Using AI Ethically in Animal Behaviour Work: A Welfare-First Framework

28 April 2026 · Rachel Trafford

Using AI Ethically in Animal Behaviour Work: A Welfare-First Framework

AI is entering animal behaviour work quickly, and most of the conversation about it is either breathless enthusiasm or blanket suspicion. Neither is useful. The technology is here, trainers and owners are already using it, and the real question is not whether but how, and specifically, how to use it in a way that puts the animal's welfare first.

This sets out a framework for that. It comes from building these tools and worrying about exactly this question. It is written for professionals, but the principles apply to anyone using AI to make decisions about an animal's behaviour.

Why this needs saying at all

Generic AI tools that analyse dog behaviour are now easy to build and easy to find. Many of them are made by people who understand the technology but do not understand animal behaviour. They produce confident, professional sounding assessments with no welfare framework behind them, no grounding in behavioural science, and no awareness of how a wrong assessment can harm an animal.

The harm is not hypothetical. An AI that misreads a fearful, appeasing dog as calm and relaxed could lead an owner to push the dog into a situation that frightens it. An AI that misses the precursor signals before a bite could give false reassurance about a genuinely dangerous dog. An AI that frames a stressed dog's behaviour as naughty rather than frightened could steer someone toward punishment when the dog needs the opposite. When the output of a tool shapes how a living animal is treated, the ethics are not optional.

Principle 1: The animal's welfare comes before the convenience of the tool

This is the foundation everything else sits on. The purpose of analysing behaviour is, ultimately, to improve the animal's life. Any use of AI that loses sight of that, that prioritises a slick output, a confident answer, or a quick result over the actual welfare of the animal, has failed regardless of how impressive it looks.

In practice this means the analysis should be built to interpret behaviour through a welfare lens, to recognise fear, stress, and conflict for what they are, to avoid framing distress as disobedience, and to flag genuine welfare concerns rather than smoothing them over for a tidier report.

Principle 2: The science has to be in the framework, not assumed

A general AI model has no commitment to any behavioural framework. Left to itself it will produce a plausible sounding mixture of folk wisdom, outdated dominance theory, and whatever was most common in its training data, which includes a great deal of unscientific content about animals.

Using AI ethically in animal behaviour work means the science is built into the structure of the analysis, not left to chance. An observational rather than interpretive default. A recognised structure such as antecedent, behaviour, consequence for organising what is seen. An emphasis on describing what is actually visible rather than inferring motive. And, importantly, a grounding in functional analysis, looking at behaviour in terms of what it achieves for the animal, what it produces or avoids, rather than by how it happens to look in a single frame or by an assumed inner motive. A dog that growls to create space and a dog that growls in play look different and mean different things, and functional analysis is what keeps the assessment honest about which is which. A general model has no instinct to think functionally. It has to be built to. The science should be a constraint on the analysis, because the model will not supply it on its own.

Principle 3: The AI must be allowed to be uncertain

One of the most dangerous properties of AI text is that it always sounds confident. A responsible tool deliberately works against this. It gives the model permission, and instruction, to say it cannot tell from this footage, that a signal is ambiguous, that the animals cannot be reliably distinguished, rather than fabricating a confident answer to fill the gap.

This matters for welfare because false confidence leads to wrong decisions about real animals. A tool that flags its own uncertainty lets the professional know exactly where their judgement is most needed. A tool that hides its uncertainty behind fluent prose invites the user to trust something that should not be trusted.

Principle 4: The human professional stays the decision-maker

AI in animal behaviour work should be an assistant, never an authority. The professional interprets, decides, and carries responsibility. The tool's job is to support that, to provide a thorough, consistent first pass, to surface what might be missed, to save time on documentation, not to replace the judgement, the case history, the relationship, or the accountability.

This is not only about quality, though it improves quality. It is about responsibility. When a decision affects an animal's welfare, a person has to own that decision. An AI cannot be accountable to a dog. A professional can.

Principle 5: Honesty about limitations is itself an ethical obligation

A tool used in welfare sensitive work has a duty to be honest about what it cannot do. Overclaiming, "understand exactly what your dog feels," is not just marketing puff. In animal behaviour work it is a welfare risk, because it encourages people to trust assessments they should not, and to act on them in ways that affect a real animal.

The ethical version of these tools is upfront about the failure modes, the breeds that get misread, the situations where it cannot tell animals apart, the difference between describing a behaviour and understanding its cause. Honesty about limitations is not a weakness in the product. In welfare work, it is a core feature.

A framework, briefly stated

Pulling it together, ethical use of AI in animal behaviour work means, first, welfare put before the tool's appearance of competence. Second, science built in, evidence based frameworks and functional analysis structured into the assessment rather than left to the model's defaults. Third, uncertainty permitted, the tool flagging what it cannot determine rather than fabricating confidence. Fourth, the human in charge, interpreting, deciding, and accountable. And fifth, honesty about limits, because overclaiming in welfare sensitive work is itself an ethical failure.

None of this is built into AI by default. All of it has to be deliberately constructed around the technology by people who understand both animal behaviour and the stakes. That construction, the welfare lens, the scientific framework, the functional thinking, the permission to be uncertain, the human in the loop design, the honesty, is the difference between a tool that helps animals and one that risks harming them while looking impressive.

Why this matters for the profession

As AI spreads through animal behaviour work, the profession has a choice about what standard to hold it to. The standard should be high, because the subject is a living animal that cannot tell us when we have got it wrong. Trainers and behaviourists are well placed to lead this conversation, to insist that tools used in their field meet a welfare first, science based bar, and to reject the ones that do not.

That leadership is needed. The technology will keep arriving whether or not the profession shapes it. Far better that the people who actually understand animal welfare are the ones setting the terms.

MyCanine360 is built on exactly this framework, welfare first, grounded in behavioural science and functional analysis, designed to flag uncertainty, with the trainer always the decision maker. It exists because we wanted AI in animal behaviour work done properly. [Learn more / try it].

Related reading:

• Can AI Read Dog Body Language? An Honest Look at What It Can and Cannot Do
• How AI Video Analysis Is Changing Dog Training