The Illusion Threshold: Reframing the Turing Test through Perceptual Intelligence
Toward a boundary condition for artificial general intelligence (AGI)
In the pursuit of artificial general intelligence (AGI), benchmarks have traditionally focused on task performance, linguistic fluency, and the replication of human outputs. The Turing Test remains iconic, but increasingly irrelevant—its emphasis on behavioral imitation can overlook the mechanisms by which cognition arises. A model that converses like a human may not reason, perceive, or adapt like one. Imitation is not understanding. Even more recent benchmarks—such as BIG-bench, MMLU, and ARC—extend the scope of evaluation but remain focused on externally visible behavior: task accuracy, factual recall, or pattern recognition.1 They assess what a model outputs, not how that output is generated. A system that responds like a human may not reason, perceive, or adapt like one.
Here I explore a potentially more structurally revealing benchmark for AGI: the Illusion Threshold. It is based on a simple but profound idea: if an artificial system can produce true optical illusions—novel perceptual traps that reliably deceive human observers—it is operating on principles aligned with human perception and, by extension, a core component of human intelligence. This capability would not merely reflect statistical power or mimicry; it would suggest the presence of an internal generative model of the world—one that simulates how observations arise from latent causes, rather than merely storing correlations between inputs and outputs that includes a model of the observer—a foundational property of general intelligence.
By true optical illusion, I refer to visual stimuli that systematically exploit inferential biases in human perception, reliably producing misinterpretations across individuals—not just confusion, but predictable perceptual error rooted in priors or context.
The inability of today’s generative systems to do this—not just to reliably reproduce known illusions, but to invent them—demonstrates a key limitation. Their outputs are impressive, but they lack the internal architecture that supports inference, prediction, and model-based manipulation of perception. The Illusion Threshold reframes AGI not as a capacity for task completion, but as a capacity to exploit the failure modes of minds—starting with our own.
This builds on arguments I made in a prior essay, The Quest for Artificial General Intelligence, where I discussed the ambiguity of defining general intelligence in humans and the epistemological challenges it poses for identifying AGI. There, I argued that the absence of a stable definition of intelligence makes AGI a moving target. I believe the Illusion Threshold can narrow this uncertainty by focusing on a quantifiable and biologically grounded benchmark—structured misperception.
Why Optical Illusions Reveal Intelligence
Optical illusions are not random visual tricks. They are structured violations of expectation that exploit the way biological perception operates under uncertainty. I refer to perception in this context not only as a sensory interface taught in medical schools but as a broader inferential process by which cognitive systems interpret ambiguous information. Perception is the substrate for much of cognition—vision, reasoning, and learning all require the capacity to infer structure from partial data. The brain is a predictive engine. It doesn’t simply record visual input; it constructs interpretations based on prior knowledge, contextual cues, and probabilistic inference. Illusions occur when this machinery is systematically misled.
To generate a novel illusion—one that successfully deceives a human—a system must do more than generate plausible visuals. It must understand how perception fails. It must possess a model of the human perceptual system and be able to identify conditions under which it will converge on the wrong interpretation.
This requires a depth of generative modeling that today’s AI lacks. It is not enough to produce outputs that resemble known illusions. True illusion synthesis involves:
Constructing a hypothesis about human perceptual inference
Modeling how sensory input will be interpreted
Introducing structured ambiguity that causes reliable misinterpretation
Creating an illusion is a test of observer modeling and counterfactual generative reasoning—the ability to simulate how another agent will interpret a scene, and to manipulate that interpretation by design. A system that can induce illusions is not merely intelligent in the behavioral sense—it is cognitively aligned with the inferential architecture of human minds.
Why Current Generative AI Falls Short
Despite their fluency and versatility, today's generative models cannot pass the Illusion Threshold. I can summarize their limitations into three core categories:
Disembodiment: Human perception is embodied. It involves continuous interaction with a physical environment—eye movements, locomotion, and sensorimotor feedback shape what we perceive. Generative models, by contrast, are disembodied samplers. They generate outputs without engaging with or updating their world models through action. At least not yet at scale.
Lack of structured spatial reasoning: Human visual perception reconstructs a coherent 3D world from 2D inputs. Illusions like the Ponzo effect or hollow-mask illusion depend on these reconstructions. Most generative systems operate in pixel or token space, lacking internal representations of geometry or causality. They may produce images that look 3D, but do not reason over 3D structures.2
No predictive coding: The human brain constantly generates predictions about incoming sensory data and adjusts its internal model based on discrepancies. This feedback loop is essential for perception—and for illusions. Although some architectures implicitly encode visual priors through denoising objectives or attention-weighted reconstruction, they do not implement prediction-error feedback loops in the cortical sense—where perception is shaped by dynamically updated expectations.
Because of these architectural differences, generative AI is currently incapable of experiencing or constructing the conditions for structured misperception. It does not perceive, it does not expect, and it does not err in the way that humans do.
The Illusion Threshold as a Cognitive Benchmark
While not universally necessary across all cognitive modalities, the ability to construct illusions reflects a form of internal model manipulation that can be thought of as a core characteristic of general intelligence. A system that can invent illusions is engaging in the kinds of representational manipulation that are necessary for true cognitive alignment, including:
Observer modeling
Internal simulation of alternative perceptual outcomes
Manipulation of latent scene structure—such as spatial layout, lighting, or occlusion relationships—to induce misinterpretation
These are not trivial capabilities. They imply a convergence between artificial and biological cognition at the level of inference mechanisms. Such a system would no longer be simply imitating intelligence—it would be participating in the same kinds of perceptual reasoning that define it.
This benchmark is also resistant to superficial optimization. Unlike performance on standard benchmarks, which can be gamed through scale, memorization, or prompt engineering, the ability to generate functional illusions requires generalization in latent model space—a genuine test of internal structure.
Toward Machines That Can Be Fooled—and Fool Us
I believe bridging the illusion gap will require more than scaling. It will demand architectural shifts such as:
Predictive processing: Models must support feedback loops where expectations shape interpretation, and prediction errors update internal beliefs.
3D generative world models: AI must reason over latent variables representing geometry, lighting, occlusion, and object permanence.
Embodied interaction: Intelligence is grounded in action. Agents must explore, manipulate, and observe consequences to form robust generative models of sensory input.
These ideas are beginning to emerge across multiple research threads. Predictive coding networks, 3D-aware diffusion models, and embodied reinforcement learning agents all reflect early steps. But no current system integrates them in a way that would allow for structured perceptual manipulation.
Our recent research at Project Data Sphere exemplifies this gap. In evaluating large language models for clinical trial interpretation and statistical reasoning, we observed consistent divergences in conclusions across independent uses—despite identical inputs. These differences were not random, but driven by the probabilistic nature of the models and their lack of stable internal representations. This is the same failure mode that prevents them from building coherent perceptual experiences. They approximate distributions, but they do not model causal or perceptual processes.
Mathematical Foundations (for readers who want the formalism)
It is important to clarify what I mean when we refer to a system "producing an illusion." Today’s generative AI models—such as diffusion models or large-scale GANs—can occasionally generate images that resemble illusions. However, these instances are incidental. They emerge from correlations in the training data, not from any understanding of how or why human perception is being deceived.
A system that truly crosses the Illusion Threshold must not only generate perceptually deceptive content, but also model the perceptual expectations and failure modes of the observer. That is, it must simulate the internal process of perception, not just emulate its outputs. In mathematical terms, this involves:
Human perception as Bayesian inference:
(Compared to LLMs, which optimize next-token prediction without latent state inference)\(% Bayesian inference P(z \mid x) \propto P(x \mid z) \cdot P(z) \)Where:
x
: sensory inputz
: latent world stateP(z)
: prior probability over possible world statesP(x | z)
: likelihood of observingx
given statez
Active inference for perception-action loops:
(Models epistemic exploration in embodied agents—not present in passive generators)\(a* = argminₐ E[F(x, z, a)] \)Where:
a
: actionF(x, z, a)
: expected free energyx
: sensory inputz
: latent state
Predictive coding and feedback error:
(Absent in feedforward-only transformers; essential in cortical visual processing)
\(E(t) = S(t) − P(t)\)Where:
E(t)
: prediction error at timet
S(t)
: actual sensory signalP(t)
: predicted signal
Rendering-based scene construction:
(Used in NeRF and 3D-aware GANs; rarely integrated with inference pipelines)
\(I = R(V, L, θ) \)Where:
I
: rendered imageR
: rendering functionV
: 3D scene structureL
: lighting parametersθ
: viewpoint or camera angle
These mathematical relationships describe, in broad strokes, the architecture required to support illusion-capable intelligence. For instance, predictive coding systems implement a feedback loop between higher-level priors and lower-level sensory input. This recursive error correction is what enables humans to misperceive in systematic, lawful ways. A generative model that can intentionally produce an illusion would need to reverse-engineer this process—it must generate an input that, when passed through a modeled perceptual system , yields a consistent misinterpretation .
No current AI system exhibits this capability. The images produced by generative systems may look like illusions, but they are not crafted with reference to a human perceptual model, nor are they iteratively refined based on feedback from predicted perception. As such, the accidental production of visual ambiguity in generative art is not a refutation of the Illusion Threshold—it is evidence for the architectural gap the benchmark is designed to reveal. They represent not just algorithms, but hypotheses about the structure of cognition.
Rethinking the AGI Question
I think the Illusion Threshold compels us to ask a different question: not “Can a machine perform like a human?” but “Can a machine misperceive like a human?”
This reframes AGI in terms of internal model structure, not behavioral surface. The ability to hallucinate, misjudge depth, or generate a perceptual contradiction that fools another mind signals not just surface-level proficiency, but structural similarity.
The Illusion Threshold thus emerges not only as a novel benchmark but potentially a core component of what it means to cross into general intelligence. It operationalizes one of the most elusive aspects of human cognition: the lawful, structured errors that arise from inference under uncertainty. These errors—far from being flaws—are integral to how humans see, reason, and adapt.
As such, the Illusion Threshold does not merely augment existing AGI evaluation strategies. It also serves as a bridge: a way of narrowing the space between human and machine intelligence, not by forcing imitation, but by identifying shared computational vulnerabilities.
Where perception falters predictably, cognition is exposed—and alignment becomes measurable.
Though here I focus on vision, the Illusion Threshold can be generalized to other perceptual domains. Auditory illusions, linguistic mispragmatics, or somatosensory misjudgments could similarly reveal structural alignment between artificial and biological agents.
AGI is not just about performance across tasks. It is about possessing the machinery to construct, revise, and sometimes distort interpretations of reality in ways that are lawful, structured, and experience-driven. The ability to invent a new illusion—to knowingly lead another perceiver astray—is a proxy for that machinery.
BIG-bench (Beyond the Imitation Game) is a diverse set of language tasks designed to test generalization and reasoning in large language models. MMLU (Massive Multitask Language Understanding) evaluates knowledge and reasoning across 57 academic subjects, from mathematics to law. ARC (AI2 Reasoning Challenge) focuses on commonsense reasoning and scientific knowledge, requiring models to answer multiple-choice science questions that often demand inference beyond surface patterns.
This is why success on benchmarks like BIG-bench or MMLU, while impressive, can be misleading. These tasks do not require models to reason over physical structure, simulate observer-dependent perspectives, or reconcile conflicting interpretations across perceptual contexts.