Metamaterial-based information extraction (sensing and imaging using dynamic metasurface antennas, DMAs, and reconfigurable intelligent surfaces, RISs) and information processing (metamaterial wave computers) are to date rarely combined, and the essential role of noise in information theory is mostly neglected. To combine wave-based sensing with wave-based signal processing, we interpret the sensor’s measurement process as an over-the-air analog computation performed on the information contained within the scene. The plurality of programmable meta-atoms in DMAs and RISs allows us to tailor the realized over-the-air analog operation to a specific sensing task. Thereby, the sensor’s measurement process simultaneously becomes a task-specific processing step. This allows our measurement to preferentially extract task-relevant information, such that the sensor’s latency improves remarkably. Using a discrete-dipole approximation of the programmable meta-atoms, we optimize the metamaterial configurations using deep-learning tools for a prototypical object recognition task under different noise conditions. We show that we can understand “macroscopic” properties of the optimized task-specific metamaterial configurations, for instance, how the overlap and intensity of the resulting scene illuminations evolve with the noise level. Our results are directly relevant to 6G wireless networks that require low-latency sensing capabilities under realistic, hence noisy, conditions.