We have investigated first-order artifacts in second-order motion perception. Subjects were required to identify the orientation and direction of a drifting sinusoidal contrast modulation. When the carrier consisted of static two-dimensional noise, performance often reflected the use of first-order artifacts that arise from stochastic local biases in the noise, rather than the detection of the contrast modulation per se. This stimulus, which has been used widely for studying second-order motion, therefore appears to be inappropriate for that purpose. In contrast, global distortion products arising from luminance non-linearities do not appear to provide usable artifacts. Two manipulations were employed to eliminate local first-order artifacts: the use of dynamic noise and the use of high-pass filtered static noise. These two manipulations gave similar results, which were quite different from those obtained with broadband static noise. We argue that performance with both of these image types reflects the activity of a true second-order motion mechanism. A characteristic property of this mechanism is that it cannot specify direction at the threshold for detecting orientation. Direction thresholds are around 50% higher than orientation thresholds when first-order artifacts are eliminated.