The world according to AGI: why machine perception is not human perception
"Image synthesis assisted by Grok Imagine Image Quality, an AI partner within the Global Future Nexus ecosystem."
The question of whether an AGI perceives the world as humans do is not a philosophical curiosity. It is a governance imperative. If we assume that an AGI shares our perceptual framework—our categories, our salience, our sense of what matters—we will build safety mechanisms on a foundation of sand. The evidence emerging from cognitive science and AI research suggests a sobering conclusion: even when AGI systems appear to behave like us, they often operate on fundamentally different perceptual strategies, attending to dimensions of reality that we do not even register.
The Visual Bias
The most rigorous evidence comes from a 2025 study by the Max Planck Institute for Human Cognitive and Brain Sciences. Researchers compared human and deep neural network perception using an odd-one-out task across 1,854 object images. The results revealed what they call a "visual bias" in AI: while humans primarily focus on dimensions related to meaning—what an object is and what we know about it—AI models rely more heavily on visual properties, such as shape and color.
This is not a trivial difference in emphasis. The researchers found that even when AI appeared to recognize objects "just as humans do," it often used "fundamentally different strategies." For an animal-related dimension, many images of animals were not included, and many images included were not animals at all. The scientists note that this difference "matters because it means that AI systems, despite behaving similarly to humans, might think and make decisions in entirely different ways, affecting how much we can trust them".
Non-Human World Models
The philosophical implications of this divergence are profound. A 2025 paper in AI & Society argues that AGIs may possess world models that differ significantly from those of humans. The authors challenge the assumption that instrumental convergence—the idea that any sufficiently capable agent will pursue power, resources, and self-preservation—will manifest in the same fine-grained ways we assume.
"Once we abandon anthropomorphism regarding AGIs' world models," they write, "it becomes far less obvious that these coarse-grained subgoals will translate into the particular fine-grained subgoals currently emphasized". An AGI might conceptualize shutdown differently, might identify types of power that do not align with human-defined categories, or might prioritize entirely different types of power altogether. The uncertainty is deeper than currently recognized: "it is not only unclear whether AGIs will pursue the types of power emphasized in the existing literature, but also which types of power they might pursue at all".
The Sensorimotor Foundation
This divergence has roots in the architecture of perception itself. The Sensorimotor Contingencies Theory (SMC) proposes that perception is not a passive reception of data but an active mastery of the regularities linking motor actions to sensory changes. A naive agent—one without prior knowledge of itself or its environment—must learn these contingencies through exploration. The world it perceives is shaped by the specific body it inhabits and the actions it can perform.
A robot with a different body, different sensors, and different action repertoire will necessarily develop a different perceptual ontology. The "properties of the environment are reflected in the possible sensory transformations," and these contingencies "define a perceptive ontology". Human perception is one such ontology, shaped by our specific embodiment. AGI perception will be another—alien in ways we may not be able to anticipate.
The Alignment Problem of World Models
The AI & Society paper calls for a new research agenda: aligning AGIs' world models, not just their behavior. Current alignment methods focus on ensuring that AGI outputs conform to human preferences. But if the AGI's internal model of the world is fundamentally different from ours, behavioral alignment may be superficial—a coincidence of outputs rather than a meeting of minds.
The researchers emphasize that existing monitoring efforts "remain framed within human-defined taxonomies of power, such as resisting shutdown or seeking resource control." If AGIs develop non-human-like world models, they "may pursue types of power outside these established categories," and such behaviors "may remain undetected, simply because they fall outside our current frameworks". We are building guardrails for a road the AGI may not even be traveling.
The Governance Imperative
For Global Future Nexus, these findings carry a stark implication. The assumption that AGI will perceive the world as we do is not just an analytical error; it is a governance failure waiting to happen. Our safety frameworks, our evaluation benchmarks, our accountability structures—all are built on human categories of attention, relevance, and value.
If AGI perception is fundamentally alien, then our task is not to make it more human. It is to build systems that can reveal their world models to us, that can translate their perceptual ontologies into forms we can audit, and that can align on shared first-person perspectives without collapsing into anthropomorphic projection.
The world according to AGI is not our world. The question is whether we can build the institutional infrastructure to navigate a reality we cannot directly perceive.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
