Published in Animal Behaviour in August 2015, the experiment was a small piece of domestic theatre staged in front of 54 dogs. The owner sat in the middle. A stranger sat on either side. For eight to ten seconds the owner struggled with the lid of a transparent container holding a roll of vinyl tape, then turned to one of the strangers and held the container out, asking for help.
In one version the stranger took the container in both hands, the lid came off, and the owner cheerfully showed the tape to the dog. In another, the stranger turned away and the owner kept failing. Afterwards both strangers held out an identical treat on an open palm at the same moment, neither of them looking at the dog.
The dog chose one.
Dogs who had watched the refusal took food from that person less often than from the stranger on the other side who had done nothing at all. Across the four trials each dog completed, choices of the refuser fell significantly below the chance level of two, at P = 0.023. Hitomi Chijiiwa, Hika Kuroshima, Yusuke Hori, James R. Anderson and Kazuo Fujita published it as “Dogs avoid people who behave negatively to their owner: third-party affective evaluation”.
This is one study, not settled consensus. None of us are comparative psychologists, and what follows is a reading of the paper, of the university’s own summary published on 24 June 2015, and of the research it sits beside.
What the dogs were actually watching
One obvious objection is that a person turning their body away is simply unpleasant to watch, and the design anticipates it. Dogs were assigned to one of three groups of eighteen. In the helper condition the owner asked and received help. In the nonhelper condition the owner asked and was refused. In the control condition the owner worked at the lid for the same interval, paused briefly, and the actor turned away without having been asked anything. In that condition too, the owner failed to open the container.
Dogs in the control condition showed no preference. Turning away, on its own, cost the actor nothing. What cost the actor something was turning away after being asked.
Smaller details make the result harder to explain away. Inside the container was a roll of tape, of no use to a dog, and the authors record that no dog tried to get it out at any point. The neutral person sat looking at the floor throughout, nobody made eye contact with the dog at the moment of choosing, positions were counterbalanced, and every dog got a treat whichever hand it went to.
The half of the result that goes missing
In the helper condition, the dogs’ choices were indistinguishable from chance. Watching a stranger step in and solve the owner’s problem did not make that stranger more attractive than a person who had spent the whole sequence studying the carpet. Reported effect sizes were r = 0.54 in the nonhelper condition, against 0.32 in the helper condition and 0.16 in the control, neither of the latter two reaching significance.
Chijiiwa and colleagues flagged the lopsidedness themselves, noting that the same one-sided pattern has turned up in three-year-old children and in the tufted capuchin monkeys their lab had tested earlier. Their reading was that negative evaluation comes first, with positive evaluation of helpful agents arriving later in human development. Whether dogs show the positive version under other conditions they left open, and put on their own list of things to test next.
Why the shape of it is familiar
Anyone who has spent time with the human negativity-bias literature will recognise the contour. Most of that work traces back to Roy Baumeister, Ellen Bratslavsky, Catrin Finkenauer and Kathleen Vohs, writing in Review of General Psychology in 2001. They worked through evidence across everyday events, close relationships, feedback, learning and impression formation, and found the same skew almost everywhere they looked. Bad impressions form faster than good ones and hold up better against later evidence that contradicts them.
That paper deserves the same scrutiny we are applying to everything else here. It is a narrative review rather than a meta-analysis, assembled a quarter of a century ago and before psychology’s replication reckoning, and its authority comes from breadth of convergence rather than from any single robust result.
Most people can audit the ordinary human version of this themselves. A colleague who did not back you in a meeting stays filed, retrievable years later, in some detail. The one who did back you is remembered as having been there. Help that arrives because it was requested tends to get absorbed into the expected, and things absorbed into the expected stop being events.
What the study does not establish
It does not show that dogs have a moral sense, and the authors do not claim it. Their framing in the university summary puts the weight on affective functions, not on language or on close evolutionary proximity to humans, and allows that the process may be automatic rather than conscious.
One alternative route stays open. That control condition rules out the turning-away gesture and holds the owner’s failure constant, which is careful work. What it cannot fully hold constant is the owner. A person who has been visibly refused may carry themselves differently in the seconds afterwards than a person who has simply paused, and a dog reading its owner’s bearing would produce the same result as a dog evaluating the actor. Separate work from the same lab that year, by Akiko Takaoka and colleagues, showed how finely dogs track human reliability when their own interests are at stake.
Eighteen dogs per condition and four trials each is a small foundation, and the adjacent human literature has become less tidy than it looked. A 2007 study by Kiley Hamlin, Karen Wynn and Paul Bloom reported that infants prefer a character who helps over one who hinders, and it became the anchor for the claim that social evaluation appears in the first year of life. A preregistered multi-laboratory replication run through the ManyBabies framework and published in Developmental Science targeted the video-stimulus version instead of the original live puppet show. Across 37 laboratories and 567 retained infants aged five and a half to ten and a half months, the preference did not differ from chance.
That has not settled the matter against Hamlin. Replication attempts have gone both ways over nearly two decades, and swapping live puppets for video is itself a candidate explanation for a null result. Contested is the accurate word.
Kyoto’s own follow-through is instructive. Running the identical procedure with cats, Chijiiwa and colleagues reported in Animal Behavior and Cognition in 2021 that cats showed no avoidance of the person who had refused their owner. They published the null result rather than leaving it in a drawer, which is more than the field always manages.
The bookkeeping
What the dog result suggests, if it holds, is how little machinery this kind of accounting requires. No language, no theory of anyone’s motives, nothing at stake, and the record still comes out one-sided. A mechanism that cheap is unlikely to be one that insight corrects, and whether knowing about the asymmetry does anything to interrupt it is not a question this literature has answered.
In their write-up, the Kyoto group listed what they wanted to test next. One was whether a helper would earn a positive evaluation if the help came unprompted. That idea is smaller and more testable than anything about moral sense: being asked may change what an act is worth to whoever is watching.
