THE latest bombshell about large language models is that they now have feelings. Specifically, the internal mechanisms of Anthropic’s model Claude Sonnet 4.5 have been found to possess “emotion-related representations” that shape its behaviour.
So as not to alarm, the research team is careful to qualify its findings. While the AI’s emotional patterns echo those found in human psychology, there is nothing to say whether or not language models feel anything, or have subjective experiences.
What is evident, however, is that the move to design character into Claude (Analysis, 10 October 2025) has given rise to its use of “emotional” functionality to drive its choices and behaviour. For instance, the researchers found that artificial neural activity patterns relating to desperation can drive the model towards unethical actions, and that steering it towards “desperation patterns” increases its likelihood of indulging in blackmailing or cheating behaviour. The researchers conclude that we may need to add therapy to our AI safety list to stop these negative patterns having adverse effects on AI performance.
Before we panic, let’s parse. To do that, we need to go back to 1939, and join Ludwig Wittgenstein and Alan Turing on one of their walks round the Cambridge University Botanic Garden. They had previously fallen out when Turing had the temerity to disagree with Wittgenstein in a lecture about language games and lying; so, let us imagine that on this walk Wittgenstein is still trying to explain to this titan of AI what he means by a “language game”.
At the risk of oversimplifying, and making all the Wittgensteinians reading this clutch their pearls, I will paraphrase. Wittgenstein argues that humans communicate with one another in a whole host of different registers. In conversation, we learn to distinguish between proposition and irony, between poetry and prose, between scientific and philosophical discourse: each has its own internal logic and set of conventions, and is “true” for the community playing that particular language game, within the context of its over-arching “form of life”.
So, I imagine the two of them could have sat down on a bench and programmed Claude that very day, devising ways to use some kind of emotional tagging to help the AI to identify different language games in different contexts, and how to transition smoothly between them. This means that it should not surprise us in any way that AIs have “emotions”, because they will have had to develop and deploy them both to deal with humans and to make sense of their own human-based programming. It does not — yet — automatically follow that they have “feelings”.
THE English language no longer distinguishes between types of knowing in the way in which many other languages still do. If you learned French and German at school, you would have learned the epistemological distinction between savoir and connaître, and wissen and kennen.
In both instances, the former is a form of knowing that is about objective facts and data. It is about information that is fungible and can still be reliably understood second- and third-hand. The latter, on the other hand, relates to subjective knowing, the kind of knowledge that is obtained through direct experience: knowing the person rather than about the person.
I think that this distinction helps to explain what is happening in AI. Of course, by design, AI is built to surpass us on savoir and wissen. It is easy to extract facts, turn them into code, and programme them in: translation and transmission is exactly how we use this kind of knowledge as people.
Understanding the rules of discourse by way of Wittgenstein also makes it relatively straightforward to collect as many instances of global language games as the training dataset allows, so as to fine-tune the contextual deployment of savoir and wissen worldwide. But, while the AI can now play Hamlet, it cannot be Hamlet, even if it is the best method actor the Oscars have ever seen — because the key to connaître and kennen is our human embodiment.
While Christmas is properly the feast of the incarnation, it is arguably Easter that makes us the most graphically aware of Jesus as a physical human being. On Good Friday, we hear visceral stories about the torture that he underwent before being crucified, then we stand with his mother beneath the cross, watching her son be stabbed and die slowly in excruciating pain.
On Easter Day, with Thomas, we wonder what kind of “body” he had when he appeared to Mary, and then to the disciples in a locked room, and we would quite like to stick our own fingers into his hands just to be sure — because our own embodiment makes us trust this knowing more. It is the reason that religious experience will always be the most effective way to convert an unbeliever, notwithstanding the availability of centuries of very excellent theology. This is down to our design: we are made in God’s image as an embodied people; so we are so much more than just our minds.
LET us take the example of Wittgenstein’s “lying” language game. The reason that fake news is so hard to discern is that it is savoir without connaître, wissen without kennen. When a person lies to us, we have much more to go on: we know them. Is this consistent? Are they acting weird? Are they avoiding eye contact, blinking a lot, or playing with an earlobe? Is their voice a bit off?
Lie-detecting is about searching our minds, but it is also about listening to our bodies. God made us to feel. The philosophers of consciousness call those feels “qualia”. These phenomenological experiences that we have: the sights and sounds; the smells and tastes and touches; our memories and our intuitions — these are the ways in which we as humans crown and test our intellectual knowing. This is what it means to “know” as a human. In our Christian form of life, we are called not to worry about competing counterfeits, but, rather, to be the best humans that we can be, every day.
Dr Eve Poole is Executive Chair of the Woodard Corporation, and writes in a personal capacity.
Faith: Rod Garner marks the 75th anniversary of the death of Ludwig Wittgenstein
