Cognitive Bias in Dogs: Measuring Optimism, Pessimism and Mood

1. The Measurement Problem

1.1 Why Behavior Is Not Enough

A dog that has stopped struggling may be calm or may have given up. A dog with a low tail may be relaxed or worried. The same visible behavior can arise from different emotional states, and the same emotional state can produce different behaviors depending on context. Physiological measures do not resolve this either: cortisol moves with arousal rather than valence, so it rises under excitement as readily as under distress (which is why operational definitions matter).

The gap this leaves is specific: there is still no validated indicator of positive emotional state in dogs (Csoltova & Mehinagic, 2020), and the standard indicator set differentiates arousal only under negative valence (Flint et al., 2024). Absence of visible distress is not evidence of wellbeing.

1.2 The Idea Borrowed From Human Psychology

In humans, mood biases judgement. Anxious and depressed people interpret ambiguous information more negatively — the same neutral face, the same ambiguous sentence, read as threatening rather than benign. This is well established and, crucially, it is measurable through choices rather than through self-report.

If the mechanism is general rather than specifically human, then an animal in a negative mood should also treat ambiguity more cautiously. That is a testable claim, and testing it does not require the animal to say anything.

1.3 The Founding Experiment

Harding, Paul and Mendl (2004) trained rats to press a lever for food when they heard one tone and to avoid pressing when they heard another that signalled an unpleasant noise. Once the discrimination was solid, they played tones of intermediate pitch that the animals had never encountered.

Rats housed in unpredictable conditions responded to the ambiguous tones as though they predicted the bad outcome. The finding launched an entire field, and the underlying framework was developed further into a general account of animal emotion and mood (Mendl, Burman, Parker & Paul, 2009; Mendl, Burman & Paul, 2010b).

2. How Cognitive Bias Tests Measure Dog Emotion

2.1 The Spatial Version Used With Dogs

Nearly all canine work uses a spatial task. A bowl in one location always contains food; a bowl in another location is always empty. Once the dog reliably runs to the first and hesitates at the second, bowls appear in ambiguous locations between the two — typically three: near-positive, middle, and near-negative.

The measurement is latency: how quickly the dog approaches the ambiguous bowl. Faster approach is read as an "optimistic" response — the dog behaving as though food is likely. Slower approach is read as "pessimistic."

2.2 What the Terms Actually Mean

"Optimistic" and "pessimistic" here are technical labels for response patterns, not claims about canine belief or personality. A "pessimistic" response does not mean the dog expects the worst; it means the ambiguity was resolved in a direction statistically associated with a negative affective state. The dog is not making a prediction about its life — it is responding to a cue whose meaning it cannot settle, and the direction of that response is used as a proxy.

The distinction matters because the everyday sense of the words is what most secondary coverage runs with, and it overstates what the data supports.

2.3 Why Ambiguity Is the Whole Point

Trained cues tell you only that the dog learned the discrimination. The untrained middle cue is where prior expectation has to fill the gap, and prior expectation is what mood is hypothesized to shift. The test is therefore a controlled way of asking what the dog expects when the evidence runs out (which connects it to prediction error more broadly).

3. What the Canine Studies Show

3.1 Separation-Related Behavior

One of the earliest and most influential canine studies tested dogs in a rehoming kennel and compared those that showed separation-related behavior when left alone with those that did not. Dogs displaying separation-related behavior responded to ambiguous bowl locations more slowly — a pessimistic pattern consistent with a negative underlying affective state (Mendl et al., 2010a).

The interpretive value is real. It provides evidence that separation-related behavior is not a training failure or a bid for attention but sits on top of a genuinely negative state (as the neurobiology of that presentation indicates).

3.2 Training Methods

Dogs whose owners reported using two or more aversive training methods approached ambiguous locations more slowly than dogs trained with reward-based methods (Casey et al., 2021). The design was cross-sectional and the training data owner-reported, so causation is not established — but the result converges with independent welfare findings (on the neurological effects of aversive methods).

3.3 Chronic Conditions and Pain

Cavalier King Charles Spaniels with syringomyelia showed response patterns consistent with a negative affective state (Cockburn et al., 2018), and dogs with epilepsy differed from controls on judgement and attention bias measures (Hobbs et al., 2020). Both fit a wider pattern in which physical conditions shift emotional state in ways that are easy to miss clinically (as with visceral pain).

3.4 Housing and Age

Shelter dogs behaved more pessimistically than pet dogs in one carefully analysed comparison of 51 shelter and 40 pet dogs — though, importantly, only under one method of scoring (Burani et al., 2020). Age also affects performance on discrimination, reversal and cognitive bias tasks in family dogs (Piotti et al., 2018), which means age has to be controlled rather than assumed away (and cognitive decline complicates it further).

4. What Changes Cognitive Bias in Dogs

4.1 Treatment Changes It

The most clinically important canine result concerns a question that behavior outcome measures cannot answer. When a dog stops howling and stops destroying the door frame, two very different things may have happened: the dog may feel better, or the dog may simply have stopped doing the thing the owner complained about. Owner-reported improvement cannot distinguish them, and neither can a checklist of observed behaviors.

Dogs with separation-related problems treated with fluoxetine alongside a behavior modification plan showed a shift toward less pessimistic responding over the course of treatment (Karagiannis, Burman & Mills, 2015). The authors framed it as the first demonstration that clinical treatment of a negative affective state in a non-human species produces a measurable shift in cognitive bias.

That is a substantial claim, and its significance is methodological rather than pharmacological. It establishes that the question did the animal actually get better? is answerable in principle, using a measure that does not depend on the owner's judgement or on the behavior that prompted the referral (which is exactly the distinction suppression obscures). One study without an untreated control group cannot settle how well the method performs — but it sets the standard any outcome claim should be measured against.

4.2 Another Person's Stress Changes It

Dogs exposed to body odour collected from an unfamiliar stressed person responded more pessimistically to ambiguous bowl locations than dogs exposed to odour from a relaxed person (Parr-Cortes et al., 2024). The person was a stranger the dogs never saw.

Two implications follow. Chemical signals of human stress are detectable and affect canine decision-making (through olfactory channels), and the transfer of emotional state between humans and dogs is not limited to visible or familiar cues (the emotional contagion picture).

4.3 What This Adds Up To

The manipulations that move canine judgement bias are the ones a welfare framework would predict: chronic stress, aversive handling, unresolved separation distress, physical illness, and proximity to a stressed human. Nothing in the canine literature contradicts the framework. The question is how reliably any single test detects it.

5. The Null Results

5.1 Brief Owner Absence Did Nothing

Pet dogs briefly separated from their owners showed no negative judgement bias afterwards (Müller et al., 2012). The manipulation was mild and short, and the result is best read as a boundary condition: the test does not respond to every stressor, and a brief separation in a familiar setting is not equivalent to the chronic separation distress in the kennel study.

5.2 Why Null Results Deserve Space Here

A field where positive findings are widely reported and null findings are not produces a distorted picture, and this literature has that shape. Null results define where a method works. Reporting only the successes leaves practitioners believing the test is more sensitive than it is (the same problem that affects reconsolidation research).

6. Limits of Cognitive Bias Tests in Dogs

6.1 The Cross-Species Effect Is Small

The largest systematic review and meta-analysis of judgement bias across species found that non-pharmacological manipulations of affect do alter judgement bias in the expected direction — animals in better conditions responded more optimistically — but that the overall effect size was small, and that effect sizes were highly heterogeneous between studies (Lagisz et al., 2020).

That is a genuine validation and a genuine limitation in the same sentence. The paradigm detects something. It detects it weakly, and inconsistently across designs. A parallel meta-analysis of pharmacological manipulations reached a similarly qualified conclusion (Neville et al., 2020).

6.2 The Scoring Method Changes the Answer

The canine-specific methodological analysis is the one practitioners should know about. Examining 51 shelter and 40 pet dogs, the shelter–pet difference appeared when raw approach latencies were analysed and disappeared under the adjusted score that most published studies use (Burani et al., 2020).

A result that depends on which of two standard analyses is applied is not a robust result. It does not mean the difference is imaginary; it means the measurement is not yet stable enough to carry the weight often placed on it.

6.3 Dogs Drop Out of the Task

The same analysis flagged a further problem: a non-negligible proportion of dogs never completed the training phase and were excluded (Burani et al., 2020). If the dogs that fail to learn a food-location discrimination differ systematically from those that succeed — in motivation, in food value, in anxiety, in age — then the tested sample is not the sample of interest (and individual differences here are substantial).

6.4 Latency Is Not Only About Mood

Approach speed is affected by hunger, general activity level, how much the dog values the food, prior reinforcement history, and arousal (which has its own effects on performance). Studies control what they can. The measure remains indirect, and the inferential chain from a walking speed to a mood state is longer than the phrase "optimistic dog" suggests.

7. What It Means for Practice

7.1 It Is Not a Test You Can Run

The task requires many training trials, controlled conditions, careful timing, and ambiguous probes presented rarely enough that the dog does not learn what they mean. Improvised versions measure nothing. This is a research instrument, and treating it as a practitioner tool misrepresents both.

7.2 What Is Genuinely Usable

Three things transfer. First, the concept: a dog's interpretation of ambiguous situations is shifted by its background emotional state, which reframes the "he's being difficult in new places" complaint as information about mood rather than obedience (the layer beneath trained behavior).

Second, the findings themselves, which give evidence-based weight to positions that otherwise rest on preference: aversive methods and unresolved separation distress are associated with a measurably worse affective state.

Third, the treatment standard. The fluoxetine result establishes that it is possible in principle to ask whether an intervention improved the dog or only quieted it — the question worth asking after any behavior program (including desensitization and counterconditioning).

7.3 What Not to Claim

Not that a specific dog is an optimist or a pessimist. Not that a single observation reveals mood. And not that the absence of a pessimistic response demonstrates a dog is happy — given the small effect sizes and the absence of any validated positive-emotion indicator, that inference runs well past the evidence (as the anxiety literature also shows).

The wording that survives scrutiny is narrower than the wording that circulates. Not "dogs trained with aversives are sad," but "dogs whose owners reported using two or more aversive methods evaluated ambiguous situations more cautiously in a controlled task." The second sentence is less quotable and is the one the data supports.

8. Summary at a Glance

The problem it solves — There is no validated indicator of positive emotion in dogs (Csoltova & Mehinagic, 2020), and standard indicators track arousal rather than valence under positive conditions (Flint et al., 2024).

The logic — Mood biases how ambiguity is interpreted. Trained in rats first (Harding et al., 2004), formalized as a general framework for animal affect (Mendl et al., 2009; 2010b).

The canine task — A bowl in a rewarded location, a bowl in an unrewarded location, and ambiguous positions between. Approach latency is the measure.

What it found — Pessimistic responding in dogs with separation-related behavior (Mendl et al., 2010a), in dogs trained with two or more aversive methods (Casey et al., 2021), in dogs with syringomyelia (Cockburn et al., 2018) and epilepsy (Hobbs et al., 2020).

What shifts it — Treatment with fluoxetine plus behavior modification moved dogs toward less pessimistic responding (Karagiannis et al., 2015), and the odour of an unfamiliar stressed person moved them the other way (Parr-Cortes et al., 2024).

Where it fails — Brief owner absence produced no effect (Müller et al., 2012); the shelter–pet difference appeared under one scoring method and vanished under another (Burani et al., 2020).

The overall effect is small — Across species, manipulations of affect shift judgement bias in the predicted direction, with small and highly heterogeneous effect sizes (Lagisz et al., 2020).

9. Research Gaps and Critical Appraisal

The effect is real but small. The cross-species meta-analysis supports the paradigm's basic validity while showing that effect sizes are small and heterogeneity is high (Lagisz et al., 2020). Any account that presents the test as a reliable mood-reading device overstates what the aggregate data supports.

Analytical choices change conclusions. A canine population difference that appears under raw latencies and disappears under the standard adjusted score (Burani et al., 2020) indicates the measurement has not been methodologically settled.

Sample attrition is unquantified in most reports. Dogs that fail the training phase are excluded, and the ways in which they differ from included dogs are rarely characterized. This is a selection problem sitting underneath a welfare measure.

The valence interpretation is inferred, not demonstrated. Slow approach to an ambiguous bowl is consistent with negative mood. It is also consistent with lower food motivation, higher general caution, or a different reinforcement history. Studies control for some of this; none can eliminate it.

Positive states remain the hard case. Most canine findings concern negative states. Demonstrating that a dog has become positively content — as opposed to less distressed — remains the unsolved problem the paradigm was partly invented to address.

Canine sample sizes are small and mostly cross-sectional. Individual studies typically involve tens of dogs, with limited longitudinal work. The fluoxetine result is the notable exception in design, and it is one study without an untreated control group.

Test–retest reliability in dogs is underreported. Whether a given dog produces a similar bias score on separate occasions is a basic psychometric question that the canine literature has not adequately answered, and without it the measure cannot support individual-level claims.

10. Conclusion

The cognitive bias test is the most serious attempt yet to answer a question that behavior alone cannot settle — whether a dog is in a good or a bad state — and its canine results have been genuinely useful. It supplied evidence that separation-related behavior sits on top of a negative emotional state rather than being a nuisance to be managed, that dogs trained with multiple aversive methods differ measurably in affect and not only in observable stress signals, and that treatment can shift the underlying state rather than merely quieting the symptom. It also demonstrated that a stranger's stress odour is enough to move a dog's judgement, which is worth sitting with. At the same time, the aggregate effect across species is small, effect sizes vary widely between designs, a canine population difference can appear and disappear depending on which of two standard scoring methods is applied, and an unquantified proportion of dogs never gets far enough into the task to be tested at all. Both halves are true, and the useful stance holds them together: this is the best instrument the field has for a question it cannot otherwise ask, and it is not yet precise enough to tell you about the dog in front of you. The concept is what practitioners should carry — that background emotional state shapes how a dog reads an ambiguous situation, which reframes a great deal of behavior that otherwise gets attributed to stubbornness (as chronic stress research also implies).

Key Insights (Takeaways)

• The test exists because behavior alone cannot distinguish a calm dog from a shut-down one. There is no validated indicator of positive emotion in dogs (Csoltova & Mehinagic, 2020), and standard indicators track arousal rather than valence under positive conditions (Flint et al., 2024). Judgement bias measures how a dog interprets an ambiguous cue and uses that as a proxy for mood.

• The canine findings converge on welfare. Pessimistic responding appeared in dogs with separation-related behavior (Mendl et al., 2010a), in dogs whose owners used two or more aversive training methods (Casey et al., 2021), and in dogs with painful or neurological conditions (Cockburn et al., 2018; Hobbs et al., 2020). Nothing in the canine literature contradicts the welfare framework.

• Treatment moved the underlying state, not just the behavior. Dogs with separation-related problems treated with fluoxetine plus behavior modification shifted toward less pessimistic responding (Karagiannis et al., 2015) — the first demonstration in a non-human species that clinical treatment changes cognitive bias, and the model for asking whether a dog is better or merely quieter.

• A stranger's stress is enough to shift it. Dogs exposed to body odour from an unfamiliar stressed person responded more pessimistically than those exposed to odour from a relaxed person (Parr-Cortes et al., 2024). The transfer of human emotional state to dogs is not limited to visible cues or familiar people.

• The method is weaker than the headlines. Across species the effect is small and highly heterogeneous (Lagisz et al., 2020); in dogs, a shelter–pet difference appeared under raw latencies and disappeared under the standard adjusted score, with a meaningful proportion of dogs excluded for failing training (Burani et al., 2020). It is a research instrument, not a test to run on an individual dog.

References

Burani, C., Barnard, S., Wells, D., Pelosi, A., & Valsecchi, P. (2020). Using judgment bias test in pet and shelter dogs (Canis familiaris): Methodological and statistical caveats. PLoS ONE, 15(10), e0241344. https://doi.org/10.1371/journal.pone.0241344

Casey, R. A., Naj-Oleari, M., Campbell, S., Mendl, M., & Blackwell, E. J. (2021). Dogs are more pessimistic if their owners use two or more aversive training methods. Scientific Reports, 11(1), 19023. https://doi.org/10.1038/s41598-021-97743-0

Cockburn, A., Smith, M., Rusbridge, C., Fowler, C., Paul, E. S., & Murrell, J. C. (2018). Evidence of negative affective state in Cavalier King Charles Spaniels with syringomyelia. Applied Animal Behaviour Science, 201, 77–84. https://doi.org/10.1016/j.applanim.2017.12.019

Csoltova, E., & Mehinagic, E. (2020). Where do we stand in the domestic dog (Canis familiaris) positive-emotion assessment: A state-of-the-art review and future directions. Frontiers in Psychology, 11, 2131. https://doi.org/10.3389/fpsyg.2020.02131

Flint, H. E., Weller, J. E., Parry-Howells, N., Ellerby, Z. W., McKay, S. L., & King, T. (2024). Evaluation of indicators of acute emotional states in dogs. Scientific Reports, 14(1), 6406. https://doi.org/10.1038/s41598-024-56859-9

Harding, E. J., Paul, E. S., & Mendl, M. (2004). Cognitive bias and affective state. Nature, 427(6972), 312. https://doi.org/10.1038/427312a

Hobbs, S. L., Law, T. H., Volk, H. A., Younis, C., Casey, R. A., & Packer, R. M. A. (2020). Impact of canine epilepsy on judgement and attention biases. Scientific Reports, 10(1), 17719. https://doi.org/10.1038/s41598-020-74777-4

Karagiannis, C. I., Burman, O. H. P., & Mills, D. S. (2015). Dogs with separation-related problems show a "less pessimistic" cognitive bias during treatment with fluoxetine (Reconcile™) and a behaviour modification plan. BMC Veterinary Research, 11, 80. https://doi.org/10.1186/s12917-015-0373-1

Lagisz, M., Zidar, J., Nakagawa, S., Neville, V., Sorato, E., Paul, E. S., Bateson, M., Mendl, M., & Løvlie, H. (2020). Optimism, pessimism and judgement bias in animals: A systematic review and meta-analysis. Neuroscience & Biobehavioral Reviews, 118, 3–17. https://doi.org/10.1016/j.neubiorev.2020.07.012

Mendl, M., Brooks, J., Basse, C., Burman, O., Paul, E., Blackwell, E., & Casey, R. (2010a). Dogs showing separation-related behaviour exhibit a 'pessimistic' cognitive bias. Current Biology, 20(19), R839–R840. https://doi.org/10.1016/j.cub.2010.08.030

Mendl, M., Burman, O. H. P., Parker, R. M. A., & Paul, E. S. (2009). Cognitive bias as an indicator of animal emotion and welfare: Emerging evidence and underlying mechanisms. Applied Animal Behaviour Science, 118(3–4), 161–181. https://doi.org/10.1016/j.applanim.2009.02.023

Mendl, M., Burman, O. H. P., & Paul, E. S. (2010b). An integrative and functional framework for the study of animal emotion and mood. Proceedings of the Royal Society B: Biological Sciences, 277(1696), 2895–2904. https://doi.org/10.1098/rspb.2010.0303

Müller, C. A., Riemer, S., Rosam, C. M., Schößwender, J., Range, F., & Huber, L. (2012). Brief owner absence does not induce negative judgement bias in pet dogs. Animal Cognition, 15(5), 1031–1035. https://doi.org/10.1007/s10071-012-0526-6

Neville, V., Nakagawa, S., Zidar, J., Paul, E. S., Lagisz, M., Bateson, M., Løvlie, H., & Mendl, M. (2020). Pharmacological manipulations of judgement bias: A systematic review and meta-analysis. Neuroscience & Biobehavioral Reviews, 108, 269–286. https://doi.org/10.1016/j.neubiorev.2019.11.008

Parr-Cortes, Z., Müller, C. T., Talas, L., Mendl, M., Guest, C., & Rooney, N. J. (2024). The odour of an unfamiliar stressed or relaxed person affects dogs' responses to a cognitive bias test. Scientific Reports, 14(1), 15843. https://doi.org/10.1038/s41598-024-66147-1

Piotti, P., Szabó, D., Bognár, Z., Egerer, A., Hulsbosch, P., Carson, R. S., & Kubinyi, E. (2018). Effect of age on discrimination learning, reversal learning, and cognitive bias in family dogs. Learning & Behavior, 46(4), 537–553. https://doi.org/10.3758/s13420-018-0357-7