Ambient Awareness of Who Knows What: Spontaneous Inferences of Domain Expertise

ABSTRACT Being aware of others’ expertise is essential for knowledge exchange. Browsing social media has been shown to foster such awareness, often referred to as “ambient awareness”. However, the underlying psychological mechanism has not yet been established. In a series of three preregistered experiments (total N = 445), we tested the notions that individuals make inferences about others’ domains of expertise when being exposed to domain-implying social media posts, and that they do so spontaneously – that is, unintentionally and efficiently. Regardless of whether individuals were carefully reading, browsing through, or memorizing domain-implying posts, and whether they had an explicit goal to form an impression or not, we found consistent evidence that they made inferences about others’ expertise domains, as evidenced by responses in a (false) probe recognition paradigm and explicit domain identification (Experiments 1 − 3). We also showed that they did so unintentionally (Experiment 2) and efficiently (Experiment 3). Spontaneous domain inferences are a plausible psychological mechanism for how ambient awareness of who knows what is achieved through regular social media use. Our findings provide causal evidence for the proposed mechanism underlying the link between social media use and expertise localization and highlight the role of social media posts in this process.

. However, the psychological processes that contribute to these informational benefits are not well understood.
Generally, to successfully exchange information, localizing expertise in one's network is fundamental (Austin, 2003;Wegner, 1987).Gathering this knowledge of where expertise is localized is typically a time-consuming and effortful process and limited to the people in one's immediate social environment.By providing access to people around the globe -and even to people one has never met, social media reduce this restriction.However, the abundance of information available on social media makes it necessary to process this information efficiently to be able to fully benefit from it.This is exacerbated by the fact that content on social media is typically brief, momentary, and delivered to other users in real time, resulting in a stream of disconnected updates from different people.Consequently, single pieces of information are often not particularly informative and rarely given more than a passing glance.Therefore, processing and accumulating the fragmented information in an efficient way is considered key to really benefitting from the potential of social media.
One concept that has been proposed to be relevant in explaining how some of the informational benefits come about through social media use is so-called ambient awareness.Ambient awareness has been described as "the awareness that an individual has about the communications occurring among those around them" (Leonardi, 2015, p. 758), and later more broadly as the "awareness of social others, arising from the frequent reception of fragmented personal information, such as status updates and various digital footprints, while browsing social media" (Levordashka & Utz, 2016, p. 147).Thereby, it refers to the phenomenon that by regularly skimming social media updates, people seem to inadvertently gain a sense of what others in their online social networks are like, even in the absence of direct communication.Importantly for the present research, it can also include an awareness of "who knows what", that is, a sense of others' domains of expertise among one's coworkers (Leonardi, 2015;Leonardi & Meyer, 2015) and beyond (Levordashka & Utz, 2016).While correlational studies suggest that using social media is associated with awareness about others and what they know in both professional and nonprofessional settings (Leonardi, 2015;Leonardi & Meyer, 2015;Levordashka & Utz, 2016;Rice et al., 2017), and while this awareness is even considered to be one of the affordances of social media (Rice et al., 2017) highlighting its relevance -little is known about the psychological mechanisms through which it develops.
Importantly, two key characterizations of ambient awareness are that it develops spontaneously, that is, without the explicit intent to gather this knowledge and efficiently -without the need to exert effort much beyond what is required for browsing itself (Leonardi, 2015;Levordashka & Utz, 2016;Thompson, 2008).These propositions are crucial, because given the size and complexity of social networks and the sheer amount of information on social media, a deliberate process would be far less potent and efficient than a spontaneous one.Nevertheless, this central idea of ambient awareness of who knows what that others' expertise is inferred both unintentionally and efficiently, merely by browsing social media content, has not been tested experimentally to date.
Notably, there is a well-established phenomenon from social psychology that shares key characteristics with the concept of ambient awareness and that may be a plausible candidate for being a psychological mechanism underlying the development of ambient awareness on social media: Decades of research have established that individuals make inferences about others when observing or reading descriptions of brief instances of their behavior as part of their impression formation process (Ambady & Rosenthal, 1992;Uleman et al., 2008).In accordance with the key characteristics of ambient awareness, they do so spontaneously, that is, unintentionally and efficiently, and even in conditions of superficial processing or under high cognitive load (Uleman et al., 2008).Research has found people to make these spontaneous inferences for various types of social information including traits, beliefs, goals, values, states, and social roles (e.g., Chen et al., 2014;Kruse & Degner, 2021;reviewed in Uleman et al., 2008).While profession-related social roles are often closely linked with a certain expertise domain (e.g., finance for bankers, science for professors, and psychology for therapists), it is currently unknown to what degree the phenomenon of spontaneous social inferences extends to information about others' expertise and if so, whether expertise-related social media posts (rather than behavior descriptions) can trigger them.Both would be required if they are the mechanism underlying ambient awareness of who knows what.
Importantly, regarding the question of whether spontaneous social inferences may be the process underlying the emergence of ambient awareness on social media, first evidence suggested that individuals make spontaneous trait inferences based on social media posts.Levordashka and Utz (2017) conducted a series of experiments in which they presented participants with others' ostensible self-generated social media posts that implied certain traits without explicitly mentioning them (e.g., "Just spilled coffee all over my laptop!", implying the trait "clumsy").This exposure phase was followed by a probe recognition task in which the same actors were presented alongside words that either matched the implied trait (here: "clumsy") or not (e.g., "honest").The participants' task was to decide whether the displayed word had been part of the actor's post.Prior research demonstrated that when the probe word is the implied trait, participants are more likely to falsely recognize it as having been presented alongside the actor (e.g., Todorov & Uleman, 2002, 2003).If more trait-matching than trait-non-matching false recognitions occur, this indicates that trait inferences were made.Because the outcome measure is based on false recognitions (i.e., mistakes), this paradigm allows us to rule out that the effect occurs merely because participants memorize the sentences (otherwise they would make less rather than more mistakes).The results obtained by Levordashka and Utz (2017) overall suggested that as expected, individuals make spontaneous trait inferences based on brief and mundane social media posts.Here, we intend to examine whether the same is true for expertise domain inferences.
In summary, social media users have been demonstrated to develop an awareness of others' expertise in their network, but the psychological mechanism underlying this phenomenon has not been established.Here, we seek to establish spontaneous domain inferences as a potential psychological mechanism underlying the emergence of ambient awareness of who knows what on social media by addressing the question of whether individuals unintentionally and efficiently make inferences about other individuals' expertise domains when being exposed to their domain-implying social media posts.

The Present Research
In the present series of experiments, we set out to test whether individuals make inferences about others' domains of expertise when being exposed to their domain-implying social media posts (Experiment 1), and if so, whether they do so spontaneously, that is unintentionally (Experiment 2) and efficiently (Experiment 3).Each of these aims was addressed in one targeted experiment, relying on established, validated experimental paradigms specifically designed to test whether social inferences are being made, and whether they are made spontaneously (i.e., unintentionally and efficiently).We adapted Levordashka and Utz's (2017) design in a way that others' ostensible posts implied expertise domains (e.g., photography) instead of traits.In accordance with their studies, our primary dependent variable was the percentage of false (domain) word recognitions participants made after being exposed to domain-implying social media posts.This indirect measure of spontaneous domain inferences allowed us to rule out that effects merely occurred because participants had memorized the sentences (otherwise they would make less rather than more mistakes).In addition, as secondary, arguably more face valid, dependent variables, we moreover included a direct assessment of domain identification, where participants were asked to identify each actor's expertise domain from a list of domains, and recorded their nonresponse in the same task, indicating that no conscious domain attribution had occurred.
In Experiment 1, we aimed at generally establishing that individuals make inferences about others' expertise domains when being exposed to their social media posts (H1.1).To do so, we used an instruction that prompted participants to use a controlled processing mode ("read carefully") and compared whether more domain inferences occurred after domain-implying compared to domain-neutral posts.To test boundary conditions, we additionally varied the presentation mode (single post at a time vs. all posts at once like on a social media timeline) to examine the prediction that the effect would also be observed for posts presented on a timeline (H1.2) -a prerequisite if spontaneous domain inferences were indeed the process underlying ambient awareness on social media.
In Experiment 2, we set out to focus on whether individuals make these inferences about others' expertise domains without intention to do so.Specifically, as a stringent test, we examined whether giving individuals an opposing goal would prevent them from making expertise domain inferences (RQ2.1).To this end, we compared a condition in which an explicit task goal was provided that was consistent with making inferences (browse through posts and try to form impressions) with a condition in which an explicit task goal was provided that was inconsistent with making inferences (memorize posts).In the absence of domain inferences, memorizing the posts should -if anything -decrease the number of false recognitions.
In line with our first aim, we again expected participants to show domain inferences, at least when they had an explicit goal to form impressions (H2.1).Because we anticipated that when merely browsing through the posts, participants might miss individual posts by chance, we presented multiple posts instead of a single post for half of the actors, each implying the same domain and tested the intuitive prediction that domain inferences are more common for actors with multiple compared to single posts (H2.2) as a potential boundary condition.
Finally, in Experiment 3, we focused on examining the degree to which domain inferences are made efficiently, that is, automatically without or with only little controlled, effortful processing beyond what is needed for browsing itself (cf.Bargh, 1994;RQ3.1).Applying a paradigm from social cognition research, here, we compared domain recognition rates between a condition in which posts only implied a domain (here, automatic processing should result in more false domain-word recognitions, while controlled processing should result in more correct rejections) and a condition in which posts explicitly mentioned the domain (here, automatic and controlled processing should both result in correct recognitions; cf.Shimizu et al., 2017).If domain inferences are made efficiently, domain-word recognitions should occur regardless of cue explicitness, that is, recognition rates should not be moderated by whether domains were explicitly mentioned or simply implied in the posts.In line with our first aim, we again also predicted that individuals infer information about actors' expertise domains (H3.1); we again additionally predicted that the effect would be larger for actors for whom participants had seen multiple domain cues (H3.2).

Transparency and Ethics Statement
All three experiments were preregistered (Experiment 1: https://osf.io/ckv4d;Experiment 2: https://osf.io/ec7ws;Experiment 3: https://osf.io/h67kp);study materials and raw data are available at https://osf.io/mfb4z(Experiments 1), https://osf.io/dv3ya(Experiment 2), and https://osf.io/ugzxj(Experiment 3).Analysis scripts are available at https://osf.io/ugzxj.Deviations from the preregistered plans are listed in Table S1 of the supplemental materials.The study procedures were approved by the local Institutional Research Ethics Board (Experiments 1: LEK 2015/029; Experiments 2/3: LEK 2018/057).We had also conducted a fourth preregistered experiment as part of this project using the same general study design but for which the randomization procedure had failed (reported in the supplemental materials).The findings of that experiment were fully consistent with the findings presented here (see Figure S1 in the supplemental materials).

Hypotheses
In Experiment 1, we hypothesized that without explicit intent to do so, individuals infer information about actors' expertise domains after carefully reading a short domain-related social media update (H1.1), as evident in more false recognitions in domain-matching versus non-matching probes (H1.1a), above-chance domain identification (H1.1b), and higher non-response (skipping) for actors whose cues were neutral versus domain-implying (H1.1c).We also expected to find these effects when reading the posts on a timeline (versus one actor-post pair at a time; H1.2a-c).

Participants
A convenience sample (N = 110) was recruited through the online participant pool Prolific (prolific.co).Prolific is available for participants residing in most OECD countries (except for Turkey, Lithuania, Colombia, and Costa Rica) and South Africa.Participants received 1.3 GBP upon completion.Sample size was planned to ensure a power of ≥ 80% to detect a small to medium effect (d ≥ .4).The effect size estimate was based on prior studies using similar designs (Levordashka & Utz, 2017;Todorov & Uleman, 2002).We excluded data of participants who had skipped the attention question or reported low attentiveness (n = 3).Moreover, due to unforeseen errors in the experimental code, certain trials in the probe recognition task were not displayed and resulted in missing data; the missing trials occurred at random.The error affected 19 participants; we retained data of those participants affected by the error who had responded to at least three (out of nine) trials (n = 16), resulting in a final sample of N = 104 participants (33 women, 64 men, 7 missing; age in years: M = 27.37;SD = 7.93; 48 employed full-time, 10 employed part-time, 15 unemployed and job-seeking, 9 not in paid work, 9 other, 12 missing).

Design and procedure
The experiment had a mixed-design with one between-subjects manipulation (presentation mode: single post per page vs. all posts at once like on a social media timeline) and one within-subject manipulation (probe type: matching vs. non-matching domain).Dependent variables were false recognitions in the probe recognition task, correct responses in the domain identification task, and skipping rate in the domain identification task.
Exposure task.Upon providing informed consent, participants were presented with pairings of photographs of 18 faces (the "actors", selected from Bainbridge et al. (2013) 10k US adult faces database; 50% female, all Caucasian; for a full list of faces used, see Table S2 in the supplemental materials), with one text cue each, written in the style of social media posts.Nine of these text cues each implied a different expertise domain without mentioning it (e.g., "Spending days and days behind the lens, trying to capture the right moment and put it in my portfolio :)", implying expertise in the domain of photography), and nine of them were domain-neutral (e.g., "Oh, just remembered … happy Friday everybody!"; for a full list of text cues used, see Table S3 in the supplemental materials).Participants were instructed to carefully read through the posts at their own pace, without any mention of impression formation or expertise domains.Participants were randomly assigned to either viewing one post at a time or all posts at once like on a social media timeline (for an example of how stimuli were presented on a timeline, see Figure 1).To avoid potential confounds related to participants' stereotypes of certain domains, we generated a unique combination of actors, actors' names (randomized within matching genders), domains, and specific post per domain for each participant by randomly selecting stimuli from a larger pool (for details, see Tables S2 and S3 in the supplemental materials).
Probe recognition task.Subsequently, participants completed a forced-choice recognition task adapted from Todorov and Uleman (2002) that assesses inferences indirectly.Participants saw each actor from the exposure task, paired with a word, and were asked to indicate whether this word was mentioned in the sentence they had previously seen with the actor.Each actor whose cue implied a domain was presented twice: Once with the domain-word implied by the cue (match) and once with the domain-word implied for another actor (mismatch).Actors whose posts did not imply any domain (neutral) were presented with a word that had been part of the sentence from the exposure task, for a total of 27 trials (nine domain-matching, nine domain-mismatching, nine neutral).In each trial, participants indicated whether the word presented next to an actor's face had been present in the sentence they had previously seen with the actor ("old") or whether it was a new word ("new").All actors for whom a domain had been implied were presented with a word that had not appeared in the sentence, therefore responses "old" were always incorrect (false recognition).

Domain identification task.
The probe recognition task was followed by a direct assessment of domain identification, where participants had to identify each actor's expertise domain from a list of nine domains, with an explicit option to skip the question.Domain identification was considered correct when the response participants provided in the domain recognition measure matched the one implied by the actor's cue in the exposure task.Separately, we coded whether participants selected "skip" rather than providing an answer.

Additional measures.
At the end of the study, participants identified their own domains of interest/expertise from a list of all domains included in the study, reported how frequently they used social media, and estimated their level of attentiveness throughout the study (0 = answered all questions carefully to 4 = responses were mostly random).We recorded the time they spent reading posts in the exposure task, as well as their total study duration.Demographic information was retrieved from the participant recruitment platform.

False recognitions
Domain-neutral trials had been fillers, so analyses of false recognitions focused on domain-related trials only.Trials within each within-subject condition were aggregated to a mean score.
In order to test H1.1a that individuals infer information about actors' expertise domains after reading a short social media update, as evident in more false recognitions in domain-matching compared to non-matching probes, and H1.2a that this effect would also be found when people read the posts on a timeline, we conducted a mixed analysis of variance (ANOVA) in which probe type (within-subjects factor: matching vs. non-matching domain) and presentation mode (between-subjects factor: single post at a time vs. timeline) were entered as independent variables; mean number of false recognitions was entered as the dependent variable.As predicted, we found significantly higher false recognition rates in the matching (50%) compared to the non-matching (39%) domain condition, F(1, 102) = 16.92,p < .001,η part. 2 = .14,90% CI [0.05, 0.25], supporting H1.1a (see Figure 2a).No significant interaction effect between probe type and presentation mode was observed, F(1, 102) = 1.25, p = .27,η part. 2 = .01,90% CI [0.00, 0.07], so in line with H1.2a, the effect of probe type was found regardless of presentation mode.We also did not observe a main effect of presentation mode, F(1, 102) = .03,p = .87,η part.
Because participants were asked to select from 10 options (9 domains and "skip"), we considered the probability of correctly guessing the implied domain to be 10%. 1 We used one-sample t-tests to compare the observed number of correct responses to what could be expected by chance.In both presentation modes, participants' responses were significantly higher than chance at recognizing the domains implied by actors' posts (single post condition: 28%, t(52) = 6.34, p < .001;timeline condition: 30%, t(50) = 6.09, p < .001),thereby providing support for H1.1b and H1.2b.

Discussion
The results of Experiment 1 provide initial evidence that, as predicted, individuals make inferences about who knows what in a social media-like environment, without explicit intent to do so.However, based on this finding alone, it cannot be ruled out that participants developed an implicit goal to form impressions while being exposed to the fictitious social media posts.Therefore, we conducted a second experiment explicitly aiming at testing whether a goal to form impressions is required for domain-related inferences to occur.To this end, we compared a condition in which participants were explicitly instructed to browse through posts and try to form impressions (consistent goal condition) with a condition in which participants were instructed to memorize the posts -a goal that should, if anything, render false recognitions less likely (inconsistent goal condition).
In Experiment 2, we moreover aimed to extend Experiment 1 in four additional important ways: First, instead of instructing participants to carefully read the posts, in the consistent goal condition, we asked them to browse through the posts as they would on social media.This was done to induce a more superficial processing style that resembled the one typical for social media.Second, because we anticipated that when merely browsing through the posts, participants might be less likely to notice individual posts, we additionally included actors with multiple posts that implied the same domain and investigated whether domain inferences might be less common or absent for actors with single cues.Third, actors in Experiment 1 were demographically homogeneous -White women and men in their twenties, thereby limiting the generalizability of the findings.In Experiment 2, we thus selected ethnically and age-diverse actors.Fourth, the diagnostic validity of the posts in implying a certain expertise domain had not been explicitly tested.If posts are low in diagnostic validity, no meaningful inferences can be made.Therefore, we decided to conduct a series of pretests with the stimulus materials prior to Experiment 2 and adapted the stimulus material where needed to ensure high diagnostic validity across all stimuli.Finally, Experiment 2 was supposed to be a partial replication of Experiment 1 in a larger sample.

Hypotheses
In Experiment 2, we planned to first examine whether a goal to form impressions was required for inferences to occur (i.e., whether the predicted effects of probe type were moderated by goal condition; RQ2.1).We also again predicted that individuals would infer information about actors' expertise domains -at least in the consistent goal condition (H2.1), as evident in more false recognitions in domain-matched compared to domain-nonmatched probes (H2.1a) and above-chance domain identification, at least when multiple posts per actor were presented (H2.1b).We further predicted that individuals would be more likely to infer the domain expertise of actors for whom they had seen multiple domain cues in the impression formation goal condition, as evident by larger inference effects in the false recognition paradigm (H2.2).

Participants
Participants were recruited on prolific.co.They were paid GBP 2 upon completion.We restricted the sample to individuals who were currently living in the United Kingdom or Ireland, who were either LinkedIn and Twitter users, and who reported full-time or part-time work as their employment status.In accordance with the power calculations based on effect sizes obtained in an additional experiment (reported in the supplemental materials), our planned, preregistered sample size was N = 200 (with replacement for exclusions).In total, N = 211 participants completed the study; n = 4 of these were excluded from analyses because they withdrew their data upon completion of the study and another n = 2 because they either reported low attention or because they skipped the attention question altogether, resulting in a final sample size of N = 205 (143 women; 61 men; 1 gender-fluid, non-binary, and/or two-spirit; age in years: M = 33.36;SD = 9.98; primary employment status: 139 employed -mostly desk work, 29 employed -mostly manual work, 10 freelancersmostly desk work, 2 free-lancers -mostly manual work, 18 students, 4 stay-athome parents, 3 other).

Materials, design, and procedure
Design and procedure of Experiment 2 were identical to Experiment 1 unless stated otherwise.To explicitly address whether inferences are made without intention to do so, we introduced goal condition (consistent vs. inconsistent with making inferences) as an additional between-subjects factor.In the consistent goal condition, participants were instructed to browse through posts and try to form impressions, whereas in the inconsistent goal condition, they were instructed to memorize the posts -a goal that should, if anything, render false recognitions less likely.As a result, the experiment had a 2 between (impression formation goal vs. memorizing goal) by 2 within (probe type: matching vs. non-matching domain) by 2 within (single vs. multiple domain cues) mixed design.Instead of a fully randomized design, we opted for a simpler pseudo-randomized design in which each participant was randomly assigned to one of four exposure blocks in each of which different combinations of actors, expertise domains, number of cues were presented in random order.Controlling for block did not qualitatively impact the results, so analyses were collapsed across blocks.Exposure time to the posts was fixed to 2.5 minutes to make the two goal conditions directly comparable.
Each participant saw a set of 40 posts from 16 different actors in random order on the timeline (see Figure 1 for an example).The posts of eight actors implied expertise domains.Four of these eight actors had four different posts, each implying the same domain (multiple cues condition) and four had a single post (single cue condition).The remaining eight actors had domainneutral posts.Again, half of these had four posts and half had a single post.To increase the generalizability of our findings to individuals with more diverse demographic characteristics, we selected 16 new faces (50% female, 50% male; 38% White, 25% Black, 13% Asian, 13% Hispanic, 13% Middle Eastern; 44% in the age group 20 − 30 years, 56% in the age group 30 − 45 years) of similar attractiveness, competence, and memorability from the Bainbridge et al. (2013) database (for details, see Table S2 in the supplemental materials).Additionally, to ensure that text stimuli were sufficiently diagnostic regarding expertise domains, we conducted a series of pretests and adapted the stimulus materials where needed (for a more detailed description of the pretests and a full list of social media posts used in each of the experiments, see Table S3 in the supplemental materials).
The probe recognition task was identical to Experiment 1 with the exception that there were only 24 trials (eight domain-match, eight domain-mismatch, eight neutral) and 16 actors.For the domain identification task, participants identified each actor's expertise domain from a list of eleven domains, again with an explicit additional option to skip the question.

Additional measures.
After the primary study tasks, we measured participants' impressions of the actors' approachability and competence with a series of forced-choice trials as an additional measure (see Figure S2 in the supplemental materials for details and results).We also assessed participants' intraorganizational and extraorganizational networking, using the short networking behavior scale (Wolff & Spurk, 2020) Results regarding this measure are displayed in Tables S4 and S5 in the supplemental materials.Finally, participants completed some additional measures, including questions assessing their attentiveness and basic demographics.

False recognitions
In accordance with our preregistered analysis plan, we first tested whether there would be an interactive effect between probe type and goal condition (RQ2.1) when conducting a mixed ANOVA with the following specifications: probe type (within-subjects factor: matching vs. non-matching expertise domain), number of cues (within-subjects factor: single vs. multiple), goal (between-subjects factor: impression formation vs. memorizing), and their interaction terms; dependent variable: number of false recognitions.The analysis revealed that there were no significant interaction between probe type and goal condition, F(1, 203) = 0.28, p = .599,η part. 2 = .001,90% CI [0.00, 0.02], and no other two-or three-way interactions involving goal condition (all ps >.15, all η part. 2 ≤ .01).This allowed us to use the same analysis to test our hypotheses H2.1a and H2.2 across conditions: Consistent with H2.1a, we found a main effect of probe type in the predicted direction, F (1,203) = 281.57,p < .001,η part. 2 = .59,90% CI [0.53, 0.65], which was qualified by a two-way interaction between probe type and number of cues, F(1, 203) = 61.14, p < .001,η part 2 = .23,90% CI [0.15, 0.30], also in the predicted direction: As depicted in Figure 2b and in line with H2.2, the effect of probe type was larger when four cues had been presented (matching cue: 74%, non-matching cue: 25%) compared to when only one cue had been presented (matching cue: 49%, non-matching cue: 22%).

Discussion
Experiment 2 successfully replicated the results obtained in Experiment 1, demonstrating that individuals make inferences about who knows what in a social media-like environment.Importantly, Experiment 2 also showed that individuals make these inferences unintentionally, even when they have an opposing goal.Additionally, they did so even when the instructions requested to merely browse through domainimplying social media posts, which was intended to induce a rather superficial, effortless processing style.However, neither of the two experiments provided a stringent test as to whether these inferences are made with limited effort -which would be required if they underlie the phenomenon of ambient awareness.Therefore, we conducted Experiment 3, in which we examined the degree to which the domain inferences are made automatically/efficiently, that is, with little controlled, effortful processing beyond what is needed for browsing itself (cf.Bargh, 1994).For this, we compared domain word recognition between a condition in which posts only implied a domain (cf.Experiments 1 and 2; here, automatic and controlled processes work in opposition: while automatic processing should result in more false recognitions, controlled processing should result in more correct rejections) and a condition in which posts explicitly mentioned the domain (here, automatic and controlled processing work in concert and should both result in correct recognitions; cf.Shimizu et al., 2017).

Hypotheses
Consistent with Experiments 1 and 2, we again predicted that individuals infer information about actors' expertise domains (H3.1), and, in accordance with the findings from Experiment 2, that the effect would be larger for actors for whom they had seen multiple domain cues (H3.2).In addition, we also asked to what degree recognition rates differ when domain cues were explicitly mentioned (i.e., correct recognition) or only implied (i.e., false recognition; RQ3.1).design, and procedure of Experiment 3 were identical to Experiment 2 with the following exceptions: All participants received the instruction to browse through the posts as they would on social media (that is, they had no explicit goal that was consistent or inconsistent with forming domain-related impressions).The posts of four out of the eight actors who had domain-related posts (two who had one single post and two who had four posts) finished with a hashtag explicitly mentioning the expertise domain (e.g., "#photography"), while the posts of the other four actors who had domainrelated posts again only implied the expertise (but were otherwise identical; see Figure 1 for an example).As a result, Experiment 3 had a 2 (explicit vs. implicit domain cue) by 2 (probe type: matching vs. non-matching domain) by 2 (single vs. multiple domain cues) repeated-measures design.
Probe recognitions results of Experiment 3 are depicted in Figure 2c.In accordance with our preregistered analysis plan, we tested our H3.1,H3.2, and RQ3.1 in a repeatedmeasures ANOVA with the following specifications: domain type (withinsubjects factor: matching vs. non-matching expertise domain), number of cues (within-subjects factor: single vs. multiple), cue explicitness (within-subjects factor: implicit vs. explicit), and all interaction terms; dependent variable: number of "yes" responses.As predicted in H3.1, we found a main effect of probe type in the predicted direction, F (1,135) [0.08, 0.25], again in the predicted direction (cf.H3.2):As depicted in Figure 2c, the effect of probe type was larger when four cues had been presented (matching cue: 73%, non-matching cue: 26%) compared to when only one cue had been presented (matching cue: 55%, non-matching cue: 27%).Regarding RQ3.1, there was no significant interaction effect between cue explicitness and probe type, F(1, 135) = 0.38, p = .539,η part.
2 < .01.This means that we did not find any indication that inferences about others' expertise domain when browsing through domain-relevant posts require controlled processing beyond what is needed for the browsing itself.

General Discussion
In a series of three experiments (total N = 445), we investigated whether individuals make inferences about what others' expertise domains are (i.e., what they know) when being exposed to their social media posts, and whether they do so spontaneously (i.e., unintentionally and efficiently).We found that this was consistently the case regardless of whether posts were presented one at a time or on a timeline (Experiment 1), whether individuals had an explicit goal to form an impression (Experiments 2) or not (Experiments 1 − 3), or even an opposing goal (Experiment 2), indicating that the inferences were made unintentionally.Furthermore, the effect was both observed when instructions prompted deeper (Experiment 1 and 2) and more superficial processing (Experiments 2 and 3).Finally, the effect was not modulated by whether domains were explicitly mentioned or simply implied in the posts (Experiment 3), suggesting an automatic, efficient process.Relying on established, validated experimental paradigms specifically designed to test whether social inferences are being made (as opposed to results being an artifact of participants memorizing posts), and whether they are made spontaneously (i.e., unintentionally and efficiently), strengthens conclusions that can be made based on the findings regarding the mechanisms at play.Together, the evidence presented here indicates that when being exposed to others' expertiserelated social media posts, individuals spontaneously make inferences about what these people's expertise domains are.
The implications of these findings for media psychology are substantial.First, in demonstrating that individuals make expertise domain inferences based on social media posts and that they do so spontaneously, the evidence provided here suggests a plausible causal mechanism for how awareness of who knows what can be achieved through regular social media use, and that this awareness can indeed be "ambient."Because localization of expertise is an important antecedent of knowledge exchange, ambient awareness of who knows what may be one pathway through which the positive associations between social media use and informational benefits, demonstrated both concurrently and over time (Leonardi, 2015;Leonardi & Meyer, 2015;Utz, 2016;Utz & Breuer, 2016, 2019), come about.
Second, the evidenced process points to a beneficial role of social media posts when it comes to localizing expertise.Previous research alone can hardly speak to whether posts are a valuable feature of social media.One concrete example would be the question whether posting should be a feature of enterprise social media.While browsing social media is intrinsically gratifying, our research shows that it can also be useful.This is particularly relevant against the background that prior research in the context of news consumption on social media indicates that individuals who perceive that browsing social media is sufficient for them to stay up to date (often referred to as newsfinds-me perception) in fact show inferior results on news-related knowledge tests (Gil de Zúñiga & Diehl, 2019;Gil de Zúñiga et al., 2017), demonstrating how important it is to systematically study these effects.Whether the benefits of browsing posts overall outweigh potential costs are a critical next question.
Our research further revealed some noteworthy nuances in the process of ambient awareness.In line with our hypotheses, the effect of browsing through domain-implying posts had a stronger effect on false recognitions when multiple domain-implying posts per actor were presented than when only a single domain-implying post per actor was presented (Experiments 2 and 3).In line with Levordashka and Utz's (2016) definition of ambient awareness, this supports the notion that repeated exposure to fragmented domain-implying information on social media fosters the formation of spontaneous domain inferences while single exposure may not always be sufficient -especially when domain-implying posts vary in how diagnostic they are as was the case in Experiment 1 (and as is arguably likely the case in most real social media feeds).Experiments 2 and 3 however powerfully demonstrate that when domain-implying posts are highly diagnostic, spontaneous inferences about another person's expertise domain are formed after as little as browsing through a single post by about 50% of people in a controlled experimental setting with fictitious actors -an effect that gets even stronger (over 70%) after multiple exposure.
A key methodological strength of this research is in combining indirect (false domain-related word recognitions) and direct measures (domain identification) of spontaneous domain inferences in our experiments.We found that the effects observed between the two measures were highly consistent within and across experiments.Adding an indirect measure allowed us to rule out the possibility that the observed effect was merely an artifact of participants memorizing the stimulus materials.While memorizing the materials would result in better domain identification in the direct measure (by remembering the posts' content in order to identify the respective domain when prompted), as it indeed did in Experiment 2, it should -if anything -decrease false recognitions in the indirect measure unless spontaneous inferences were indeed formed during the social media browsing phase.

Limitations and Future Directions
In our experiments, participants knew that the social media streams they were presented with were fictitious, potentially limiting generalizability to real-life social media skimming.However, we would argue that if these spontaneous inferences occur in the artificial setting of an online experiment featuring fictitious people as stimuli, individuals would (if anything) be more ready to spontaneously make those types of inferences about real people in their own social online networks as these have a much higher relevance for them.Indeed, prior research demonstrated that individuals have a sense of the profession/ professional interests of members of their real social online networks that they have never come across or interacted with outside of social media (Levordashka & Utz, 2016).
Our interpretation that domain inferences are made efficiently is based on the observation that no significant difference was found between a condition where automatic and controlled processing would oppose each other and a condition where automatic and controlled processing would lead to the same outcome, with very small effect sizes.However, it should be noted that strictly speaking, the absence of a significant effect does not provide support for the absence of an effect in the population.Therefore, future studies should test whether there is indeed no effect using confirmatory equivalence testing (see e.g., Lakens et al., 2018;Weber & Popova, 2012).
While demonstrating that individuals make spontaneous inferences about others' expertise domains when browsing through a social media feed-like environment is very relevant from a media psychological perspective, it should be acknowledged that spontaneous domain inferences, as introduced here, seem to conceptually overlap with social role inferences, as described by Chen et al. (2014).These included different professions (e.g., therapist, banker, reporter) in addition to non-profession-related social roles (e.g., neighbor, student, friend).To what degree spontaneous domain inferences can be considered a separate type of social inferences, may rather be a subtype (or secondary effect) of spontaneous social role inferences, or are part of a broader inference type encompassing both spontaneous role inferences and spontaneous domain inferences will need to be established in future research.
Furthermore, the extent to which individuals can make inferences about others' professional expertise depends on the degree to which they are exposed to expertise-implying content.Although our stimulus materials were adapted from real social media posts, showing that individuals do indeed post expertise-implying content on social media, we have no empirical data on how common it is for people to post such information and to what degree people differ in how often they do so.Relatedly, our results clearly show that more spontaneous inferences about others' expertise domain are made the more diagnostic the respective person's posts are.Automated analyses, which identify expertise-related content on a large scale, ideally combined with measures of how diagnostic posts are for the specific expertise (e.g., through linguistic markers), would be an interesting means for exploring this question.
Finally, social media users may often be aware of their contacts' broad expertise domains prior to being exposed to their posts regularly by adding them to their contact list.Therefore, some may argue that inferring others' expertise on social media is of limited value.However, many social media platforms, including enterprise social media, but also the micro-blogging platform Twitter and the professional social networking platform LinkedIn do not only expose users to posts of their direct contacts but also to posts of network members that they are only indirectly connected with, whether it be through a common contact, by following the same hashtag, being in the same group, or alike, which they can then decide to follow/connect with if they appear to have a relevant expertise.This may even be more pronounced in the decentralized micro-blogging platform Mastodon where users are exposed to timelines with three levels of abstraction: home (containing posts by followed users), local (posts from users of the same home instance), and federated (posts from all users; cf.La Cava et al., 2022).Here, users may overall be exposed to more posts by strangers whose domain expertise they do not know.Relatedly, similar mechanisms as the ones observed here might allow to infer more fine-grained information about others' expertise including changes of specializations over time and level of expertise.While the present series of experiments demonstrates that it is possible for individuals to infer other social media users' broad expertise domains based on their expertise-related posts and that individuals do so unintentionally and efficiently, future studies will be needed to test whether users can also infer other important aspects of expertise such as one's specialization within a broader domain and one's level of expertise.

Conclusion
Across three pre-registered experiments, here we provide consistent evidence that individuals make inferences about what others' expertise domains are when being exposed to their domain-implying social media posts, and that they do so spontaneously.These unintentional, efficient inferences are a plausible psychological mechanism that may underlie the phenomenon of ambient awareness of who knows what in one's online social network that can ultimately foster successful information exchange and benefits facilitated through social media use.

Figure 1 .
Figure 1.Example snippet of "social media" timeline used in experiments 1 − 3. Note.Example snippet of "social media" timeline during the exposure phase using a publication-friendly selection of faces from the Bainbridge et al. (2013) 10k US adult faces database; different faces from the database were used in the actual experiments.The example timeline shows a post implying (but not mentioning) a domain of expertise (Experiments 1 − 3), posts explicitly mentioning domains of expertise (Experiment 3 only), and a domain-neutral post (Experiments 1 − 3).

Figure 2 .
Figure 2. Results of probe recognition task for experiments 1 − 3. Note.False recognition rates for a) Experiment 1 and b) Experiment 2 by conditions; c) correct (left panel) and false (right panel) recognition rates for Experiment 3. Dots depict mean values and whiskers depict 95% confidence intervals.