The AI Journal

A horsefly approaching a zebra can fly towards it normally and yet, in the final moments, fail to land. That small behavioural discontinuity matters. In experiments comparing captive zebras with domestic horses, tabanids did not simply avoid the vicinity of zebras. They approached both animals, but their trajectories differed close to the body: flies approaching zebras tended to fly faster, failed more often to decelerate normally, and landed less frequently (Caro et al., 2019). Put a striped coat over a horse and the number of landings on the covered region falls. The effect therefore requires neither a zebra’s smell nor its behaviour nor its African habitat. Something about the visible surface itself changes the fly’s decision or performance at close range.

The tempting explanation is that stripes do something spectacular to vision. Repeated high-contrast lines seem almost designed to generate illusion. They might produce misleading optic flow, apparent movement, spatial aliasing or an aperture effect as the animal and fly move relative to one another. The hypothesis is intuitively satisfying because it treats the extraordinary appearance of a zebra as evidence of an equally extraordinary perceptual mechanism. I argue that this intuition has led the mechanistic discussion in the wrong direction. The most important property of zebra stripes may be subtractive rather than illusory: striping breaks a dark animal into visual elements too small to provide the large homogeneous dark target that attracts a horsefly during its final approach.

The broader evolutionary case for an anti-fly function is strong. Caro et al. (2014) compared geographical and ecological variables across equids and found the strongest association between intense striping and regions where biting-fly activity is prolonged. Experimental work independently established that striped surfaces can be unusually unattractive to tabanids. Egri et al. (2012), using patterned surfaces and horse models, found that attractiveness declined as stripes became narrower; importantly, zebra stripe widths fell within the range associated with low attractiveness. Later experiments produced a wonderfully literal demonstration: painting black-and-white zebra-like stripes on cattle reduced fly landings and fly-repelling behaviour relative to controls (Kojima et al., 2019). Whatever else stripes might do, an effect on biting flies no longer depends on an evolutionary just-so story.

But establishing function is not the same as establishing mechanism. One proposed mechanism has been disruption of optic flow. An approaching fly must estimate its motion relative to a surface and regulate its approach if it is to land without collision. A field of repeated stripes could plausibly corrupt that information. In particular, the aperture effect can make motion viewed through an oriented contour appear to occur in a direction different from its true direction. If a zebra’s stripes transformed the optic-flow field experienced by a tabanid, a fly might reach the animal still travelling too quickly to execute a controlled landing.

How et al. (2020) devised an elegant test. Horses wore grey, striped or checked rugs while the trajectories of approaching horseflies were recorded. Flies approached striped and checked surfaces less closely and landed on them less often than on grey surfaces. The checks are crucial. A checked pattern lacks the continuous parallel contours required for the proposed aperture-effect mechanism, yet it deterred flies too. The experiment therefore retained the phenomenon while removing a supposedly necessary component of its explanation. Stripes plainly interfere with landing, but they need not do so because parallel lines create an aperture illusion.

The pattern-element hypothesis makes sense of this result. A striped zebra and a checked horse rug look quite different if the salient variable is pattern geometry. They look much more alike if the salient variable is the absence of a large uninterrupted dark patch. Caro et al. (2023) tested this possibility using patterned coats on domestic horses. High-contrast stripes discouraged landings whereas low-contrast stripes did not. Regular checks were somewhat more effective than irregular checks, but the difference did not provide strong support for an aliasing explanation. Instead, landing behaviour tracked the size of dark pattern elements: larger dark regions were more attractive. The authors consequently proposed that thin, sharply bounded stripes work because they eliminate the large monochrome dark patches that horseflies find attractive at close range.

That interpretation also changes what requires explanation. Asking why a fly is “confused” by a zebra presupposes that the fly has already categorised the animal as a landing target and then experiences navigational failure. But attraction and landing control need not be separable stages. A visual surface can cease to provide the cues that sustain an approach before a landing manoeuvre is initiated. The zebra would then function less like a moving optical illusion and more like a badly specified destination. The fly’s behaviour near the animal could look like failed navigation even if the critical event were weakening target salience.

This account fits an older result that otherwise risks being treated merely as an interesting precursor. Egri et al. (2012) found not simply that “striped” targets were unattractive but that decreasing stripe width reduced attractiveness. That is exactly what one would predict if segmentation itself matters. Every additional division of a black region reduces the dimensions of the uninterrupted dark elements available to the approaching insect. Under this interpretation, the relevant unit is not “a stripe” at all. Checks, spots or other high-contrast patterns could produce similar effects if they divide attractive regions below the spatial scale at which the fly responds strongly to them. Zebra striping would be one evolutionary solution to a more general problem of visual target fragmentation.

The strongest objection is now unusually strong because it comes from a mechanistic model published after several of these experiments. Mouy (2025) argues that aliasing was dismissed too readily. Modelling image formation and fly motion processing from first principles, he shows that striped patterns can generate spatial and motion-analysis errors at distances consistent with observed fly behaviour, and that randomised checks can also produce substantial directional errors. The fact that checks deter flies therefore does not, by itself, eliminate aliasing. On this account, the experimental evidence that appeared to undermine an optical-illusion explanation may actually be compatible with a more sophisticated version of it.

That objection prevents a categorical claim that target fragmentation is the sole mechanism. It does not, however, restore illusion as the best organising explanation of the behavioural evidence. A model demonstrating that aliasing can occur establishes perceptual plausibility, not that aliasing supplies the causal variable to which landing probability is most sensitive. The 2023 experiments manipulated contrast, regularity and dark-patch dimensions on actual horses and found a direct relationship between the presence of substantial dark areas and fly attraction. Conversely, the failure of simple aperture-effect predictions in the 2020 experiment shows the danger of inferring mechanism from what stripes look capable of doing to a visual system. The decisive experiment would have to vary predicted aliasing while independently holding dark-element size constant, and then reverse that manipulation. Until then, the simpler behavioural regularity deserves priority: horseflies disproportionately land where sufficiently large dark targets remain available.

Nor does thermoregulation provide a compelling reason to assign the pattern another primary function. Horváth et al. (2018) tested water-filled barrels covered with zebra, black, white and grey hides under field conditions and found no evidence that striped coverings kept the contents cooler than homogeneous grey coverings. That result cannot exclude every possible physiological interaction between living zebra pelage, sweating and air movement, but it removes the easy version of the convection argument. The anti-fly evidence, by contrast, spans comparative distributions, artificial targets, living zebras, clothed horses and painted cattle.

The zebra’s appearance encourages us to search for something that stripes positively create: camouflage, dazzle, false motion, altered airflow. The fly experiments point towards the inverse possibility. Natural selection may have produced one of the most conspicuous mammalian coats not because the pattern adds a powerful signal, but because each stripe repeatedly prevents something else from existing. There is almost nowhere on the zebra’s body for a large dark target to persist. For an approaching horsefly, that absence may be the pattern’s most consequential feature.