People seem all too ready to anthropomorphize LLMs and AI chatbots like ChatGPT, Claude, and Gemini. Some humans even admit to ‘relationships’ with one or more of the various examples of machine intelligence. To illustrate how flawed this instinct could be, a Microsoft AI researcher built a tiny neural network inside Age of Empires II using goats, grass, and bridges. Adrian de Wynter shared his work in a paper dubbed If LLMs have human-like attributes, then so does Age of Empires II. The Microsoft researcher, based at the University of York, also talked to 404 Media recently about how he likes to turn absurdism up to 11 to make a point.
In the research paper, De Wynter doesn’t make the argument that LLMs do or do not actually have generalized anthropomorphic attributes. Instead, he illustrates that the AoEII goats can also power the kinds of models that lay behind today's most popular chatbots. That hammers home the argument that “in no case is a machine’s activity to be interpreted in terms of higher cognitive processes, if it can be fairly interpreted in terms of processes which stand lower in the scale of cognitive evolution and development.”
De Wynter also raises the well-known concept of confirmation bias. Those looking for human traits in tech like chatbots will tend to find them, he proposes. However, the big contrast between the absurdist goat example and the commercial LLM chatbot is the way people interact with them, the interface that makes the likes of Claude ‘conversation friendly.’ De Wynter’s research indicates that anthropomizing LLMs is a common trend in computer science papers. From 337 such papers De Wynter looked at, published in the last two years, he says that 57% assumed that LLMs could have human-like traits. This basic assumption could color the research, testing, and, of course, conclusions of these papers.
So, how did the Microsoft AI researcher build the goaty AoEII LLM? Well, he didn’t quite go as far as developing a full-blown LLM. Instead, De Wynter thought it sufficient to use AoEII’s scenario editor to build a working NAND gate, with a 1-bit perceptron, where the goats act as bits. This crude perceptron and the circuit to train it in-game are enough to demonstrate that the simplest building block of a modern neural network could be made this way. And if you think it is absurd that AoEII goats can embody consciousness, then it should be equally absurd to regard any of the well-known chatbots as anything more.
Companies behind the AI boom aren’t discouraging people from anthropomorphizing their wares. In many ways, they might benefit from these human perceptions. Chatbots they deploy are trained with natural language and use techniques to mimic the shape and tone of natural conversation. This makes it easy for users to project personality, emotion, or even consciousness onto them. Top AI company execs have leaned into the perception of their customers, publicly entertaining the idea that their systems could or might be exhibiting signs of consciousness. In his 404 Media interview, De Wynter also highlighted research indicating that people buy more products when they can empathize with them, and that includes AI/chatbot/LLM subscriptions.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.