Researcher: LLMs Are No More Sentient Than Age of Empires II

The sentience debate around large language models has produced confident claims and very few falsifiability tests. Adrian de Wynter, a Microsoft AI researcher, built one: he constructed a working NAND gate and 1-bit perceptron inside Age of Empires II using goats as signal carriers, then wrote a paper whose stated aim, visible in 404 Media's reporting by Matthew Gault, is "to formally show that we anthropomorphise too readily." The YouTube video on the story frames the same argument.

The mechanism is deliberately simple. Goats move along rails inside the game's scenario editor, each one carrying a signal through the logic circuit. "Only one rail is active at a time, with a goat acting as the signal carrier. When the gate fires, the bit-goats are removed (they ded), and a new bit-goat is placed in its respective output rail." The argument is a reductio: if what makes a system conscious is implementing the same computational primitives that underlie neural networks, Age of Empires II qualifies. If that conclusion seems absurd, the consciousness criteria need to be a lot more precise than much of the literature currently requires.

The literature review is the more striking part. De Wynter analyzed 337 computer science papers and found that "57% of them treated LLMs as if they had some form of consciousness or personality." He attributed this to confirmation bias: people find human qualities in AI when they expect to see them. He also named a less charitable explanation, that "implying chatbots have some kind of consciousness may just be a good marketing ploy by the companies involved." Science fiction writer Ted Chiang had made a parallel observation in an essay the 404 Media piece references: "Being open to the possibility that LLMs are conscious is the same as being open to the possibility that Microsoft Word is conscious."

What the reporting does not give you is the peer-review status of de Wynter's paper, or the exact criteria it used to flag a paper as anthropomorphizing — both matter for how much weight the 57% figure should carry. The AoE2 analogy is also by design extreme, and whether it holds as AI architectures grow more complex is a real open question. But the core challenge is hard to sidestep: a consciousness definition loose enough to cover a 1-bit goat-powered perceptron is probably too loose to do meaningful work.