Contributing to the debate whether level 1 visual perspective taking (L1-VPT) is an implicit socially rooted process (mentalizing) or rather a non-social process (submentalizing), a laboratory study by Furlanetto et al. (2016) showed evidence in favor of the mentalizing approach: Using the Dot Perspective Task (Samson et al., 2010), the authors found altercentric intrusions, when a top-down belief cover story asserted that an avatar was able to see, but not when the avatar’s vision was presumably suspended. Altercentric intrusions have been considered indicative of spontaneous perspective taking, implying that observers cannot fully inhibit another agent’s conflicting perspective. We aimed to test and transfer this finding to non-human avatars in two online experiments, each with a 2 (consistency vs. inconsistency of observer’s and avatar’s perspective) × 2 (self vs. other perspective) × 3 belief induction (seeing avatar without goggles; seeing avatar with transparent goggles; non-seeing avatar with opaque goggles). Experiment 1 with a human avatar (n = 86) and Experiment 2 with a robot avatar (n = 86) both replicated the hypothesized altercentric intrusion effects for error rates, but not for response times. Egocentric intrusions, i.e., intrusions of the self-perspective when asked to take the avatar’s perspective, however, were moderated by the belief induction, appearing only under the conditions with a seeing (human or robot) avatar. Addressing the generalizability of altercentric intrusion effects over human and robot avatars, we discuss whether this effect might represent a more explicit form of perspective taking based on submentalizing.