Spontaneous Trait Inferences on Social Media.

The present research investigates whether spontaneous trait inferences occur under conditions characteristic of social media and networking sites: nonextreme, ostensibly self-generated content, simultaneous presentation of multiple cues, and self-paced browsing. We used an established measure of trait inferences (false recognition paradigm) and a direct assessment of impressions. Without being asked to do so, participants spontaneously formed impressions of people whose status updates they saw. Our results suggest that trait inferences occurred from nonextreme self-generated content, which is commonly found in social media updates (Experiment 1) and when nine status updates from different people were presented in parallel (Experiment 2). Although inferences did occur during free browsing, the results suggest that participants did not necessarily associate the traits with the corresponding status update authors (Experiment 3). Overall, the findings suggest that spontaneous trait inferences occur on social media. We discuss implications for online communication and research on spontaneous trait inferences.

Social media allow people to communicate at virtually no cost and effort and build large online networks, which can be powerful sources of social and emotional support (Donath, 2007). The challenge lies in finding successful ways to maintain and navigate those networks, since their size and diversity render traditional maintenance strategies, such as face-to-face communication, less feasible (Resnick, 2001; Tong & Walther, 2011). Snap social processes like spontaneous inferences present a potential solution. Research has shown that people spontaneously infer traits, goals, and values from minimal exposure to information (Uleman, Saribay, & Gonzalez, 2008). Social media offer a near constant stream of information, and if inferences indeed require no more than a passing glance, browsing might help users gain awareness of their online networks (Levordashka & Utz, 2016; Thompson, 2008). However, the extent to which such snap inferences are made online is not clear. We examine key conditions that could hinder spontaneous inferences on social media and report experiments designed to test whether inferences occur under these conditions.

In research on online impression formation, it is common to explicitly ask participants to judge others, whose profiles they view at their own pace (e.g., Antheunis & Schouten, 2011; Back et al., 2010; Evans, Gosling, & Carroll, 2008; Gosling, Augustine, Vazire, Holtzman, & Gaddis, 2011; Pennington & Hall, 2014; Westerman, Van Der Heide, Klein, & Walther, 2008). In contrast to this deliberate evaluation, browsing involves skimming through information without any particular intention. Even if the encountered posts contain relevant cues, we do not know whether these cues will lead to inferences when people encounter them briefly and without explicit impression formation goals.

There is robust evidence that people spontaneously infer social information and form impressions (Ambady & Rosenthal, 1992; Uleman et al., 2008), which can persist over time and affect behavior (Todorov, Mandisodza, Goren, & Hall, 2005; Todorov & Uleman, 2004). Here, we focus on behavioral descriptions as a cue (Carlston, Skowronski, & Sparks, 1995; Uleman et al., 2008). Studies have shown that when reading about others’ behavior, people make inferences even when their task is unrelated to impression formation or when they are under high cognitive load (Todorov & Uleman, 2003). Status updates often contain trait-implying information and can therefore be expected to produce similar effects. However, the evidence for spontaneous trait inferences comes from laboratory experiments with conditions that differ from social media in important ways.

Spontaneous trait inference experiments typically use third-party descriptions (Rim, Uleman, & Trope, 2009; Saribay, Rim, & Uleman, 2012; Todd, Molden, Ham, & Vonk, 2011), which are particularly powerful in driving impressions. Even when first-person descriptions are used (e.g., Carlston et al., 1995), the information is provided to the participants with no context, and there is little reason to doubt its accuracy. Social media updates, however, are self-generated and shared voluntarily, which makes them less reliable due to strategic self-presentation (warranting principle; Utz, 2010; Walther, Van Der Heide, Kim, Westerman, & Tong, 2008).

Moreover, the information in laboratory experiments can be of rather extreme nature. Behaviors such as “She threw a chair at her classmate” (Bliss-Moreau, Barrett, & Wright, 2008) or “I kicked [a puppy] out of my way” (Carlston & Skowronski, 1994; Carlston et al., 1995; McCarthy & Skowronski, 2011) are used as stimuli, but would rarely be shared online, where information tends to be more mild and appropriate (Utz, 2014). This is problematic because extreme information is known to influence impressions more strongly (Fiske, 1980; Skowronski & Carlston, 1989). Indeed, researchers have speculated that too mild content might not result in spontaneous inferences (Skowronski, Carlston, & Hartnett, 2008).

Another major difference lies in how information is presented. Nearly, all experiments on spontaneous trait inferences present only one pair of actor and behavior at a time. Forming a distinct association under such conditions is more straightforward than on social media, where different actors and behaviors appear in parallel and users browse through without being particularly attentive. A social media user could easily be looking at one person, while still processing information about another, especially while browsing.

Since updates on social media are self-generated and often mild in content, it is not clear whether they will be sufficiently powerful as a cue to produce spontaneous inferences. Furthermore, viewing multiple updates in parallel and in “browse” mode might hinder the associative process. In a series of experiments, we tested whether people will spontaneously infer traits from mild, self-generated social media updates (Experiment 1), when information is presented in parallel (Experiment 2) and merely browsed through (Experiment 3).

General Method

Overview

We adapted an established trait-inference paradigm (false recognition; Todorov & Uleman, 2002). The paradigm assesses trait inferences via a recognition task. First, participants see a number of actors paired with brief trait-implying descriptions. They are asked to read the descriptions without any mention of impression formation (learning phase). In a subsequent recognition task, the same actors are paired with single words, and for each pair, participants indicate whether the word appeared in the actor’s description. It has been shown that if the target word is the trait implied by the actor’s description, participants make more mistakes saying that it was in the description (Todorov & Uleman, 2002). This false recognition of implied traits occurs because, while reading the descriptions, participants spontaneously infer the implied traits and associate them with the corresponding authors. When developing the paradigm, researchers adopted additional control conditions and counterbalancing to rule out alternative explanations, such as mere word activation, and effects of extraneous characteristics of the stimuli. It was consistently demonstrated that the counterbalancing had no effect, which led the authors and other researchers employing the paradigm to drop the counterbalancing.

We followed the same procedure but changed the content and presentation of stimuli. The stimuli were social media updates, ostensibly posted by the actor (self-generated) and nonextreme (appropriate) in content. Experiment 1 followed the original paradigm but with status updates as stimuli. In Experiment 2, we used the same content and varied the number of updates presented simultaneously. In Experiment 3, participants browsed through all updates at their own pace. Our primary dependent variable was the number of mistakes (false recognitions). Whether trait inferences will be made in these conditions is an open question. What we hypothesize and test in each experiment is that if trait inferences occur, participants will make more mistakes for implied traits as compared to other traits.

Prior research has assessed response times (RTs) to correct trials. Since longer RTs can be indicative of greater difficulty, it can be expected that if spontaneous inferences are made, RTs to implied traits should be longer. Although there has been some supporting evidence, RTs are not a reliable indicator of trait inferences. Nevertheless, to be consistent with prior work, we recorded and reported RTs. In addition to false recognitions, we developed an alternative assessment of impressions.

We report how we determined sample size, all data exclusions, manipulations, and measures (Simmons, Nelson, & Simonsohn, 2012). The design and hypotheses were preregistered (osf.io/jqhdz). The experiments were designed in PsychoPy (Peirce, 2007) and analyzed in R (R Core Team, 2015). The research was approved by an ethics committee.

We conducted pilot research for Experiments 1 and 2, which is not reported here, but a report is available online (Levordashka, 2016). The results are consistent with the remaining experiments and including them in the article would not have altered our conclusions.

Sample

The effect sizes in spontaneous trait inferences experiments using the false recognition paradigm are moderately large (Todorov & Uleman, 2002). A sample of 16 participants would have been sufficient to achieve statistical power of 0.95 in a two-tailed dependent-samples t-test. We have decided on larger samples to ensure power of at least 0.90 for effect sizes of dz = 0.60.

For all experiments, participants were recruited from the participant pool of a German research institute. Some experience with social media was called for during recruitment but not subsequently assessed. Since prior research has not found gender differences in spontaneous impression formation, we did not consider it necessary to discriminate participants based on their gender to ensure a balanced sample. Gender distribution is reported per experiment. Other demographic information was collected separately and reported for the sample across experiments. The participants were undergraduate students of various faculties (no discipline was represented by more than 10%). Age ranged from 19 to 34 years (M = 24, SD = 3).

Method

Participants

Thirty participants took part in the study (22 female). After providing an informed consent, participants received instructions and completed the study on individual computer screens. The total duration was 15 min. At the end, participants were debriefed and reimbursed (2 EUR).

Procedure

Participants saw 36 social media updates (Figure 1), which they were asked to read carefully (learning phase). On each trial, a single update was presented for 5 s. In the recognition task, participants saw each face from the learning phase, paired with a word, and had to indicate as quickly as possible whether the word had appeared in the status update of the person (old word) or whether it is a new word that was not in the update. For example, if person A’s description was “Just spilled coffee all over my laptop!” and in the test phase person A appeared with the word laptop, the correct response would be old; any word that was not in the update (e.g., clumsy) would be a new word. The presentation order of single trials was randomized in each task and for each participant.

Figure 1.

Example of status update stimulus. The actual stimulus had a different face and name, and the text was in German.

For the updates, we used 36 sentences (in German), 12 of which mentioned a personality trait and 24 implied a trait without explicitly mentioning it. The associations between sentences and implied traits were pretested. The number of positive and negative sentences was balanced.

The sentences were randomly paired with 36 faces from Bainbridge, Isola, and Oliva (2013). We used equal number of male and female faces of similar attractiveness and memorability (based on the ratings from the database).

For the recognition task, each face was paired with a word. Everyone faces whose status update mentioned a trait were paired with this same trait (presented trait). The remaining 24 sentences were split into three groups: eight faces were paired with the trait implied by the status update of the same actor (implied-trait condition); eight faces with the trait that was implied by a status update of another actor (other-trait condition); eight faces with a novel trait (control-trait condition). Presented-trait trials served as fillers and were not analyzed. The correct response to all other trials was new. Responses old were coded as errors (false recognitions).

Memory of Stimuli

We assessed participants’ memory for the sentences form the learning phase in a recall task. Participants saw a random selection of six faces and had to type the corresponding updates. The responses were coded by research assistants unaware of the study design. The scale for accuracy was: 0 = no response, 1 = not at all accurate, 2 = accurate meaning; mistakes in wording, 3 = mostly accurate, 4 = accurate. Another item assessed whether participants recalled a sentence but paired it with the wrong target. If a participant’s response was a somewhat accurate recollection of one of the stimulus sentences but not that of the target face, the response was coded as mistaken target. Intercoder reliability was estimated on 20% of the responses, with ratings from two coders (intraclass correlation coefficient [ICC] = .93, 95% confidence interval [CI] [0.83, 0.97]).

Direct Assessment of Impressions

Participants were asked to evaluate some of the actors (n = 12 per participant) from the learning phase. On each trial, they saw a face, paired with two traits: The trait implied by the status update and another trait of the same valence and had to choose one of the traits to evaluate the person. This assessment of impressions was an additional measure of trait inferences.

Excluded Cases

Our a priori criterion was 3 SD above or below the sample mean for each measure in the study. There were no outliers. One participant made no correct responses in the implied-trait condition. Since the RT analyses were based on correct trials, this resulted in missing data. To ensure a balanced design, we excluded this participant from the RT analysis. Alternative approaches (replacing the missing data with the participant’s average RTs in the other conditions or with the average RTs of other participants in the missing condition) yielded similar results.

Results

We used paired-sample t-tests to test our hypothesis regarding false recognition rates (Table 1). Participants made more mistakes when an actor’s face was paired with the trait implied for this actor (implied-trait condition; M = 0.55, SD = 0.24) as compared to a trait implied for another actor (other-trait condition; M = 0.34, SD = 0.22) or a novel trait (control condition; M = 0.23, SD = 0.17). This pattern suggests that people spontaneously inferred traits when reading status updates with mild content, written from a first-person perspective.

Table 1.

Results of Dependent-Samples t-Tests Comparing False Recognition Rates across Conditions (Experiment 1).

Comparison	df	t	Sig. (Two Tailed)	Hedges’ g [95% CI]
Implied–other	28	4.15	<.001	0.89 [0.41, 1.4]
Implied–control	28	6.21	<.001	1.47 [0.89, 2.13]
Other–control	28	2.93	.007	0.52 [0.15, 0.91]

Note. CI = confidence interval.

Average RTs to correct trials ranged between 1.03 and 3.24 s (M = 2.15, Mdn = 2.22, SD = 0.54). The data violated assumptions of normality. We performed a log10 transformation on the row data (RTs per trial). To avoid negative values, we added 1.0 to all values prior to the transformation. Following the transformation, the data no longer violated assumptions of normality and equal variances (Shapiro Wilk’s ps > .06; Levene’s test ps > .1). There were no significant differences in RTs (all ps > .087).

The recall of the stimulus material was low. For 40% of the trials, participants were not able to recall anything and 42% of the recalled sentences were classified as “not at all accurate.” As we expected, participants occasionally recalled the status update of a person other than the one whose picture was displayed. This however occurred in only 16% of the cases, which indicates that overall retention of the stimuli was indeed low.

The occurrence of spontaneous inferences was apparent in the direct evaluations measure. When asked to evaluate the actors from the learning task, participants selected the implied trait over another trait of the same valence 62% of the time. An exact binomial sign test indicated that this was significantly higher than chance, 95% CI [0.55, 0.69], p = .001. Given the low recall, this preference is likely due to implicit evaluation based on the actor’s update rather than direct recall of the update.

Discussion

The results of this experiment show that the content typically encountered on social media can result in spontaneous trait inferences. These inferences were reflected in the outcome of a word recognition task: Participants were more likely to falsely recognize a word as having been previously mentioned in an actor’s status update, when the word was the trait implied by the update, as compared to traits implied for other actors and to novel traits.

We also found that when directly asked to evaluate a person, participants were more likely to choose the trait, implied by this person’s update. Importantly, this likelihood was higher than what the memory of stimulus materials would suggest. That is, participants’ evaluations were driven by status updates they had previously seen but could likely no longer recall.

In sum, we provide evidence for trait inferences on the basis of ostensibly self-generated status updates. However, the possibility remains that if too much information is presented at once, there will not be clear associations and person-specific inferences. The following experiment was designed to address this concern.

Nineteen participants took part in the experiment (16 female). The recruitment, procedure, stimuli, and measures were identical to that in Experiment 1, with the exception of how we presented the stimuli. Instead of one at a time, participants saw nine updates in each of the four trials (60 s per trial). In the test phase, the implied-trait condition remained unchanged. There were two comparison conditions: Faces paired with traits implied by other updates from the same trial and faces paired with traits from other trials.

There were no outliers based on the three SD criterion. One participant had to be excluded from the analysis of RTs because they made no correct responses in the implied-trait condition, which resulted in missing data.

The number of false recognitions differed significantly across conditions (Table 2). Participants made considerably more mistakes when an actor’s face was paired with the trait implied for this actor (implied-trait condition; M = 0.57, SD = 0.23), as compared to a trait implied for another actor in the same trial (M = 0.3, SD = 0.15) or another trial (M = 0.29, SD = 0.18). There was no significant difference between the two control conditions. Crucial here is the difference in false recognitions between traits implied by status updates that were presented during the same trial. Although the updates appeared simultaneously, the trait implied by each update was specifically associated with the person who posted the update.

Table 2.

Results of Dependent-Samples t-Tests Comparing False Recognition Rates Across Conditions (Experiment 2).

Comparison	df	t	Sig. (Two Tailed)	Hedges’ g [95% CI]
Implied–other (st)	18	6.17	<.001	1.32 [0.75, 2]
Implied–other (ot)	18	5.14	<.001	1.3 [0.67, 2.03]
Other (st)–other (ot)	18	0.35	.734	0.08 [−0.38, 0.54]

Note. st = same trial; ot = other trial; CI = confidence interval.

RTs to correct trials ranged between 1.1 and 5.59 s (M = 2.46, Mdn = 2.41, SD = 0.99). Due to violated assumptions of normality, we performed a log10 transformation on RTs per trial (1.0 was added to all values). Following log transformation, the data no longer violated the assumptions of normality and equal variances (Shapiro Wilk’s ps > .09; Levene’s test ps > .79). RTs for correct responses for implied-trait trials (M = 3.11, SD = 0.92) were shorter than in the other two conditions (same trial: M = 3.49, SD = 0.97; other trial: M = 3.46, SD = 1.01; Table 3).

Table 3.

Results of Dependent-Samples t-Tests Comparing RTs across Conditions (Experiment 2).

Comparison	df	t	Sig. (Two Tailed)	Hedges’ g [95% CI]
Implied–other (st)	17	−4.91	<.001	−.42 [−.67, −.21]
Implied–other (ot)	17	−2.92	.010	−.39 [−.72, −.1]
Other(st)–other (ot)	17	0.25	.805	.03 [−.22, .28]

Note. st = same trial; ot = other trial; CI = confidence interval.

The recall of status updates was low. On the majority of trials, participants provided no response (47%) or a highly inaccurate response (28%). There were some cases of recalled update but a mistaken target (18%), which shows that participants recalled more than their raw memory scores suggest. Even when considering these cases, the overall recall was low.

Despite having poor recall of the seen updates, when asked to evaluate an actor, participants chose the implied trait over another trait of the same valence 75% of the time, which was significantly higher than chance, 95% CI [0.66, 0.82], p < .0001. These results provide evidence that participants made actor-specific trait inferences.

Experiment 2 provided evidence that the simultaneous presentation of status updates can result in distinct, actor-specific inferences. Participants were more likely to falsely recognize an implied trait as having been previously presented, which indicated that they spontaneously inferred the trait. They were also more likely to associate an actor with the trait implied by this actor’s update, without necessarily being able to recall the update. We did not find the expected differences in RTs.

Unique to this experiment is the demonstration that when participants viewed updates simultaneously, the traits they inferred from each update were distinctly associated with the author of the update and not with the other actors whose picture and update they viewed at the same time. However, the experimental setup does not rule out the possibility that browsing poses a boundary condition to this association. We therefore conducted a third experiment, in which participants browsed through the updates at their own pace in a setup that closely resembled social media.

Forty-five participants took part in the study (37 female). Recruitment was identical to that in previous experiments and we used a similar procedure. The stimulus material was the same as in the previous experiments, but instead of presenting the status updates in discrete trials for fixed amount of time, all updates appeared on a time line and participants could scroll through for as long as they like. They were instructed to mark the updates they found interesting and proceed when ready (browsing time was recorded). In the test phase, the experimental conditions were the same as in Experiment 1: implied trait, other trait, and a novel trait.

Scenario-Based Evaluation Measure

We developed a novel scenario-based measure of direct impressions. On each of four trials, participants read a situation and were asked to make a choice between two people. The situations were such that a person with certain trait would be preferred (e.g., “Who would you rather give your house key to?” would likely be somebody who is trustworthy rather than unreliable). For each question, participants had a choice between two actors: one for which the desired trait was implied and one for which an opposite trait was implied. We compared the likelihood of choosing a person with a desired trait over a person with an undesired trait. To control for the influence of faces, for every pair of actors in a scenario, we swapped the traits implied during the learning phase. This manipulation was done between participants. That is, already in the learning phase, for half of the participants, person A was paired with the desired trait and, for the other half, person A was paired with the undesired trait.

There were no outliers in the false recognitions measure and one RT outlier, who was excluded from the RT analysis. We excluded one more participant from the RT analysis due to missing data: They had no correct responses in the implied-trait condition.

On average, participants spent 2.42 (SD = 1.19) min browsing the updates, which is an average of 4 s per update. They marked 20% (SD = 15) of the updates as interesting. Browsing time was positively correlated with memory of the stimulus updates (r = .32, p = .032) but not with error rates (r = .18, p = .243). The number of interesting updates was positively correlated with browsing time (r = .51, p < .001). All zero-order correlations are reported in the supplemental material.

The pattern of means was consistent with our previous findings: higher number of error rates in the implied-trait condition (M = 0.43, SD = 0.25) as compared to other-trait (M = 0.37, SD = 0.21) and control (M = 0.24, SD = 0.19). The difference between implied and other-traits was small and only approaching significance (Table 4). Given that the means are in line with our hypothesis and a one-sided test (justified by our preregistered directional prediction) would have been significant, we consider that spontaneous inferences did occur. There was also a significant difference between the other- and control-trait condition, which we did not find in the previous experiments. If we are to interpret this pattern according to Todorov and Uleman (2002), we would conclude that inferences occurred (the least mistakes were made for novel traits) but were not successfully bound to corresponding faces (no difference between correctly paired trait-face trials, i.e., implied trait condition, and mismatched pairs, that is other-trait condition).

Table 4.

Results of Dependent-Samples t-Tests Comparing False Recognition Rates across Conditions (Experiment 3).

Difference	df	t	Sig. (Two Tailed)	Hedges’ g [95% CI]
Same–other	45	1.94	.058	0.28 [−0.01, 0.57]
Same–control	45	5.33	<.001	0.86 [0.5, 1.24]
Other–control	45	4	<.001	0.62 [0.29, 0.96]

RTs to correct trials ranged between 0.84 and 3.78 s (M = 2.04, Mdn = 1.96, SD = 0.69). Following a log10 transformation (per trial; 1.0 added to each value), the data no longer violated the assumptions of normality and equal variances (Shapiro Wilk’s ps > .02; Levene’s test ps > .11). Only the difference between implied and novel traits was significant, t(43) = 2, p = .052 (all other ps > .146).

The recall of status updates was low. Responses were mostly absent (45%) or highly inaccurate (41%). Occasionally, participants recalled the sentence of a different actor from the data set (17%).

The results of our scenario-based assessment offer support for the occurrence of trait inferences. Across four different scenarios, participants selected the actor whose status update implied a trait that would be desirable in the particular scenario 61% of the times, which was significantly higher than chance, 95% CI [0.54, 0.68], p = .002.

Evidence for spontaneous inference during browsing was found in the direct evaluation measure. In the false recognition measure, the difference between the implied- and other-trait conditions was not significant. According to the traditional interpretation of the paradigm, this pattern suggests that inferences occurred (the least mistakes were made for novel traits) but were not successfully bound to corresponding faces, as indicated by the small difference between correctly paired trait-face trials (implied-trait condition) and mismatched pairs (other-trait condition). However, it is not clear whether this pattern is robust. The pattern of means corresponds to what was found in the first two experiments and we found evidence for inferences in the alternative measure, which renders the possibility that the effect was present but weaker.

General Discussion

The information social media users encounter online is rich in social cues. Psychological research on first impressions suggests that even without attending to the social aspects of information, users might form impressions based on that information. Despite this possibility and its implications, spontaneous inferences have not been studied in the context of social media. We provide evidence that trait inferences occur spontaneously from content and under conditions resembling social media (Figure 2).

Figure 2.

Visual summary of results from the false recognition paradigm. Higher number of false recognition rates in the implied-trait condition (relative to the other two conditions) indicate occurrence of spontaneous inferences. Center lines show the medians; box limits indicate first and third quartiles; whiskers extend to the highest and lowest value within 1.5 times the interquartile range; outliers are represented by dots. The data are plotted using ggplot (Wickham, 2009).

Participants spontaneously formed impressions on the basis of single, nonextreme, ostensibly self-generated status updates without being instructed to do so. Our research supports the robustness of spontaneous trait inferences and their relevance for online communication.

Prior research has shown that person inferences are unintentional, cognitively efficient, long-lasting, and can be of traits but also values, goals, or intentions (Uleman et al., 2008). This makes them highly relevant for online communication, where information is often merely glanced at. Our work is an important first step toward examining the impact of spontaneous impression formation on relational and informational processes online.

We used an established measure of trait inference. Using a previously validated measure strengthens the conclusions we can draw from the research. This particular measure requires the use of cues that do not explicitly contain the target impression, which limits the possibility of conducting research with nonconstructed stimuli (e.g., posts directly taken from social media). We therefore included alternative assessments, which involved choosing traits to describe an actor and choosing between two actors for a scenario that calls for the trait implied for one of the actors. The outcomes of these alternative measures were consistent with the false-recognitions measure, which offers support for their effectiveness.

Framing a spontaneous-inferences paradigm in a social media setting is another innovation, which can be helpful for future research. Social media provides a realistic setting for research on spontaneous inferences, with possibility to manipulate social context and integrate additional tasks in a smooth way that is intuitive and familiar for the participants.

The experiment in which we investigated impressions during browsing provided evidence for actor-specific trait inferences on scenario-based measure of inferences but had inconclusive results on the recognition measure. Participants were more likely to falsely recognize previously implied traits (as compared to novel traits), regardless of whether they were paired with the person who posted them or another person from the stimulus set. This pattern suggests that people inferred traits from status updates but did not necessarily associate these traits with their corresponding authors. One explanation of this would be that while browsing, people paid more attention to the content itself rather than who posted it. This would be an interesting result, but we cannot firmly assess its robustness, which we acknowledge as a limitation of our present work. The extent to which people form actor-specific inferences while browsing is an important future direction. If future research reproduced the pattern where only novel traits are less likely to be falsely recognized, it would be important to investigate the conditions under which actor-specific inferences do occur.

One limitation of this and prior research is the handling of extraneous characteristics of the stimulus material (e.g., faces). Prior research has included replication conditions with differently paired stimuli and shown that the pairing does not influence the overall effects, which led the researchers to drop the replication conditions from future work (Todorov & Uleman, 2003). However, the means of assessing that (through replication conditions which double the sample size) are not optimal. It would be possible to program studies in a way that randomizes stimulus pairs at each run thus reducing extraneous effects to random noise.

Bringing together research on snap social judgment and online impression formation opens up important directions for future work. Clearly, there is more to social media than what our experiments aimed to capture. Under conditions of true information overload, the ability of cues to attract attention might matter more than it did in the present research, thus it will be important to reproduce the effect under even more cognitively demanding conditions. Another direction for future studies we have already mentioned is whether and how spontaneous inferences can help users navigate and maintain their vast online networks. Examining the accuracy of inferences will be important for understanding such interpersonal effects.

Conclusion

There are many reasons to browse the streams of updates on social media sites—to kill time, to catch up on current events, to have a laugh at a friend’s joke. The present research suggests that, without necessarily having the intention to do so, people form impressions of others. These spontaneous processes paint a different picture of what browsing social media might entail. While scrolling down to catch a colleague’s most recent pun about conference deadlines, a social media user might also be picking up bits of information that shape her perceptions of people from various corners of her vast online networks.