Castelli et al., 2001 Castelli L.
Vanzetto K.
Sherman S.J.
Arcuri L. The explicit and implicit perception of in-group members who use stereotypes: blatant rejection but subtle conformity.
Anobile et al., 2014 Anobile G.
Cicchini G.M.
Burr D.C. Separate mechanisms for perception of numerosity and density.
Castelli et al., 2001 Castelli L.
Vanzetto K.
Sherman S.J.
Arcuri L. The explicit and implicit perception of in-group members who use stereotypes: blatant rejection but subtle conformity.
Participants took part in a modified version of the number of letter estimation task (). The procedure of Study 2 is schematically summarized in Figure 4 . Participants had to estimate the number of letters presented on the screen at different densities. Importantly, at each trial and before each density display, they saw one of the faces used in Study 1 (either a GAN or a REAL face) for 3 s; furthermore, after the face disappeared and on top of the letter display, they were also provided with the estimate given by that face for the same letter display. Participants decided if they wanted to use such an estimate as an anchor to provide their own estimate. They were informed that “real faces tried very hard to provide a correct estimate”, and therefore the number could be informative, but the number provided by GAN faces was randomly generated by an algorithm, and therefore was not necessarily informative. Therefore, we implicitly asked participants to judge whether they believed a face was real or fake. To incentivize participants to use all available information, we rewarded their accuracy on a random trial with a bonus. The letter displays were generated using Matlab 2019b. The monitor was a Dell UP2516D (55.29 × 31.10 cm; 2560 × 1440 resolution; 0.216 × 0.216 mm pixel pitch). Each display consisted of 200 letters As randomly distributed within an imaginary circle. We varied the radius to generate 500 displays with 5 different densities to manipulate the numerosity perception (). More specifically, 5 radii varied from 39 to 78 pixels (with equal step). For each radius we generated 100 random displays by sampling the letter positions as Cartesian coordinate pairs from a normal distribution with mean zero and standard deviation equal to the radius. To account for extreme locations, we repeated the sampling until all letter positions were within 3 standard deviations from the mean of all position radii. The density of each display was computed as the number of letter within a unit area divided by the unit area. Therefore, for each radius a letter could have appeared at any angle of the circle and at any distance between the centre of the circle and the radius. In total, we generated 500 displays (100 displays per radius) from where a display was randomly sampled at each trial (N = 100 trials). The estimate on top of the letter display, that was supposedly given by the face displayed just before, was actually a random number generated from a normal distribution with a mean of 200 (i.e., the number of actual dots in the display) and a standard deviation of 25. These values were chosen to match the value used in. The sampled numbers were then rounded to the nearest integer because the number of dots is an integer. The task involved seeing 100 faces (50 GAN and 50 REAL) and 4 attention check trials which were already included in Study 1, each for 3 s and had 10 s to provide their estimate of the number of letters on the screen. It was divided into two blocks, with a 3-min break in between. In Study 3, participants were divided into the Knowledge and the NoKnowledge groups. Both groups started with the Conformity task followed by the Realness task, but differed on the type of information that they received at the beginning of the Conformity task. The Knowledge group was first given the same definitions as in Study 1 followed by a paragraph to provide a social context: “Such technology of generating artificial faces of non-existing people can be used in various useful contexts (e.g. improve the quality of old photos, generate new model images for commercial websites, etc.), but is also being used with malicious intentions, such as generate fake social media profiles that could influence social and political behavior. It is therefore important and timely to investigate how we process such faces”. This was added to reinforce the importance of the distinction in Realness and to effectively frame the Conformity task. Participants were told that some of the faces had put a lot of effort to provide a good estimate of the number of letters on the screen, others just responded randomly and that they had to “Use any information available to you to consider whether a face has given an informative response that you may or may not want to take on board”. The sentence intended to heighten the participants’ awareness and effort in gathering information from non-verbal cues of the faces, but was purposefully vague about what “cues” meant. This intended to increase ecological validity, where people might be aware of the existence of fake faces (e.g. on social media), but are not told that the faces they are looking at might not be real. The participants then proceeded to the Realness task. The NoKnowledge group did not receive any indication about the nature of the presented faces in the initial Conformity task. The only piece of information provided was that some of the faces on the images tried to make good estimates, whereas others just responded randomly and that they should use all the information available to them, when estimating the number of letters on the screen. After the Conformity task, we again introduced the definitions of Real and Fake faces with social context and proceeded with the Realness task (see section 2.9 of the Data S1 file for further details).