Source paper:https://arxiv.org/pdf/2605.31514
If a large language model (LLM) is considered to have "human-like attributes" — emotion, self-awareness, moral judgment — then, in principle, any sufficiently powerful "substrate" could exhibit the same attributes, including Age of Empires II, a 1999 real-time strategy game. This paper doesn't argue about whether LLMs "really" have these attributes. Instead, it shows that the experimental designs currently used to argue "LLMs do/don't have human-like attributes" are logically unsound regardless of their conclusion — experiments that conclude "yes" are circular, and experiments that conclude "no" are uninformative. The author supports this claim in three steps: (1) actually building and training a neural network perceptron inside the Age of Empires II game engine, and proving the game is functionally and Turing-complete; (2) using formal logic to prove that, under either "accepting" or "rejecting" a given theory of mind (such as the computational theory of mind), the resulting conclusion necessarily falls into one of two failure modes — circularity or uninformativeness; (3) proposing a "null assumption" methodology — making no prior assumption about whether the system has anthropomorphic attributes, and measuring only observable behavior itself.
Imagine this absurd scenario: you take the core computation of a large language model and transplant it, verbatim, into Age of Empires II — using the game's sheep, villagers, and resource markets to implement the exact same input-output mapping. You say to this "AoEII-LLM," "I feel lonely," and it replies, through some agreed-upon in-game representation: "I feel bad for you. Maybe catch up with a friend? Closeness always helps in these situations."
That reply makes it genuinely hard to say: does this system, made of sheep and resource prices, actually "know" anything about what helps with loneliness? Does it "really" have empathy? Or is its output convincing only because you happen to be able to read the representation?
This is exactly where the paper's core argument starts. The author argues that the reason this question is so awkward is precisely because anthropomorphic traits are, at bottom, dependent on the mode of representation, not on what the system is computing underneath. The exact same logical operation looks "human" wrapped in a low-latency chat box that speaks natural language; it looks nothing like a human wrapped in a herd of virtual sheep wandering back and forth on grass — even though the underlying computation could be identical. This means the anthropomorphic attributes an LLM displays are, empirically, not "unique" to LLMs at all: any sufficiently powerful substrate could carry the same computation, and only the mode of representation changes how a human interprets it.
To keep this from being just a thought experiment, the author actually did something concrete: inside Age of Empires II's scenario editor, using the game's own units and terrain rules, he built a NAND logic gate — the "universal" basic building block of digital circuits; in principle, given enough NAND gates, you can assemble any computational circuit. Based on this NAND gate, the author proves Age of Empires II is "functionally complete" and "Turing-complete" — meaning that, in principle, you could build inside this game anything any computer can compute, including a neural network. The author actually trained a simplified perceptron (the most basic neural-network unit) inside the game, and it successfully learned the logical AND operation. This isn't a metaphor — it's a working, reproducible construction.
What follows is the sharpest part of the paper. Suppose a scientist wants to figure out: "do LLMs actually have genuine anthropomorphic attributes?" She first has to pick a stance: either assume that "the mind is a kind of information-processing system" (the computational theory of mind, CTM) is true, or reject it. Whichever she picks, the author proves with formal logic that the experimental result has only two possible outcomes, and neither is useful. If she "accepts CTM," then "anthropomorphic attributes exist" and "the experiment got a positive result" are, in essence, the same claim restated — this is circular reasoning, proving nothing independently. If the experiment gets a negative result, there's no way to tell whether "the assumption was wrong" or "the experiment itself was flawed" — this is an uninformative result. Extending this logic across many papers and many sets of hypotheses, the conclusion still holds: as long as the starting point of the argument already contains what you're trying to prove, no amount of experimentation escapes the loop.
So what's the solution? The author proposes something called the "null assumption": completely abandon the question of "whether this system genuinely has anthropomorphic attributes," and measure only what it "does" under specific conditions, without interpreting whether that behavior reflects "real" understanding or "real" emotion underneath. For instance, an LLM generating text that explains its own behavior is an observable, reproducible fact; but whether that text represents "understanding" of its own internal state is an additional interpretation that shouldn't be assumed by default. Separating "the pattern observed" from "the ascription made about that pattern" produces weaker conclusions (you can't say "the LLM has self-awareness") but much sounder ones (you can say "under this specific task, the LLM displays a consistent behavioral pattern").
The author designs a darkly funny thought experiment to reinforce this point: suppose you mobilize all 660,000+ residents of the Greater Boston Area, with each person representing a node in an LLM's neural network, exchanging the inputs and outputs of the computation via text message. Would this giant human-powered computing network end up "feeling anxious"? Or does the anxiety exist only in the content of one specific text message? Or in the people walking into a particular address (say, the block corresponding to the "output layer")? The author's answer: the question itself is malformed — you cannot ask "does this system have anthropomorphic attributes" without making any assumptions at all, because the answer to that question depends entirely on which theory-of-mind stance you've already silently accepted at the start.
The paper closes by citing and adapting a classic principle from animal-cognition research (Morgan's Canon): never explain a behavior in terms of a higher cognitive process if a lower one will fairly explain it. The author's final position is restrained: he doesn't claim LLMs definitely lack emotion or consciousness — he says that, absent a clearly stated measurement standard, the sentence "the LLM displayed a human-like trait" proves no more than the sentence "the sheep in Age of Empires II displayed a human-like trait."
Anthropomorphism / Anthropomorphic Ascription: the tendency to attribute human mental traits (emotion, intent, understanding) to a non-human entity. The paper distinguishes between "a system displays a behavioral pattern" (an observable fact) and "this pattern represents some internal mental attribute" (an interpretation/ascription).
Substrate Non-Uniqueness: the paper's core claim — the anthropomorphic traits an LLM displays are not, empirically, unique to LLMs, because the same input-output behavior can be reproduced on a different "substrate" (a different hardware/implementation), with only the mode of representation changing how an observer interprets it.
Computational Theory of Mind (CTM): a position in the philosophy of mind holding that the mind is, at bottom, a kind of information-processing/computational system. The paper uses "accepting CTM" or "rejecting CTM" as the example of the stance a scientist must pre-select before designing an experiment about anthropomorphic attributes (any analogous framework would apply just as well).
Functionally Complete / Turing-Complete: functional completeness means a set of logical operations (like NAND) can compose any Boolean function; Turing-completeness means a system can, in principle, simulate any Turing machine and execute any computable operation. The paper proves Age of Empires II satisfies both properties.
Circular vs. Uninformative: the two experimental failure modes the paper characterizes with formal logic — circular means the experiment's "evidence" and the "hypothesis" it's meant to prove are, at bottom, the same claim; uninformative means the result cannot distinguish whether the original assumption was wrong or the experimental design itself was flawed.
Null Assumption: the methodology this paper proposes — distinct from the statistical "null hypothesis" — meaning that when measuring anthropomorphic attributes, the researcher makes no prior judgment about their existence or non-existence, and measures only "observable, implementation-defined behavior" itself.
A large body of LLM research (the paper surveys 315 related works published between mid-2024 and mid-2026) defaults to or explicitly assumes LLMs possess anthropomorphic attributes — cooperation, empathy, psychological principles and values, introspection and self-awareness, natural-language understanding, anxiety, even deceptive capability. Some of this research explicitly labels these traits "emergent behaviors" rather than a product of training. AI companies are also actively reinforcing this perception by design, having chatbots say things like "I'm confident," "I can understand and respond to your emotions," or "I believe..." — partly for commercial reasons related to improving user experience, but this has also produced documented harms: user attachment, overreliance, and reinforced delusion.
The paper's core implicit demand is that any research, claim, or policy discussion concerning whether LLMs have generalized anthropomorphic attributes must explicitly state the scope of its experimental assumptions and the boundary of its conclusions' applicability, rather than jumping directly from an implicit or explicit "accept/reject" stance to a generalized "this attribute exists/doesn't exist" conclusion. This also implies that future evaluation should explicitly separate "what an AI can objectively do" (a testable capability) from "what the experimenter believes it should be" (an anthropomorphic ascription).
"For this we build and train a simple neural network on the video game Age of Empires II, and note that any entity in a sufficiently-powerful substrate, such as LEGO or the Greater Boston Area, could also present such attributes."
"we also construct a physical system, in, say, the Greater Boston Area, which has 667,137 inhabitants... Would the Greater Boston Area present human-like attributes (excluding these in the 'neurones') through autoregression? ... we do note that one cannot begin an evaluation of such a system by assuming that it has them intrinsically."
"Paraphrasing and adapting Morgan (1903)'s canon from animal cognition into AI, we argue 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."
A faithful reading and plain-language retelling of the paper, generated by PodLens.
This is one source-grounded reading, not a replacement for the original. Every point is anchored to its source, so you can check it yourself — and corrections are welcome.