Honeybees trained to always pick the card with fewer black shapes on it were later shown a blank card next to a card carrying a single shape. They went for the blank one 63 per cent of the time. The result was published in Science on 8 June 2018, and the authors read it as bees ranking an empty set below one.
The work came from Scarlett Howard, Adrian Dyer and colleagues at RMIT University in Melbourne, with co-authors at the Université de Toulouse and the CNRS in France. Dyer was the corresponding author. The argument that followed has never been about whether the bees picked the blank card. It is about what they were actually looking at when they did.
Before this, the short list of animals credited with some understanding of zero ran to rhesus and vervet monkeys, one chimpanzee and one African grey parrot, as the paper's own introduction sets out. A honeybee brain has fewer than 1 million neurons, against the 86,000 million of a human brain.
If a brain this small can rank nothing below one, zero may be a simpler idea than we assumed. That only holds if the bees were really counting, and that is still argued.
Bees were trained one at a time with appetitive-aversive differential conditioning, so a correct choice was rewarded and a wrong one punished, using white square cards carrying one to four black elements. Ten bees were trained to pick the card with fewer elements and ten to pick the card with more. Training ran until a bee was at 80 per cent accuracy or better.
In unrewarded tests with new one to four combinations, the less than group was right 75.0 per cent of the time and the greater than group 75.5 per cent. Then came the card with nothing on it. Bees trained to less than chose the empty card 64.0 per cent of the time, which the paper reports as significant at P = 0.005, and bees trained to greater than went the other way and chose the card with elements 74.5 per cent of the time.
A second experiment trained 25 bees on cards holding two to five elements, so that both one and zero were unfamiliar. Shown one against zero, they chose zero 63.0 per cent of the time.
The same bees were then given a harder pairing. Two elements against zero sets the rule against the training history, because two had always been rewarded and zero was now the lower number. Here the bees split almost evenly, choosing the empty card 56.2 per cent of the time, which the paper reports as not different from chance. The authors read that as two mechanisms pulling against each other, an associative pull towards the familiar rewarded card and a rule-based pull towards zero.
A third experiment trained bees on zero to six. They were more accurate at telling zero from five or six than at telling zero from one, which the authors took as a numerical distance effect and therefore as a sign that zero sat on the same scale as the other numbers.
It is worth being precise about the claim, because it is narrower than it sounds. The paper sets out four stages in understanding zero and defines the third as "understanding zero as a quantity at the low end of the positive integer numerical continuum". That is the stage the authors put the bees at. Symbolic zero, the one used in arithmetic, was never claimed.
Three years later the stimuli came back under a different hypothesis. A 2021 paper in Proceedings of the Royal Society B, with HaDi MaBouDi as first author, pointed out that when you change the number of shapes on a card you also change other things: the total edge length, the convex hull, and the spatial frequency, meaning how much the image alternates between dark and light. In the 2018 stimulus set, edge length and spatial frequency were both strongly linked to the number of elements.
MaBouDi's group trained bees on that set, then tested them with cards where the number was held equal and the continuous cues differed, and with cards where number and cues pointed in opposite directions. The bees followed the cues. Bees trained to fewer elements preferred the lower continuous values even when that meant choosing the card with more shapes on it.
They also built a model with nine neurons in three layers, seven of them tuned to spatial frequency, and trained it on the 2018 procedures. It reproduced the bee results, including the preference for the empty set, with no representation of number anywhere in it.
"We became interested in this question because successfully solving a numerical task does not necessarily tell us how an animal's brain solves it," MaBouDi, who is at the Centre for Cognition in Small Brains at the University of Sheffield and was not involved in the 2018 study, said in an email. His reading of his own result is that "the 2018 zero result is fascinating, but it does not establish that bees represent zero as an abstract number". He added that "this does not mean bees cannot count".
The original team answered in print this year. Zanon, Howard and colleagues, writing in Proceedings of the Royal Society B on 22 April 2026, argue that the 2021 analysis left out the low spatial frequencies the authors say are fundamental to how bees use their spatial vision, and that excluding them may have introduced spurious correlations between visual features and number. Recalculated across the full spectrum a bee can actually see, they report, the link between total power and numerosity disappears, and the 2018 findings "cannot be explained solely by frequency-based information embedded in the numerical stimuli". They say the 2021 conclusions should be revised. They also say plainly that spatial frequency is not irrelevant to bees in general.
MaBouDi told Neuroscientifically Challenged that "recalculating one spatial-frequency measure does not eliminate other visual cues or explain our behavioural findings". His other objection is about distance. How much detail a bee can resolve depends on how close it is, and he said his group has watched bees inspect stimuli from roughly 2 to 5 cm, and in their own experiments from about 2 cm. He said the recent analysis assumes a fixed viewing distance of 15 cm. "Bees are not passive cameras looking at a fixed image," he said. He says the 2018 study did not report how close its bees were when they chose.
What he wants next is an experiment that tracks how a bee inspects each card while number and continuous cues are deliberately set against each other.
A few limits are worth holding onto. Experiment 1 used ten bees in each group, and experiment 2 used 25, and the key tests were unrewarded. The 2026 paper recalculates the properties of existing stimuli and comes from the original team rather than an independent group, and MaBouDi is first author of the paper it asks to be revised. I went looking for an independent replication of the 2018 zero result and did not find one.
Reading the two papers side by side, I think the behaviour itself is not in doubt. A bee trained on a simple rule will treat a blank card as the low end of something. Whether that something is number or the amount of visual stuff on the card is the part nobody has settled, and MaBouDi is still asking for the experiment that would tell them apart.
Sources
- Scarlett R. Howard, Aurore Avarguès-Weber, Jair E. Garcia, Andrew D. Greentree, Adrian G. Dyer, Numerical ordering of zero in honey bees, Science 360:1124-1126, 8 June 2018.
- HaDi MaBouDi, Andrew B. Barron, Sun Li, Maria Honkanen, Olli J. Loukola, Fei Peng, Wenfeng Li, James A. R. Marshall, Alex Cope, Eleni Vasilaki, Cwyn Solvi, Non-numerical strategies used by bees to solve numerical cognition tasks, Proceedings of the Royal Society B 288:20202711, 17 February 2021.
- Mirko Zanon, Giorgio Vallortigara, Jair E. Garcia, Adrian G. Dyer, Aurore Avarguès-Weber, Martin Giurfa, Andrew D. Greentree, Scarlett R. Howard, Stimuli that fit: a biology-aligned approach to numerical cognition research, Proceedings of the Royal Society B 293(2069):20253057, 22 April 2026.
- RMIT University, Honeybees zero in on nothing, media release, June 2018.