On This Day in AI: January 2, 2024
On January 2, 2024, in Mountain View, California, Google DeepMind researchers Gamaleldin Elsayed and Michael Mozer published a detailed account of evidence that images engineered to mislead artificial neural networks can also subtly influence human vision. The work addressed “adversarial” images: ordinary photographs altered by carefully calculated pixel changes so small that a person may see only harmless noise, while an image-classification system makes a confident error. DeepMind’s explanation brought a counterintuitive result to a wider audience. Humans did not suddenly mistake a vase for a cat, as a machine might, but their forced choices could be nudged in the same direction as the model’s mistaken classification.
The research grew from a long-running AI-security problem. Modern vision systems can identify objects with remarkable accuracy, yet minute changes to an input can expose brittle behavior. In one illustrative case, a network correctly recognized a vase until a patterned perturbation pushed it toward the label “cat.” The research team created pairs of altered natural images and asked participants questions such as which picture appeared more cat-like. To the volunteers, the paired pictures often looked virtually identical. Across controlled experiments, however, selections favored the image modified toward the named target more often than chance would predict, and the measured bias grew as the perturbations became stronger.
The result did not erase the large difference between human and machine perception. People continued to recognize the original subject, while the neural network could abandon the correct category entirely. The paper also tested possible artifacts and compared perturbations produced by different model architectures. Its authors concluded that the shared sensitivity was more likely tied to higher-order statistics in natural images than to one particular neural-network design. That qualification mattered: the experiments showed a small, systematic influence on judgment, not proof that people and computers see identically or that a hidden image can wholly control a viewer’s perception.
The January 2 explanation gave the finding significance beyond a laboratory curiosity. Adversarial examples are central to efforts to make computer vision safer in areas such as transportation, identity systems and medical imaging. If some machine-targeted signals also affect people, researchers cannot assume that human observers provide a perfectly independent reference. At the same time, comparing where biological and artificial vision agree may help engineers build models whose representations are more aligned with human perception and less vulnerable to deceptive inputs. The event marked a useful shift in the AI-safety conversation: robustness is not only a contest between attackers and algorithms, but also a question about how technology interacts with the people asked to judge its outputs.
An adversarial attack calculates small changes to image pixels that push a neural network toward a wrong label while leaving the underlying scene recognizable to a person.
The DeepMind examples included natural photographs paired with different perturbations, allowing researchers to compare human choices with the direction of the model’s error.
Participants usually felt that paired pictures were nearly identical, yet their forced choices were biased toward the target category encoded by the adversarial change.
The effect was subtle and did not make people adopt the network’s full misclassification; human observers still perceived the original object clearly.
The findings suggest that people and artificial vision systems can share sensitivity to some statistical cues in natural images despite major differences in their processing.
Studying those overlaps may improve robustness tests and help researchers evaluate whether computer-vision models are aligned with reliable human perception.
Explore more of "On This Day ..."
Discover more events from the same date across news, politics, technology, sports, and other fields. Each link highlights significant moments that shaped history on different fronts.