with different endings
Change the setting to a fictional two-day workshop. Day one lasts four hours and earns an average momentary usefulness rating of six on a ten-point scale. Day two lasts two hours and earns an average rating of four. A time-weighted average across all six hours is (4×6+2×4)/6=(24+8)/6=32/6≈5.33. An unweighted mean of the two daily averages would be (6+4)/2=5, which is different because the days are unequal lengths.
Now suppose a final twenty-minute recap is rated nine out of ten and is included within the second day’s two-hour average. A respondent might later remember the workshop especially favourably because of that strong ending. We cannot calculate the respondent’s actual global memory rating from these three reported numbers; doing so would require an observed memory measure or an explicitly hypothetical heuristic. This is an important limitation, not missing mathematical sophistication.
Imagine the organiser chooses between improving the long, weaker middle segment and adding another impressive closing speech. A purely peak-end-driven strategy might favour the closing speech. A learning-first strategy should inspect whether participants can use the concepts, where they got stuck and what the longer weak period cost them. A strong closing explanation may be useful, but it cannot replace correction of prerequisites.
Suppose 30 attendees attempt a changed-context test: 18 succeed on the first attempt, eight improve after guided feedback and four still need assistance. The original independent first-attempt success share is 18/30=60%, while 26/30≈86.67% have succeeded after feedback. Those are different conditions and should be reported separately. A survey about how attendees remember the ending is yet another measure.
This changed-context case teaches how to keep units and goals straight: minutes and hours measure exposure, scores measure experience or recall depending on the question, and independently checked solutions measure one aspect of learning. A useful decision weighs all of them rather than selecting the most flattering figure.
What the research does not prove across all people and situations
The peak-end rule is not a claim that duration never matters. Some studies and contexts show strong effects of duration, sequence, expectations, meaning or remembered details. The original cold-water choice has a particular design and participant group; generalisation requires additional evidence. People’s responses can differ, and individual experiences need not obey an aggregate tendency.
A retrospective preference is not identical to wellbeing. A person might prefer to repeat an episode with a better ending while nonetheless experiencing more total unpleasantness during it. Neither measurement should be suppressed. Ethical evaluation may give special weight to serious harms regardless of a favourable overall impression.
Models of memory cannot be evaluated with only a story that sounds familiar. A proper test needs recorded momentary measures, time boundaries, a well-defined retrospective question, suitable comparisons and uncertainty. If the research question concerns causal improvement, allocation of participants and the handling of confounding matter.
The numerical peak-end scores in this guide were constructed as teaching heuristics, not fitted behavioural predictions. No empirical claim follows that any real learner, patient or customer will rank the two fictional episodes in the same way. The distinction between hypothetical demonstration and observed behaviour must remain explicit.
Good application means thoughtful attention to endings, real improvement throughout a process, and independent evidence of outcomes. It never means that a pleasant closing moment can legitimately stand in for every other measure of quality.
Frequently asked questions
What is the peak-end rule? It is an observed tendency in some retrospective evaluations to place disproportionate weight on an experience’s emotionally salient peak and ending rather than giving equal weight to every moment.
Is the peak always the best moment? No. For aversive experiences researchers may study the most unpleasant moment; for positive experiences a particularly rewarding moment may be salient. Define the measure before calculating a score.
Does duration neglect mean time is unimportant? No. It describes patterns in some remembered evaluations, not an absence of physical, practical or ethical consequences from longer exposure.
Can a pleasant final minute erase a harmful experience? No. An improved memory response does not negate harm, unmet rights or the obligation to improve the entire process.
Why were participants sometimes willing to repeat a longer painful trial? In the original experiment the added portion ended less unpleasantly than the earlier period, and many participants chose that trial when asked to repeat one.
Does the formula peak plus end divided by two predict everyone’s memory? No. It is a useful illustrative heuristic, not an individual-level psychological law with universal precision.
How could a teacher use this responsibly? Provide a truthful recap and independent transfer question while also measuring understanding and correcting earlier gaps. Do not substitute a positive mood for learning evidence.
What should a service organisation report? Total durations, interruptions, completion outcomes, abandoned users and retrospective ratings where relevant, each with its own denominator and time window.
How is this different from affective forecasting? Peak-end research concerns evaluation of a remembered episode; affective forecasting concerns predictions about future feelings. The mechanisms can interact but are not identical.
Practical reader routes and the original research
For students, distinguish the time-weighted average from the toy peak-end number in the worked example. Explain why Episode B can have a lower average and a better illustrative ending-based evaluation. Then practise calculating the changed-workshop denominator without mixing hours, minutes and people.
For teachers and tutors, create a meaningful ending that helps students organise ideas and leave with a clear next step. Verify improvement through independent questions and delayed retrieval. Treat student mood as one relevant signal rather than a complete grade or proof of mastery.
For researchers and service managers, collect in-session measures where practical, survey at clearly stated times, include abandoned journeys in denominators and distinguish memory, future choice and objective completion. Check ethical implications before running any intervention based on a cognitive tendency.
Primary source: Kahneman, Fredrickson, Schreiber and Redelmeier (1993), When More Pain Is Preferred to Less: Adding a Better End, Psychological Science 4(6). Related verified eduKate learning routes include Affective Forecasting and the How Intelligence Works hub.
All workshop and service numbers are labelled invented teaching cases, not results from a real school, clinic or business. The purpose of the examples is to sharpen reasoning about what the evidence can establish.