How bee cognition is tested

The proboscis extension response

A bee is harnessed so that only its head and antennae can move. A scent is presented, and immediately afterwards a droplet of sugar touches the antennae, which makes the bee extend its tongue. After a few pairings the scent alone produces the extension. The response is binary, easy to score, and can be run on hundreds of individuals.

It is the workhorse of olfactory learning and memory research, and it is where most of what is known about memory phases and their molecular requirements comes from. Its limitation is that a harnessed bee cannot move, choose, or do anything the experimenter did not build into the apparatus.

Free-flying choice tasks

A bee flies into a Y-maze or an arena and meets two or more stimuli. One is paired with sugar; the other with plain water or, in better designs, with quinine, which bees find aversive. After training the bee is tested on stimuli it has never seen that share the trained property.

The transfer test is where the science actually happens. A bee that has memorised two pictures will fail it; a bee that has learned a rule will pass. Every claim about concepts, quantity and relations rests on transfer performance rather than on training performance, and any report that quotes only the training result has left out the finding.

Observation and manipulation tasks

A bee has to pull a string, roll a ball, rotate a lid or push a tab to reach a reward. These designs test problem solving, social learning and behavioural flexibility, and they are where most of the recent striking results come from.

They are almost all done with bumblebees rather than honey bees, for practical reasons: bumblebee colonies are small, individuals are large, easily marked and will work in an arena for long periods. That practical choice is the reason so much popular coverage attributes bumblebee abilities to honey bees.

What none of these designs can do

• They cannot ask a bee what it perceives. Every result is inferred from a choice under reinforcement.
• They cannot distinguish understanding from a successful strategy, because a successful strategy is all the apparatus can see.
• They cannot address subjective experience. Judgement-bias designs detect a state that shifts decisions; they are silent on whether anything is felt.
• They cannot rule out an unmeasured cue. They can only rule out the cues somebody thought to control.

How to read a bee cognition paper in five questions

1. Which species? Apis mellifera and Bombus terrestris are not interchangeable.
2. Was there a transfer test with novel stimuli, or only training performance?
3. What were the controls for correlated low-level cues?
4. How many individuals, from how many colonies?
5. Has anybody independently replicated it?