A 25-Year-Old Game Passes the Same AI Consciousness Tests LLMs Do
Researchers built a neural network inside Age of Empires II and used it to argue that every anthropomorphic claim made about large language models applies equally to medieval virtual peasants farming grain. The paper's punch line: if substrate doesn't matter, neither does the silicon.
Adrian de Wynter at arXiv trained a simple neural network using AoE II's in-game logic to demonstrate that "human-like" behavioral attributes are empirically non-unique to any particular substrate.
The paper proposes a "null assumption" framework: treat LLM anthropomorphism as the hypothesis needing proof. By that standard, a glass of water and a 1997 RTS share the same epistemic standing as GPT-4.
HN erupted. Critics called it "genuine crackpottery"; supporters said it lampoons Substack essays that deduce consciousness by analogy alone.
The Paper That Made AoE II a Consciousness Substrate
Adrian de Wynter's arxiv paper arrived on HN with a title that reads like satire but isn't: If LLMs Have Human-Like Attributes, Then So Does Age of Empires II. The paper reached 104 points and 103 comments in the thread, its ratio suggesting deep engagement rather than casual upvotes.
What the Paper Actually Argues
De Wynter's move is surgical. He does not argue that LLMs are or aren't conscious. Instead, he constructs and trains a neural network inside Age of Empires II — using the game's scripting and in-game mechanics as a substrate — and then shows that any claim of anthropomorphic uniqueness made for LLMs applies equally to that absurd setup. His term: LLM anthropomorphism is "empirically non-unique" across substrates.
The conclusion follows from Turing completeness: if AoE II can be shown to be Turing-complete (the paper provides a specific proof requiring two markets, a town centre, a trade cart, six villagers, five farms, and a suicidal scout), then any computation an LLM performs can in principle run there. The paper proposes a "null assumption" — researchers studying LLM cognition should assume non-uniqueness by default and design experiments that actually falsify it, rather than reading human-like qualities off the output layer.
The broader targets are the academic papers and Substack essays that run down a list of brain-like processing events in LLMs, declare functional equivalence, and stop there.
The Thread Tears Into Both the Paper and the Anthropomorphizers
The sharpest critique came from a long-tenured HN contributor who called the paper's logic incoherent from the ground up:
"They conflate the substrate with the presentation layer. Then they point out that Turing equivalence means you can run an LLM on anything… Yes, this seems to imply if your partner writes something heartwarming to you using LEGO, you're meant to laugh at them and point out how LEGO isn't intelligent so this isn't evidence of anything." — a commenter active since 2014
A reader who spotted the paper's actual target audience offered a different read:
"It's showing flaws in methods other academic studies used to determine behavior, so it's aimed at people that create and review academic studies. It's not a very large audience." — a commenter active since 2024
The most pointed philosophical reversal came from a reader who noted the paper's target may already agree with it:
"I think the majority of people who are happy to anthropomorphise LLMs… are already completely happy with the notion that a computer game might have elements that are human like. They've already accepted that key aspects of being a human are substrate independent, so why would the idea of a computer game as substrate be disconcerting to them? There's no bullet left to bite here." — a commenter active since 2012
One commenter extended the reductio further than the paper does:
"There exists some isomorphism between the velocities of the molecules in a glass of water, and the states of a Turing machine simulating a human mind. So is the glass of water conscious? Actually there are many such isomorphisms to many possible conscious minds, so is every glass of water simultaneously having every possible conscious experience?" — a commenter active since 2016
Meanwhile, a commenter who traced the social context diagnosed the epidemic the paper is responding to:
"LLM enhanced essays in deep language about functional equivalence between mental states as they're known in humans and in human brains… Something about it seems to abuse the power of analogies to draw connections, treating view from 10,000 feet comparisons like they're proof of identity." — a commenter active since 2008
The Rule the Paper Reveals
The paper's contribution is less a proof than a diagnostic. Turing completeness has always implied that substrate is the wrong place to look for cognitive uniqueness — but nobody bothered to build the reductio ad absurdum in a medieval farming game before. De Wynter's null assumption is the useful residue: if your methodology for detecting consciousness can't distinguish an LLM from Age of Empires II, the methodology is the problem. That's the standard academic psychology and AI cognition research are failing to meet, and the paper names it directly without pretending to solve it.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Netlify CTO says coding is dead — HN asks if she ever coded
Netlify CTO Dana Lawson told an audience that "writing code is no longer the job" and predicted a billion new applications by 2029. The HN thread responded by checking her LinkedIn and concluding she may have never held a developer role.