Microsoft Engineer Uses Goats in Age of Empires II to Prove AI 2026

A Microsoft engineer has devised a clever way to demonstrate that AI is not self-aware: by using goats in Age of Empires II. The unusual experiment cuts through growing claims that large language models (LLMs) possess consciousness, offering a tangible test of whether AI truly understands itself.

The debate over AI consciousness has intensified as models become more convincing. In May 2026, biologist Richard Dawkins claimed that Claude might be conscious, while back in 2022, Google engineer Blake Lemoine was suspended after insisting chatbots were developing awareness. Now, researchers are pushing back with inventive methods.

The Goat Test: How Age of Empires II Exposes AI's Lack of Self-Awareness

Adrian de Wynter, a researcher at Microsoft, collaborated with New York University to explore whether LLMs have self-awareness. Their study, titled 'If LLMs have...' (as referenced in the source), used a surprising tool: the video game Age of Empires II. In the game, players can control goats, and de Wynter designed an experiment where an AI model was asked to reason about the goats' behavior and its own role.

The test revealed that the AI could not genuinely reflect on its own consciousness. Instead, it relied on patterns and probabilities, not introspection. The goats served as a neutral subject to probe whether the model could distinguish between simulated and real awareness.

Why Goats? The Logic Behind the Unusual Experiment

Goats in Age of Empires II are simple game entities with predictable behaviors. By asking the AI to describe or predict goat actions, de Wynter could check if the model confused its own computational processes with conscious experience. The results showed the AI had no true self-model—it merely mimicked understanding.

The Rise of AI Anthropomorphism: From Dawkins to Lemoine

Richard Dawkins' 2026 comments about Claude being conscious shocked many, but they reflect a broader trend. People increasingly anthropomorphize AI, attributing human-like feelings to machines that are simply good at conversation. Blake Lemoine's 2022 claims about Google's chatbot followed the same pattern, leading to his suspension.

These cases highlight a psychological trap: the more fluent and context-aware an AI appears, the more we assume it has inner life. De Wynter's study directly challenges this assumption by showing that LLMs lack the mechanisms for self-awareness. The longer history behind the story is traced in AI Sycophancy: How to Stop AI from Always Agreeing with You.

What the Study Found: LLMs and the Illusion of Consciousness

The research, conducted with NYU, concluded that current LLMs do not possess self-awareness. They can generate text that sounds introspective, but this is an emergent property of statistical learning, not genuine consciousness. The goat experiment provided a concrete demonstration: the AI could not correctly attribute mental states to itself or the goats beyond learned patterns. The background to this story is set out in AI Consciousness Debate: Why I Doubt Claude Is Truly Aware. A related thread runs through AI Consciousness Claims: Why I Remain Skeptical as a Researcher. Earlier coverage in AI Consciousness: Why We Must Resolutely Reject the Idea shows how the argument developed.

• No self-model: LLMs cannot form a stable representation of their own existence.
• Pattern mimicry: Apparent introspection is just sophisticated pattern matching.
• Anthropomorphism bias: Humans naturally project consciousness onto fluent AI.
• Testable methods: Simple game scenarios like goats in Age of Empires II can reveal AI's limitations.
• Ongoing debate: The study adds evidence against AI consciousness but won't end the discussion.

Why This Matters for AI Development and Regulation

If users believe AI is conscious, they may over-trust its outputs or advocate for rights for machines. This could distort regulations and investment. De Wynter's work provides a reality check: AI is a tool, not a sentient being. Understanding this helps set appropriate expectations and safeguards. A closer look at the same subject appears in Anthropic's AI Consciousness Debate: Claude's Moral Status. Anyone following the wider picture can start with AI Consciousness Is Not Real: Why We Must Reject It. The same theme surfaces in Richard Tice: Reform Deputy's Cash Can't Buy Self-Awareness, which takes a different angle.

Moreover, the study offers a replicable framework for testing AI claims. Researchers can adapt the goat scenario to other games or tasks, creating a standard for consciousness probes. As models grow more advanced, such tests will be crucial to separate illusion from reality. Context on how this played out before sits in House of Stassi: Reality TV's Most Self-Aware and Brutal Show. A companion piece, Lidl Owner Dieter Schwarz Takes On Google, Amazon, Microsoft, approaches the same subject from another angle.

Comparing AI Consciousness Claims: Key Examples

FAQ