A “3,” “5” or “7” can be sent toward this device, but the camera always receives an “8.” This is the principle behind the “lying mirror” designed at the University of California, Los Angeles. It does not first recognize the digit and then create a false image: its surface directly transforms the light carrying the image.
In the experiment, the researchers projected optical patterns representing different handwritten digits. The light carrying these patterns then reached a surface made up of numerous micro-mirrors. After reflection, a camera placed in the intended area recorded the resulting image. Whatever digit was tested, the device had been adjusted to make the same fake digit appear: an “8.”
The principle of the “lying mirror,” which transforms different input images into fake patterns. The device remains effective despite rotations, changes in scale, shifts, noise or unseen images.
To understand what is happening, we need to forget the idea of a screen hidden behind the mirror. Each small area of the surface slightly alters the path of the light wave. The different parts of the light then reinforce one another in certain places and weaken one another elsewhere. This redistribution, through diffraction, makes it possible to concentrate brightness according to the desired pattern.
The calculation takes place before this transformation. A deep-learning program determines how each area of the mirror should act to obtain the target image from many different inputs. In the visible demonstration, the surface contained 150 × 150 adjustable areas. Once this adjustment has been established, each new image is transformed through the propagation of light, without having to be recalculated numerically.
In the simulations, one version made a bag appear from a variety of clothing items. Another produced the digit “8” from handwritten digits, and a third produced a panda from drawings.
The tests also show that the process tolerates certain variations in the input images. The researchers shifted, enlarged, reduced or rotated the patterns, and added noise. They also tried images that were absent from the training set. Quality declined in some cases, particularly with very unstructured patterns, but the fake image often remained recognizable.
The experimental validation used three visible wavelengths: 480, 550 and 600 nm, close to blue, green and red. They worked separately and together. In the version published in August 2026, the team also says it designed a mirror covering a continuous range of wavelengths. The goal is to accommodate different lighting conditions.
The device remains a controlled optical experiment, very different from a mirror installed in a room. The authors nevertheless envision uses for masking visual information, creating entertainment effects or complicating certain observations. An important next step is now to adapt these surfaces to less controlled lighting, particularly natural light.