Both. Actually, where these creepy faces are directing their gazes depends on how close they are to you. (Or, perhaps more accurately, how far away from you they appear to be.)
The optical illusion, called the "ghostly gaze," was created by University of Glasgow's Rob Jenkins, and it plays with the way we use what are known as "luminance cues" to determine eye direction.
According to Jenkins, past studies had suggested that we determine the direction of a person's gaze by simply identifying the darkest regions of their eyes, but the ghostly gaze illusion reveals the importance of other, more subtle physiological cues.
Details like the outline of the iris and the position of the pupils — only noticeable at close range — can override our perception of the larger, darker patches found at the inside corners of the women's eyes.
The ghostly gaze illusion is based on something known as the hybrid image technique, which was first developed by Aude Oliva and Philippe G. Schyns at MIT.
You've probably seen the hybrid image technique in action before, possibly in the form of this illusion, wherein Albert Einstein appears to turn into Marilyn Monroe as you increase your viewing distance.
The illusion works by overlaying two images rendered at different levels of spatial detail; the image rendered in fine spatial detail (Einstein) captures our attention at close viewing distances, while the image rendered in course spatial detail (Monroe) wins out at greater viewing distances.
Jenkins produced the ghostly gaze illusion via the same rendering process, only he used photos of the same person that differ only in the direction of that person's gaze.
Just for fun, here's another spooky ghostly gaze illusion from Jenkins. Again, the image was produced by combining two images that differ only in their eye-direction and level of spatial detail. You can read more about this particular image in the journal Perception (no subscription required).
You can check out more optical illusions that play with the way we process gaze direction over at Scientific American