The Neuroscience of Reward: What Prediction Error Can Tell Us About Horse Training

What does dopamine actually do in your horse’s brain?

Dopamine has a reputation it does not quite deserve. Most people, if they know anything about it, know it as the “pleasure chemical”, the brain’s reward signal, the molecule that makes things feel good. This is not exactly wrong but it is misleading enough to obscure what dopamine is really doing in any mammalian brain, including your horse’s.

One major function of phasic dopamine signalling is to encode a reward prediction error^{1, 2}: the difference between what the brain expected and what actually happened. When an unexpectedly good outcome occurs, dopamine neurons in the midbrain show a brief burst of activity. Think about how happy you’d be if you finally won lotto after many years of buying tickets. When the outcome is fully predicted, there is little or no prediction-error response at the moment the reward arrives. When an expected reward does not occur, dopamine activity can briefly fall below baseline. Disappointment, anyone?

Importantly, that does not mean a predicted reward has stopped being rewarding. As learning develops, the dopamine response can shift from the reward itself to the cue or event that predicts it. The brain has learned the relationship. Dopamine, in this context, is therefore not simply encoding ‘yum’. It is helping signal whether an outcome was better than expected, worse than expected, or about what the brain predicted.

Applying this framework directly to horses still requires some extrapolation. Direct recordings of dopamine prediction-error signals in behaving horses are not currently available, so equine researchers infer the likely role of these systems from behaviour, physiology, equine neurobiology and comparative mammalian neuroscience^{3, 4}.

Recent equine research does, however, fit well with the broader framework. Work at Nottingham Trent University has shown that physiological arousal is associated with performance on learning tasks⁵. McBride and colleagues have reviewed how the equine basal ganglia and dopaminergic systems are likely to contribute to learning⁶, while Henshall and colleagues have discussed equine learning, including aversive learning, using prediction-error mechanisms as part of the explanatory framework⁷.

The cautious conclusion is not that horses have been shown to produce exactly the same dopamine signals recorded directly from laboratory rodents and primates. Rather, the wider mammalian evidence provides a biologically plausible and useful framework for interpreting equine learning, while direct equine neurophysiological evidence remains more limited.

Dopamine is not simply a ‘pleasure signal’. In reward learning, prediction-error signalling helps the brain detect when an outcome is better or worse than expected and update its predictions accordingly.

What’s the trap most people fall into?

Now, before anyone reading this starts varying their cues or their timing in pursuit of more dopamine, let me be very clear, because this is the part that goes badly wrong if it is misunderstood.

If you choose to introduce variability once a behaviour is well established, it belongs in the reward rather than in the cue or the timing of the consequence.

The cue you give your horse for forward should mean the same thing every time. The cue for halt should mean the same thing every time. The release of pressure or the marker for a positive reinforcer should occur promptly and consistently when the correct response begins. Where a marker is used, it can bridge the short delay until the primary reinforcer is delivered. These three pillars of clarity of cue, timing, and consistency are the foundation on which all training rests, and they do not change because I am now banging on about varying something.

Once a behaviour is well established, what can vary is the type, value or frequency of any additional reward that follows the release or marker. Sometimes the established reinforcer may be sufficient. Sometimes you may add a wither scratch, food, rest, enthusiastic praise or another outcome that the individual horse values. Variation can be used to thin reinforcement appropriately, maintain engagement and occasionally make an especially good response produce a better-than-expected outcome.

A 2025 narrative review by Bradshaw-Wiley, Henshall, McLean and Freire examined combined positive and negative reinforcement in equine training¹¹. The authors identified combined reinforcement as a promising area for further investigation, but also highlighted how limited the direct evidence remains and how inconsistently reinforcement methods have sometimes been described in equine research. The review therefore does not establish that combining pressure release with a positive reinforcer is inherently more powerful than pressure release alone. Instead, it highlights the need for better research into how combined approaches affect learning and horse welfare.

If you take one practical idea from this section, take this: keep your cues and timing consistent. Once a behaviour is established, the rewards around that behaviour can sometimes vary. The cues are the words. Timing and contingency are the grammar. Rewards are the feedback. Keep the words and grammar clear, while the feedback can sometimes vary.

Keep your cues consistent. Rewards can vary once the behaviour is established.

Does the same principle apply to pressure-release?

Most reward-prediction-error research has focused on appetitive rewards, but dopamine is also involved in learning about aversive events, escape and relief. The underlying circuits and signals are not identical to appetitive learning, so it is better not to describe positive and negative reinforcement as neurologically interchangeable. The important point is that prediction and prediction error can also contribute to learning when an animal discovers how its behaviour can terminate or avoid an aversive event.

Prediction errors help update learning. Clear contingencies tell the animal how its behaviour changes the outcome.

The neuroscience

Until recently, the role of dopamine in negative reinforcement learning was much less well-studied than its role in positive reinforcement. That gap has been closing rapidly, and what the science is showing is actually really interesting. Trust me. You must, because you are still reading.

In 2021, a group of researchers led by Zhijun Diao recorded directly from substantia nigra dopamine neurons in mice learning to escape mild foot shocks¹⁴. They found that the dopamine system was bidirectionally engaged throughout negative reinforcement learning, not silent as some earlier theories had assumed. In the early phase of learning, when the mice had not yet figured out how to escape, the aversive stimulus suppressed dopamine activity, as you would expect. Like when we accidentally touch the electric fence, it’s not exactly fun. But once the mice had learned the escape response, the same aversive stimulus became associated with a brief increase in dopamine activity once the escape response was well learned, consistent with the stimulus having acquired a different predictive significance. Kind of like how when you let go of the fence, you know the shock will stop.

A 2025 study by Cheng and colleagues extended this work substantially¹⁵. Using more sensitive techniques, they tracked exactly how the dopamine signal shifts across the stages of negative reinforcement learning, and found a beautifully clean pattern:

• First, before the animal can escape, the dopamine response tracks the aversive stimulus itself.
• Then, as the animal begins to learn that the aversive stimulus reliably ends, the dopamine response shifts to track the termination of the stimulus. This shift is consistent with learning about stimulus termination and relief, although negative and positive reinforcement are not neurologically identical.
• Finally, once the animal has fully learned to escape on cue, the dopamine response shifts again, this time to track the onset of the aversive stimulus, because that onset now reliably predicts the relief that the animal can produce by performing the escape response.

These studies show that dopamine can participate in negative reinforcement learning and in the prediction of relief. They do not mean that release from pressure should be made unpredictable. Quite the opposite. In horse training, the animal needs control over the contingency: when the horse gives the correct response, the pressure should reliably and promptly reduce or cease.

Henshall and colleagues have discussed prediction-error processes as a useful framework for understanding equine learning, including aversive learning⁷, and reviews of equine neurophysiology identify dopaminergic systems as relevant to learning⁶. The defensible conclusion is therefore that negative reinforcement involves prediction and learning about relief, not that it stops teaching when relief becomes predictable.

Your horse is ALWAYS learning.

Something I find myself saying to my fellow horse peeps all the time is that every interaction with your horse is training, whether you intend it to be or not. Prediction error is one part of that much bigger learning picture. The horse is continually learning which cues predict which events, which responses change outcomes, and which behaviours are worth repeating.

As trainers, our job is not to keep the horse’s brain surprised all the time. It is to make the relevant contingencies clear. The horse should be able to predict what a cue means and what response will produce release or reinforcement. Within that predictable framework, unexpectedly good outcomes can sometimes provide additional information and help strengthen or refine learning.

So the practical question is ‘Was the cue clear? Was the response criterion appropriate? Did the consequence follow the behaviour I wanted? And was the horse in a physical and emotional state that allowed useful learning?’