Compassion fade
Feeling and giving are said to fall as victims are added, starting at the second, and every large preregistered test has found nothing or the reverse.
Feeling and giving are said to fall as victims are added, starting at the second, and every large preregistered test has found nothing or the reverse.
The claim is about a slope, not a contrast. Proportion dominance, entry 193, is about the denominator, and the identifiable victim effect, entry 179, is about a name and a face. Compassion fade says the curve turns down as soon as a second person is added and keeps falling. Slovic drew it as a value function that peaks at one life and declines from there, and called the decline psychic numbing after Robert J. Lifton's term for what let rescue workers function after Hiroshima.
Two mechanisms compete. The affect account says the feeling system cannot scale, so eight children move a donor less than one does. Cameron and Payne's motivated account says groups do stir feeling and people shut it down in advance when they expect the bill. The distinction matters because only the second predicts that the fade should vanish when nobody is going to ask for money.
Failed to replicate The origin's own experiments were rerun on 1,207 people and the fade did not appear, and the largest test of the mechanism found more compassion for eight children than for one.
Mayiwar and seven colleagues rebuilt Västfjäll's Experiments 1a and 3 in one preregistered design, on 1,207 US participants recruited through CloudResearch Connect. The one-versus-two contrast on hypothetical donations came out at d = -0.16, t(337) = -1.76, p = .079, against the original's d = 0.82, and the sign favours the two children. Positive affect, negative affect and perceived impact moved no further than d = 0.16, with Bayes factors of 5.75 and 5.92 favouring the null on affect and on perceived impact.
In the five-condition replication of Study 3 the one supported contrast was more money for eight related children than for one, d = -0.29, again the wrong way. That work is a preprint posted on PsyArXiv in August 2026 and not a published article, and its two posted versions carry the same numbers. Hagman, Tinghög, Dickert, Slovic and Västfjäll rebuilt Cameron and Payne's Experiment 1 with 1,177 US participants in a two by two by two design, powered to detect a small effect.
The interaction that the original rested on came out at F(1, 1,173) = 0.012, p = 0.913, partial eta squared 0.000, against the original's partial eta squared .04. The main effect of victim number ran the wrong way: compassion was higher for the eight children than for the one, F(1, 1,173) = 23.549, p = 0.000, partial eta squared 0.020, in both the help-request and no-help-request conditions.
Mood fell further after eight children, F(1, 1,173) = 15.473, p = 0.000, partial eta squared 0.013. Donation amount showed no interaction, F(1, 1,173) = 0.099, p = 0.754. The manipulation itself did not hold: 53.4% of the help-request group and 59.2% of the no-help-request group said they had thought about donating while rating their feelings. Splitting the whole sample by whether people expected to be asked, which should have been the friendliest possible test, still produced no interaction.
Hart, Lane and Chinn had already run two large tests of the individual-versus-group contrast. Their Study 1, on 924 MTurk participants, found no difference across four appeal types on feelings, F(3, 916) = 1.45, p = .23, or donation amount, F(3, 916) = .95, p = .41. Study 2, on 1,085 Qualtrics panellists, likewise found nothing on feelings, F(1, 1079) = .053, p = .82, or donation amount, F(1, 1079) = .002, p = .96.
What stands on the other side is the Butts meta-analysis, whose abstract reports the effect in the predicted direction across 95 samples and whose full text could not be read. Markowitz's pandas are often counted as support, and the F(1, 165) = 4.93 quoted for them is the interaction between environmentalism and the linear trend, not a fade across the sample.
One child on the card, or eightillustration
An appeal card carries a named, photographed child, and a second version of the card carries eight children named and photographed the same way. The claim is that the second raises less. Identifiability has been held still and only the count has moved, which is where the fade is supposed to live.
The second name on the collection sheetillustration
A street collection for one neighbour whose flat flooded is said to bring in more per household than the same sheet raising money for three. The claimed decline begins at the second person rather than at some crowd too large to picture.
Six dogs at the shelterillustration
A shelter posts a photograph and a history of one dog waiting for a home, then a post built the same way for six dogs. If feeling scales with the number, the six should move readers six times as far, and the claim on this page is that it moves them less than the one.
Cases marked as illustrations describe nobody in particular. They are there to make the pattern recognisable, not to report that a particular thing happened.
Västfjäll and colleagues ran five experiments, four of them on Swedish undergraduates and the last on 559 students at the University of Oregon. Study 1a, on 208 people, gave hypothetical donations of 37.7 SEK in the single-child conditions against 26.3 SEK for two children, t(205) = 5.88, p < .001, with affect ratings of 3.7 against 3.3, t(205) = 2.87, p < .01.
Study 1b put real money on it. Its 168 participants could give their earnings to Save the Children for Rokia (n = 47), for Moussa (n = 51), or for both (n = 70), and gave 24.5 SEK to one child against 21.5 SEK to two, t(166) = 1.67, p < .05, with affect at 3.5 against 3.0, t(166) = 2.81, p < .01.
Study 2, on 107 people, measured facial muscle activity while donations fell across one, two and eight children, 24.9, 23.4 and 21.3 SEK, linear trend F(1,103) = 12.01, p < .001. Zygomaticus activity, which tracks positive affect, ran .41, .12 and .09 across the same conditions, F(1,103) = 8.90, p < .001. Cameron and Payne supplied the mechanism study.
Their Experiment 1 showed one or eight children from Darfur to 120 students at Chapel Hill, of whom 118 were analysed, and told half that a donation question was coming. Compassion showed no main effect of help request, F(1, 116) = 1.15, p = .29, none of victim number, F(1, 116) = 0.34, p = .56, and an interaction at F(1, 116) = 4.61, p = .03, partial eta squared .04.
Markowitz, Slovic, Västfjäll and Hodges carried the pattern to animals. Their Study 3 paid 171 students $5 in single dollar bills and asked for a real donation to panda conservation. Non-environmentalists gave $3.50 for one panda, $2.83 for two and $2.25 for eight, linear contrast F(1, 82) = 5.29, p = .02, r = -.25, while environmentalists gave $2.50, $3.30 and $2.93 with no linear effect, F(1, 85) = .68, p = .41.
Across the whole sample neither the number of pandas nor environmentalism produced a main effect, both Fs below 1. Butts, Lunt, Freling and Gabriel pooled 41 studies, 95 independent samples and 13,259 participants, and their abstract reports that victim group size negatively affects helping intent, helping behaviour, anticipated positive affect and perceived impact, but not empathic concern.
5 primary research4 replication and reanalysis2 commentary and critique
Their paper carries the name and the one-versus-two data. The finding itself was in print seven years earlier as a figure: Slovic's 2007 essay shows mean affect and mean donations for Rokia, Moussa and the pair, credited to Västfjäll, Peters and Slovic, in preparation, with no numbers attached and the comment that perhaps the blurring of individuals begins at two. Markowitz and colleagues cite the same work in 2013 as a Decision Research white paper titled "Compassion fatigue: Donations and affect are greatest for a single child in need".
The claim, and the replication this entry reads its figures from.
The drop from one victim to two, which is the claim that makes compassion fade more than a restatement of scope insensitivity, entry 186, rests on a thin base. In the only one-versus-two study with real money, Västfjäll's Study 1b, the gap was 3 SEK on a t of 1.67, with 47 and 51 people in the two single-child conditions and 70 in the two-child condition.
The larger and cleaner-looking gap, 37.7 against 26.3 SEK, came from the hypothetical study. Four of the five experiments in the original paper were run on undergraduates at a single Swedish university, and Markowitz's three on students at public universities in the United States. Sample sizes at the origin are small where the effect is claimed to be a slope: Markowitz's Study 3 split 171 participants across three donation conditions and two self-identified groups.
The Butts meta-analysis is the only pooled estimate this entry located, and its correlations, confidence intervals and publication-bias tests could not be read. It is closed at Elsevier, no repository copy is listed by Crossref, Unpaywall, OpenAlex or CORE, and the publisher's page answered a plain request with a Cloudflare block on 21 September 2026, so the direction reported above comes from the publisher's abstract alone.
Meier's field study of real-world online donations could not be read either. Crossref, Unpaywall and OpenAlex all mark it open under a CC-BY licence, yet on 21 September 2026 every Sage route tried, the landing page, the PDF, the ePub and the text-mining XML, answered four different user agents with a Cloudflare bot challenge, and the one archived copy sits on a host this fetcher is barred from reaching. Nothing from it appears on this page.
Each of these is named in this entry's own prose as something it gets taken for. Where the two came out differently, that is worth knowing before you quote either.
Cameron and Payne are cited as having demonstrated compassion collapse, and their Experiment 1 did not. When help was requested, compassion toward one victim was only numerically higher than toward eight, F(1, 59) = 1.18, p = .28. The significant simple effect was the other one: with no help requested, eight victims drew more compassion than one, F(1, 57) = 3.87, p = .05, partial eta squared .06.
Their result is a reversal in the no-request condition, not a fade in the request condition. Second, compassion fade is not compassion fatigue. The clinical term describes the erosion of a carer's capacity over time; this is a cross-sectional response to how many people are in one appeal, and the original Decision Research white paper used the wrong one of the two names.
Third, the effect is not the identifiable victim effect, entry 179, renamed. That effect contrasts a named person with statistics, this one holds identifiability fixed and counts heads, and the two can be dissociated: Hart, Lane and Chinn crossed appeal type with imagery on 1,085 people and moved neither. Nor is it the singularity effect. Kogut and Ritov put that name on the gap between one identified victim and a group of eight, measured with real contributions from 153 students in Jerusalem, and that literature is covered under entry 179.
Majumder and colleagues keep singularity apart from compassion fade, treating compassion fade as the version that makes victim number a continuous measure, though they file it alongside scope insensitivity rather than opposite it. Fourth, the curve that appears in talks and textbooks, peaking at one life and declining, is a drawing. Slovic's Figure 11 is a model whose axes are labelled value of life saving and number of lives with no scale on either, published alongside a study then in preparation, and nobody has measured the function it depicts.
Fifth, a reader who takes the effect as settled should note that its own proponents do not. Hagman and colleagues, five authors who work on this literature and two of whom wrote the 2014 paper, published the failure themselves.
Every claim on this page was held against these sources on 2026-09-21. Nothing here is written from memory. Written and checked by Krishna Chagti; how entries are written and corrected. Found a mistake? Report it, with the sentence and a source.