The classic duck-rabbit illusion reveals that when visual evidence is ambiguous, the human brain must decide what it sees. But where exactly do human observers draw the line between ''duck'' and ''rabbit'', and do machine classifiers draw it in the same place? We use semantically ambiguous images as interpretability probes to expose how vision models represent the boundaries between concepts. We present a psychophysically-informed framework that interpolates between concepts in the CLIP embedding space to generate continuous spectra of ambiguous images, allowing us to precisely measure where and how humans and machine classifiers place their semantic boundaries. Using this framework, we show that machine classifiers are more biased towards seeing ''rabbit'', whereas humans are more aligned with the CLIP embedding used for synthesis, and the guidance scale seems to affect human sensitivity more strongly than machine classifiers. Our framework demonstrates how controlled ambiguity can serve as a diagnostic tool to bridge the gap between human psychophysical analysis, image classification, and generative image models, offering insight into human-model alignment, robustness, model interpretability, and image synthesis methods.
Figures & tables
Figure 1: Left: The classic duck–rabbit illusion [ 19 ] is a well-known example of perceptual ambiguity, where the same image can be seen as either a duck or a rabbit. Right: Our process for generating ambiguous stimuli.
Figure 2: Top: Illustrative psychometric function and effects on bias and sensitivity as a function of semantic mixing ratio ( α ). Bottom: Example stimuli as a function of semantic mixing ratio ( α ) and guidance scale (GS).
Table 1: Bias for humans and classifier models. White denotes zero bias. Red denotes negative bias (toward rabbits), and blue denotes positive bias (toward ducks). Color intensity is computed using global min–max normalization across all entries. All entries are non-positive, indicating a consistent bias toward rabbits across observers.
Table 2: Sensitivity for humans and classifier models. Darker red indicates higher sensitivity (steeper psychometric slope).
Figure 3: The 300 duck-rabbit stimuli used in the experiment, as a function of semantic mixing ratio ( α ) and guidance scale (GS).
Figure 4: Perceptual evaluation framework and 2AFC experimental design. Participants performed a categorical two-alternative forced-choice task using ambiguous duck–rabbit images (300 trials per participant).
Table 3: Deviance (goodness-of-fit) across guidance scales for human observers and machine vision models. Lower deviance values indicate better fits of the psychometric function.
Category
Model
Architecture
Training Objective
Decision Space
Supervised learning (ImageNet classification)
ResNet-50 [ 13 ]
CNN
ImageNet-1K classification
1000-way softmax
ConvNeXt-Base [ 25 ]
CNN
ImageNet-1K classification
1000-way softmax
DenseNet-121 [ 18 ]
CNN
ImageNet-1K classification
1000-way softmax
MobileNetV3-L [ 17 ]
CNN
ImageNet-1K classification
1000-way softmax
EfficientNet-B0 [ 41 ]
CNN
ImageNet-1K classification
1000-way softmax
ViT-B/16 [ 5 ]
Vision Transformer
ImageNet-1K classification
1000-way softmax
Table 4: Models used for machine psychometric evaluation. Eight classifiers are supervised models trained for ImageNet object recognition, and evaluated with a 1000-way softmax in the duck-rabbit categories.
Category
ImageNet-1K ID
Class Name
Duck (“a duck”)
97
drake
98
red-breasted merganser, Mergus serrator
Rabbit (“a rabbit”)
330
wood rabbit, cottontail, cottontail rabbit
331
hare
332
Angora, Angora rabbit
Elephant (“an elephant”)
385
African elephant ( Loxodonta africana )
Table 5: Selected ImageNet-1k classes used for the rabbit, duck, and elephant.
Figure 5: Psychometric functions across guidance scales for humans and machine vision models in the duck-rabbit condition. Error bars represent the standard error of the mean.
Figure 6: Summary estimates of bias and sensitivity for duck-rabbit condition. Left: Bias, computed as PSE−0.5 , reflects shifts in the semantic meaning boundary. The dashed horizontal line at zero indicates no bias; positive values indicate bias toward duck, values near zero indicate no bias, and negative values indicate bias toward rabbit. Right: Sensitivity, measured as the slope of the psychometric function at the PSE, indicates how sharply responses transition between semantic meanings.
Figure 7: Bias for the duck-rabbit condition across guidance scales for humans and classifier models. Grouped bars show bias values computed as PSE−0.5 for each model, with colors indicating different guidance scales (GS). All observers showed a negative bias, indicating they were more likely to see a rabbit than a duck even at a semantic mixing ratio α = 0.5.
Figure 8: Sensitivity for the duck-rabbit condition across guidance scales for humans and classifier models. Sensitivity is quantified as the slope of the fitted psychometric function at the PSE ( β1 ). Grouped bars show β1 for each model, with colors indicating guidance scales (GS). Higher sensitivity values indicate stronger semantic separation as the mixing parameter α varies. Sensitivity generally increases with guidance scale for both humans and models, although the magnitude and stability of this increase vary across architectures.
Table 6: Bias for humans and classifier models across guidance scales in the elephant-rabbit condition. Bias is defined as PSE−0.5 . Color intensity is computed by global min–max normalization across all entries, with negative values mapped to red (towards rabbit), positive values mapped to blue (towards elephant), and zero mapped to white (no bias).
Table 7: Sensitivity for humans and classifier models across guidance scales in the elephant-rabbit condition. Darker red indicates higher sensitivity.
Figure 9: The 300 elephant-rabbit stimuli used in the experiment, as a function of semantic mixing ratio ( α ) and guidance scale (GS).
Figure 10: Summary estimates of bias and sensitivity for the elephant-rabbit condition. Left: Bias, computed as PSE−0.5 , reflects shifts in the semantic meaning boundary. The dashed horizontal line at zero indicates no bias; positive values indicate bias toward duck, values near zero indicate no bias, and negative values indicate bias toward rabbit. Right: Sensitivity, measured as the slope of the psychometric function at the PSE, indicates how sharply responses transition between semantic meanings.
Figure 11: Bias for the elephant-rabbit condition across guidance scales for humans and classifier models. Grouped bars show bias values computed as PSE−0.5 for each model, with colors indicating different guidance scales (GS).
Figure 12: Sensitivity for the elephant-rabbit condition across guidance scales for humans and classifier models. Sensitivity is quantified as the slope of the fitted psychometric function at the PSE ( β1 ). Grouped bars show β1 for each model, with colors indicating guidance scales (GS). Higher sensitivity values indicate stronger semantic separation as the mixing parameter α varies.
We identify a previously unreported phenomenon in CLIP representations: human and AI-generated paintings spontaneously separate along the dominant principal directions of their joint embedding distribution, without any supervised objective designed to distinguish the two classes. Rather than exploiting this phenomenon for detection, our objective is to interpret it: we seek to identify the visual information underlying the separation and to trace it back from the embedding space to the image domain. We pursue this objective through a progressive investigation combining interpretable image representations with gradient-based inversion, used systematically as an experimental probe of the relationships identified in feature space. Robustness experiments and increasingly expressive statistical descriptors progressively rule out several intuitive explanations based on global image properties and simple local statistics, and point instead to distributed multiscale image structure. Multiscale scattering provides the most informative interpretable representation considered, but offers only a partial account of the phenomenon. Direct inversion provides a complementary and striking observation: substantial displacements along the dominant CLIP directions can be induced by image perturbations that remain nearly imperceptible to human observers, showing that the directions involved in the separation are highly sensitive to image variations with very low perceptual salience for humans. Taken together, these results reveal a significant difference between the visual evidence reflected in CLIP representations and that readily accessible to human perception, raising broader questions about the relationship between artificial and human vision and, ultimately, between artificial and human aesthetic judgment.
Andrea Asperti
Department of Informatics - Science and Engineering (DISI), University of Bologna, Via Mura Anteo Zamboni 7, Bologna, 40126, Italy.
Vision-language models (VLMs) exhibit a systematic bias when confronted with classic optical illusions: they overwhelmingly predict the illusion as "real" regardless of whether the image has been counterfactually modified. We present a tool-guided inference framework for the DataCV 2026 Challenge (Tasks I and II) that addresses this failure mode without any model training. An off-the-shelf vision-language model is given access to a small set of generic image manipulation tools: line drawing, region cropping, side-by-side comparison, and channel isolation, together with an illusion-type-routing system prompt that prescribes which tools to invoke for each perceptual question category. Critically, every tool call produces a new, immutable image resource appended to a persistent registry, so the model can reference and compose any prior annotated view throughout its reasoning chain. Rather than hard-coding illusion-specific modules, this generic-tool-plus-routing design yields strong cross-structural generalization: performance remained consistent from the validation set to a test set containing structurally unfamiliar illusion variants (e.g., Mach Bands rotated from vertical to horizontal stacking). We further report three empirical observations that we believe warrant additional investigation: (i) a strong positive-detection bias likely rooted in imbalanced illusion training data, (ii) a striking dissociation between pixel-accurate spatial reasoning and logical inference over self-generated annotations, and (iii) pronounced sensitivity to image compression artifacts that compounds false positives.
Wittgenstein's duck-rabbit poses a question for vision-language models: when a model captions an ambiguous image, where in the model is the commitment to one aspect made? We address this with a 3,320-generation behavioral baseline over 83 bistable stimuli that surfaces three regimes (default-dominant, force-dominant, force-balanced) under neutral vs forced-choice prompting, then probe the underlying representations using a TopK sparse autoencoder we train on the CLIP layer that LLaVA-1.6-7B actually consumes (validation EV 0.93). Across 69 bistable stimuli with both per-aspect feature pools available, 72% (50/69) show simultaneous activation of both pools at the vision tower, including 12/12 default-dominant duck/rabbit and 7/8 force-balanced young/old. Causal steering at CLIP layer 22 flips captions on default-dominant stimuli (33% rabbit-flip rate under a fluency guard) but cannot flip captions on force-balanced young/old at any tested coefficient, despite their vision-side superposition. The dominance bottleneck lives downstream of the vision tower; the gap between vision-side representation and language-side commitment is an empirical handle on the seeing/seeing-as distinction. We also flag a methodological note: rank-based statistics on TopK SAE outputs require tie-corrected ranking to avoid silent row-order bias.