Wittgenstein and why AI cannot talk to animals

Could AI Translate Animal Languages?

Claims are being made that artificial intelligence will allow us to translate animal language. Following Wittgenstein, Constantine Sandis thinks different.


The growing market for ‘pet communication buttons’ is leading to a rise in the number of extremely impressive videos and blogs from pet owners documenting how their animals – particularly canines – use talking pet buttons. I recently tried a couple of these button sets on my dog, Calypso, and she gave them a bit of a half-hearted chew before losing all interest, so we won’t be starting any TikTok trends just yet. What is more, recent headlines suggest that AI, and in particular large language models, can help us translate the sounds animals make to communicate with each other into human language. Wittgenstein's philosophy can help us understand why such far-fetched claims of animals communicating in a human language, and humans understanding animal language are misleading.

The basic premise of pet communication is simple enough. The cats, dogs, birds, and other non-human animal companions with whom we share our lives clearly have a working understanding of the words or phrases that we use to communicate with them: ‘wait’, ‘eat your food’, ‘where’s your ball?’, ‘walking home’, etc. Although philosophers of various stripes may concoct fancy arguments in defence of the view that non-human animals have no language or concepts, there are no fixed parameters to what counts as a language, and only someone in the grip of a philosophical theory would be tempted to deny that our fellow creatures have any kind of grasp (however limited) of certain concepts.


Animals may also have concepts that humans lack, for to share a concept is to share a form of life and the everyday practices that go along with this.


The main technology used by ‘pet communication buttons’ dates back to the century-old electric era of sound recording. By pre-recording ourselves uttering specific words and phrases that our animal companions already respond to, we give them the tools not only to respond to our calls but also to actively request such things as food, water, or cuddles whenever they want them. They do this not through their usual range of pet sounds and gestures, but rather by recalling the position of the appropriate button. They can literally push the button and let us know. Some non-human animals are even able to form primitive sentences using their names, as well as those of other animals in the household. By memorising which button releases which sound, the button-pressing behaviour of your animal companion replaces their more primitive physical behaviour and accompanying sounds. The process is eerily similar to that hypothesised by Ludwig Wittgenstein in relation to how a child may come to say ‘I am in pain’ rather than crying.

Here is one possibility: words are connected with the primitive, natural, expressions of sensation and used in their place. A child has hurt himself and he cries; then adults talk to him and teach him exclamations and, later, sentences. They teach the child new pain-behaviour.

Wittgenstein, Philosophical Investigations, § 244

At this point in the text, Wittgenstein’s imaginary interlocutor asks ‘So you are saying that the word “pain” really means crying?’, to which Wittgenstein replies: ‘On the contrary: the verbal expression of pain replaces crying, it does not describe it’ (ibid.) Something similar is happening in the case of pet communication buttons. Here, the animal does not itself utter the relevant word or phrase, but rather causes it to be voiced by pressing the appropriate button(s). Suppose the recorded words on one of these buttons are ‘Calypso walk’. Their transmission would be replacing my dog’s prior behaviour of looking at the door while raising her inner eyebrow, but they would not be describing it; they perform the same function. Indeed, a cat or dog can learn to express their desire, without needing a concept of desire itself.


It is tempting to think that animals talk to one another in their own languages and that if we could only decode them that we could come to communicate with them.


But animals may also have concepts that humans lack, for to share a concept is to share a form of life and the everyday practices that go along with this. The average human will do more things together with cats and dogs than with lions and cheetahs, say, but we ultimately share far more with other humans, however different our cultures or age gaps, for instance, may be. It is tempting to think that animals talk to one another in their own languages and that if we could only decode them that we could come to communicate with them in a Doctor Dolittle kind of way. And if such communication is possible, then surely AI can help us achieve it; if AI can offer up basic (albeit amusingly fallible) translations of written and spoken human languages, then why should it not also be used to decode those of other animals?

related-video-image SUGGESTED VIEWING AI consciousness cannot exist With Markus Gabriel

This is the kind of thinking that has led to headlines such us ‘Artificial Intelligence Could Finally Let Us Talk with Animals’. The technology behind the hype is impressive in its own right, helping us to better understand both wild and domesticated animals. Projects such as CETI (Cetacean Translation Initiative), for example, aims to utilise machine learning and robotics engineering to study how whales communicate. The technology will allow researchers to interact with whales in a non-invasive way, while the AI facilitates the speedy analysis of data and related pattern detection in the vocalisations of sperm whales. It does not, however, enable us to ‘understand what whales are saying’. Indeed, it would probably be a mistake to think that whales say things to one another in the same way that we humans do, let alone that they would speak to us. If a whale could speak English, it would not say things like ‘take this camera away from me’. To do so, it would need (among other things) the concepts of a camera and of taking something away.

tn 4 SUGGESTED READING Wittgenstein vs Wittgenstein By Lee Braver

‘If a lion could speak, we couldn’t understand it’, Wittgenstein famously quipped in his Philosophical Investigations (Part II, p. 235) This is not to say that when it comes to lions and whales there is nothing for us to understand, but rather that there is far more to understanding another being than having their sounds and gestures translated by some machine. At a minimum, we would need to understand why any given animal thinks, feels, and acts as it does, and to do so we would need to share certain concepts, and the practices that give rise to them (it is no accident that Wittgenstein chose a wild animal and not a domestic one).


No device could possibly translate animal sounds into anything that would make straightforward sense to the average human.


Ethologists have long been able to associate particular animal sounds with certain kinds of behaviours. While language models can be highly efficient at detecting and sorting out such patterns, they aren’t ‘deciphering meaning’, and there is no such thing as ‘ AI is decoding the animal kingdom’, as many sensationalist headlines would have it. No device could possibly translate animal sounds into anything that would make straightforward sense to the average human.

related-video-image SUGGESTED VIEWING The dream and danger of AI With Liv Boeree, Dominic Walliman

AI is vital in helping us to better understand animal behaviour and the reasons for it, if only because of the sheer speed with which it can organise huge volumes of data, thereby enabling the analyst to better detect patterns between animals’ sounds and movements in their interactions with us. But you’d be barking mad to think that AI can translate pet sounds into human languages. Consider, for example, the impressive AI system CHAT (Cetacean Hearing and Telemetry) created by Thad Starner and his team at Georgia Tech:

Working with Dr. Denise Herzing of the Wild Dolphin Project, we are creating wearable computers for conducting two-way communication experiments with cetaceans. With CHAT, one researcher uses the waterproof system to broadcast a sound, associated with an object with which dolphin’s [sic.] like to play. A second researcher, upon detecting the sound, passes the object to the first. The researchers pass objects back and forth, further associating the sound with the object. The goal is to see if the dolphins mimic the sound in order to “ask” for the play object. The wearable computer uses pattern recognition technology to detect these mimicked sounds. In a more long-term effort, UHURA uses pattern discovery techniques in an attempt to uncover fundamental units of dolphin vocalisations.

Computing-enabled and interactive technology for and with animals allows researchers to teach words to dolphins, much as one might do with a cat or dog. This in turn facilitates human–dolphin interaction. Should the dolphin succeed in mimicking the sounds with intent, its behaviour would be similar to that of a button-pushing canine. What the technology quite evidently does not do is to translate ‘dolphin language’ into any human language.

The headlines are misleading, but this is not to say that AI is not going to become an increasingly vital component to communicative interaction between humans and other creatures. What AI cannot do is to replace the key activity of doing things together. For it is only through shared behaviour that we may come to share any concepts at all, and there can be no possibility of translation without shared concepts.

Latest Releases
Join the conversation