When considering internet-based socialisation, the exact nature of communication is important. Digital technology allows for communication approaches that mirror real-life one-to-one (direct messaging), one-to-many (a company’s advertisements being shown to users on LinkedIn), many-to-one (users creating and signing petitions on official government websites), and many- to-many (collective editing and reading of entries on Wikipedia) communication [41, 66]. However, digital communication is often affected by site-specific algorithms, which can be defined as rulesets “behind the scenes” on a website that determine the specific information individuals are shown. This includes social media algorithms, which have been found to significantly alter the information we receive and retain [13, 100], in turn impacting on social beliefs and behaviours [42]. Brady et al. [13] stated that social media algorithms affect social perception by exploiting our existing biases - namely, posts involving PRIME (prestigious, in-group, moral, and emotional) social information are promoted, while those involving non-PRIME social information are suppressed. These strong social biases keep users focused and engaged with social media longer (therefore making social media companies more money from advertising revenue), but over time alter users perceptions, beliefs, and actions. For example, incorrect PRIME information is presented far more commonly (and to a wider audience) than correct non-PRIME information [13, 100]. This may lead to false information being increasingly accepted as truth [100]. Social media algorithms can then affect personal and social identities by creating “filter bubbles” [63] in which homogenous, confirmatory, information is presented that matches existing personal and in- group beliefs. Indeed, “filter bubbles” may create “echo chambers” by amplifying certain beliefs and creating confirmation biases [63]. This is much more common with individuals who hold “news-finds-me” beliefs – that is, individuals who more readily accept information that is algorithmically presented to them – especially when the information aligns with their political beliefs [100]. Over time this influences how users perceive the importance of different social groups, leading to increased social segregation even among groups that are not in active conflict [26]. This may lead to individuals’ social identities crystallising around the algorithmically boosted in-group, resulting in them abandoning groups that they had earlier belonged to, resulting in heightened tension between groups. This increased tension and segregation in turn heightens social misperceptions and intergroup conflicts [13]. In relation to health, a recent review of literature on social media use [88] determined that high levels of social media use – and likely therefore engagement with algorithmically promoted content – was associated with eight major psychological effects (anxiety, depression, loneliness, eating disorders, low self-esteem, low life satisfaction, insomnia, and stress) and three major physiological effects (physiological stress, change in physical brain structure, and affective experience state [increased anxiety, emotional valence, physiological arousal, and sustained attention]). Further, Walla and Zheng [96] determined that the P300 effect - related to attentional brain functions - was suppressed in those who watch four or more hours of short-form video content on social media sites/apps per week. These algorithmic shifts can have positive effects too; positive information is interacted with more frequently, leading social media algorithms to promote positive posts more often [8]. These positive posts include individuals acting in prosocial and civic-minded ways. Algorithmic promotion leads to these posts being seen, and reacted to positively, more often than would be the case otherwise [42]. This increases the likelihood of both the poster and postees (those who saw the post) repeating the prosocial action [42]. As such, it is important for researchers to determine how different social media algorithms work, and how they may be manipulated to reduce negative outcomes and enhance positive outcomes. This is a complicated topic as different social media websites use different algorithms, and host different kinds of postings (Bluesky for short-form communication, Facebook for longer-form communication and group membership, Tiktok and Instagram for short-form video communication, etc.), preventing a “one size fits all” solution.

Even though the specifics of what socialisation involves differs depending on whether it occurs via the internet or in real life, we treat all forms of socialisation as a unified system [29, 30]. This is based on the assumption that internet-based social networks almost perfectly replicate real-life socio-cognitive processes and social network structures. As an example, acceptance and rejection universally activate the same areas of the brain regardless of whether they occur digitally or in-person [1, 22], but are communicated differently in real-life (by body language) and internet spaces (by interactions with user posts). Social media algorithms come into play with these processes as well, as more interactions on promoted posts may artificially initiate feelings of acceptance, and fewer interactions on suppressed posts may initiate feelings of rejection [5, 37]. A person may therefore experience the cognitive impacts of social rejection from algorithmic suppression rather than true rejection. This may cause further issues such as, perceiving themselves as socially rejected and damaging their social reputation by lashing out at their friends.

Another recent development is the popularity of commercial social virtual reality experiences, such as VRChat, allowing for the examination of new social microcosms [18, 54]. These environments allow users to select and/or design avatars to represent themselves, and to communicate in real-time with others. Users commonly choose avatars that align with their self- perceptions and, when doing so, experience a heightened sense of presence and attachment compared to other avatar choices [32, 48]. For all avatars, research has determined that users experience “virtual embodiment”, which occurs when factors that are true for the avatar are accepted by the user as being, to some degree, true for themselves. For example, using an avatar associated with an individuals’ outgroup has been found to significantly reduce associated negative biases, while women using male avatars were able to avoid gender-based stereotype threat [4, 57, 65]. Others’ avatars are also accepted as true representations of them, and activate social cognitive effects associated with being around others [92, 95]. Interestingly, social dynamics can present in new ways based on program-specific elements. For example, VRChat implemented a “trust system”, where users could rate each others’ trustworthiness, as an attempt to reduce bad behaviour [18]. Studies have indicated that it was instead treated as an indicator of social rank, with stereotyping and intergroup conflict occurring between the “ranks” [18]. As such, while social virtual reality has the potential to allow researchers to control social scenarios than many real-life approaches, such as examining the interactions of globally distributed cultural diaspora (such as Romani), when provided with an avenue for easier embodied interactions, researchers must be careful in considering whether any elements unique to the simulated environment and its systems may impact their experiment.

Two core processes affected by internet use are memory and attention. The way internet has affected memory is related to the centralisation of information on the internet, increasing the ease with which information can be found compared to what was available pre-internet. This increase in the findability of information has reduced the quantum that individuals need to personally remember, moving aspects of social memory processes from semantic to transactive memory [29, 30, 67, 85]. Semantic memory - the explicit conscious long-term form of memory relating to meaning, understanding, and conceptualisation of facts built over the course of our lives, including social knowledge [77] has declined significantly since the advent of the internet, as the ability to quickly look up information undermines the necessity for memorisation [30]. Indeed, individuals familiar with the internet are significantly more likely to remember where to find information rather than to remember the information itself [86], although this is moderated by self-efficacy. Individuals high in self-efficacy are resistant to this shift as they more commonly adopt learning strategies that lead to long-term knowledge creation [97]. Transactive memory - the process by which information storage is outsourced to others in our community [30] – in this case treats the internet as a single hyper-knowledgeable individual. This centralises transactive knowledge by reducing the need for other humans to be involved [30]. As a result, traditional communally-held social knowledge may give way to internet-guided, globally standardised, social knowledge. This risks the loss of traditional knowledge that is not archived on the internet.

The way internet has affected attentional capability is related to the way information is conveyed via the internet, with current research focusing on the impact of social media. Research has determined that the amount of time an individual spends interacting with social media directly correlates with their attention span, with increased time inversely correlated with attentional ability and academic performance [64, 93]. Social media companies benefit financially from keeping people engaged; they therefore employ tactics to encourage users’ entering “flow states” where individuals’ spend copious amounts of time scrolling through the website without realising the time spent [75]. There are differences in the degree to which this is successful, even for similar social media sites. For example, while posts on both Instagram and TikTok are short-form video content, TikTok users are more likely to enter, enjoy, and remain in these flow states for a longer time [75]. If time spent engaging with social media is truly correlated with attention span, it is likely that individuals who engage heavily with TikTok are likely to have lower attentional capability than others who engage heavily with Instagram, with both being significantly worse than those who avoid social media. These findings are significant for research in social cognition, as impaired attentional ability may impact social cognition by interfering with or preventing the activation of appropriate social knowledge, therefore impacting stereotype creation and maintenance processes [40]. Individuals with lowered attention spans may fail to activate social stereotypes when interacting with others, and therefore may act differently. Other researchers have suggested that personality traits drive social media use [20, 49], with extraversion, openness to experience and, for men, emotional instability correlated with social media use [20], and with introversion, conscientiousness, agreeableness, and neuroticism correlated with problematic social media use [49]. As such, it is possible that individuals with these personality traits may also have lower attention spans. This raises the question of whether personality factors truly account for these attentional differences, or whether - while they may account for an individual engaging with social media in the first place – attentional degradation does occur. Future research may elucidate attentional degradation longitudinally, with these personality traits used as potential control variables.

The way the internet is used has led to multitasking behaviours becoming far more common [29]. Multitasking can be defined as the sequential (in order) and/or parallel (simultaneous) undertaking of two or more tasks in a fluid manner [28, 31]. Sequential multitasking commonly involves active tasks that require concentration and attention, while parallel multitasking commonly involves a passive task occurring alongside one or more active tasks [28]. Sequential multitasking occurs more frequently than simultaneous multitasking [28], but they are not mutually exclusive and often occur together [31]. For example, someone sequentially working on a series of tasks may, passively and in parallel, listen to music in the background. Studies on media-based parallel multitasking has determined that participants who frequently engage in parallel multitasking perform better in task-switching paradigms than those who infrequently multitask, with both groups performing equally well in dual-task paradigms (cite61). Fischer and Plessow [31] posited that research in multitasking has often focused on sequential and parallel multitasking in isolation, and that there is a need to examine both in combination during successful multitasking. Gaining a more accurate image of multitasking is especially important for internet-based experimentation, as nearly 60% of respondents to an internet-based questionnaire were found to lose concentration and/or to swap browser tabs away from the experiment [79].

To summarise, social and cognitive researchers must be aware that internet use significantly affects social cognition directly and through changes in memory and attentional processes, and that these changes affect social cognition in both digital and physical spaces. Furthermore, social and cognitive researchers should consider the potential impact of habitual multitasking activity, due to the high percentage of participants in internet-based studies showing evidence of multitasking and the potential of those who habitually multitask to have improved performance on specific kind of tasks. While understanding these changes is important for all social and cognitive research, it is especially important for internet- based studies that focus on cultures and/or communication, such as research examining social media [13, 25, 100] and social virtual reality [18, 54]. The considerations relating to internet use and the consequences of changes in memory and attention are addressed in the Methodology section.

In order to accurately assess the effects of internet use on social cognition, researchers must develop precise measures of how individuals spend their time on the internet. This includes an examination of the impact that different activities and websites have on wellbeing and social cognition [7] - both positively (activities/websites that increase attentional ability and/or semantic memory) and negatively (activities/websites that decrease attentional ability and/or semantic memory). This is a complex topic, and researchers must be mindful of connected issues that may confound these examinations, such as underlying differences in the populations that naturalistically choose specific activities/devices/etc. For example, older research examining users’ preferred personal-computer technology [16] identified Apple users as being more educated and having higher openness than PC users, pointing to cultural dynamics at that time – Apple computers were primarily used by those in artistic industries, which often required university degrees. More recent research into users’ preferred smartphone technology determined that personality factors did not influence smartphone choice [36], but internet use did [34]. Specifically, iPhone users generated the most web traffic per user, followed by Android users, and with users of other internet-enabled smartphones generating the least amount of traffic per user. Further, the very nature of conducting research via the internet may introduce bias. Those who spend more time on the internet are more likely to be willing to participate than those who spend little time on it. Experiments are therefore likely to be subject to self-selection bias, undermining the generalisability of the results obtained [11, 14, 71]. This is an issue for internet-focused research, as the method of collection is also the topic of examination. The participant pool is therefore likely to be skewed towards individuals with higher internet-use rates than is truly representative [14, 71]. Researchers therefore need to select recruitment approaches that ensure an equal number of participants are recruited for each category of interest and, when undertaking internet-based research, should consider inclusion of internet usage rates as one of these categories. This purposive sampling approach minimizes the impact of self-selection bias, and equal recruitment into each category allows for the results to be generalisable even when the population recruited from is not normally distributed [12, 76]. While this increases the complexity of examinations into positive and negative effects of different websites and activities, it is not a wasted effort as it greatly increases the nuances of the knowledge gained around the activities/websites examined. Until the conclusion of the study, researchers may consider controlling for the effects of internet use by employing purposive sampling and including questions that accurately profile their participants. For example, how participants split their time between activities (social media use, emails, information searching, recreational activities, etc.), or how the time spent on social media use is split between different social media sites. This will likely require more participants to be recruited for each experiment to maintain comparable sample sizes between categories of interest, or for deselecting participants while pruning uncommon categories (e.g., if only one participant uses Lemon8 for more than 4 hours a week, and they use no other social media sites/apps, removing both the participant’s responses and the Lemon8 category from analysis). Depending on the requirements of the experiment, researchers therefore need to consider whether the tradeoff between reduced experimental noise and an increase in the complexity of the experiment (including the number of required participants) makes sense for them, or whether to focus on subsets of the community (e.g., those who use social media for less than four hours a week).

Beyond merely controlling for internet use as a whole, there is a need to control for participants’ engagement with algorithmically-curated content. For example, individuals who actively seek out political information on social media (those who look to traditional sources and those who look to friends and family) are more likely to be able to identify disinformation and avoid echo chambers than those who passively receive news via their social media feeds [24]. They may also engage less with algorithmically-promoted content. Interestingly, aggravation with social media algorithms has driven the creation of certain browser- based plugins, such as Facebook Purity, that remove algorithmic influence. When these plugins are used, all posts by followees (individuals the user follows) and groups the individual is a member of are shown in reverse chronological order (newest posts first). As such, even though an individual may be on social media for many hours a week, the level to which they engage with the algorithm may be quite low. Conversely, as individuals with “news-finds-me” beliefs are quick to accept algorithmically-promoted content as true [100], individuals who may be on social media very sparingly may still engage with the algorithm quite highly. As such, researchers should consider how best to implement some measure of algorithmic interaction - both for participant selection and for use as a sliding-scale control variable. This may be quite complex, as the lack of transparency around social media algorithms suggest that participants are unlikely to know how often they are actively engaging with it. As such, these measures should focus on questions participants can easily answer, such as asking them (1) to estimate how much of their time on social media they spend interacting within familiar groups, (2) the level to which they actively seek information and check facts, and/or (3) the degree to which they enter a flow state while using social media (indicating the success of the algorithm in keeping them engaged). Further, the success of browser-based plugins that alter social media algorithms offers at least two interesting paths for research. First, researchers may explore the impact of social media without algorithmic influence on social perception and behaviour. Second, researchers may explore alternative “replacement algorithms” as controlled environments (researchers would have the needed transparency) and as an avenue for investigating the effect and effectiveness of algorithms designed to promote positive outcomes. For example, a “replacement” algorithm that suppresses posts reinforcing negative outgroup beliefs and, in line with Jung et al. [42], promotes posts showing prosocial behaviour.

It is also important to consider how to address potential multitasking issues. As mentioned earlier, 60% of respondents to a questionnaire study were found to have sequentially multitasked at least once during the experiment [79]. Parallel multitasking is also likely to impact responses, although the specific nature of this is unclear. For example, there has long been a debate on the effect of background music on cognition. On the one hand, background music has been found to be cognitively demanding during task performance [35, 53, 82, 83]. People are less likely to listen to music while engaging with complex tasks [35], finding lyrics more cognitively distracting than instrumental music [82], and finding music they enjoy more distracting than that they find less enjoyable [83]. Furthermore, it is easier to enter a flow state during task performance without music [53]. On the other hand, music can improve cognitive performance [19, 53, 82]. For example, positive music has been linked to increasing mood, which in turn boosts cognitive ability during task performance [53, 82]. Researchers wishing to control for sequential multitasking may, in line with Sendelbah [79], examine paradata, taking long delays and switching away from the experiment as proof of multitasking, but parallel multitasking is much more difficult to control for. Furthermore, the role of simultaneous sequential and parallel multitasking must also be considered [31]. This highlights three key aspects for researchers. First, research should examine whether multitasking experience is useful for experimental paradigms not yet identified. Second, research should examine the ways in which multitasking experience may interfere with responses to experimental paradigms not yet identified. For example, it may lead to earlier onset of response fatigue. In order to examine these key issues, researchers need to include measures of participants’ experience with both sequential and parallel multitasking, either using an evaluation task (testing executive functions), or a self-reported questionnaire. The latter may not always be informative, as participants have been found to be inaccurate in self-reporting their multitasking behaviours [79]. This leads to the third key point: when researchers are not able to include a direct measure of multitasking, they should take special care to stress the importance of avoiding external distractions such as music [47].

The shift from culturally-relevant semantic memory towards a globally standardised transactive memory with increasing internet use is likely to lead to issues in two ways. First, the general cultural knowledge of individuals is likely to decrease, potentially requiring certain experimental paradigms that have previously been normed for use with a specific group to require being normed again before they can be used. Second, the acceptance of globally standardised information includes information that, purposefully (“fake news” [6]) or accidentally (journalists’ factual reports including personal/cultural bias), is not aligned with the traditional cultural norms of a group. For example, if a US based social media site uses an algorithm that boosts posts in line with US cultural norms and suppresses posts in line with Norwegian cultural norms, then over time Norwegians who extensively use that social media site may begin to perceive US cultural norms as more socially appropriate than Norwegian norms. Research suggests that there have been purposeful attempts to use the nature of the internet to rewrite remembered history in specific ways [21, 44, 74]. For example, the widespread use of historical memes by far-right groups has been linked to conspiratorial beliefs in an “alternative history” [21, 74]. This includes swapping the roles of Bosnians and Serbians in the Bosnian Genocide, remembering it instead as a “white genocide” carried out by the Bosnians [74]. While many “memory changes” are likely to be less extreme, and less purposeful, than these examples, less controversial “alternative histories” may be harder to notice and challenge, before they have become generally accepted as truth. Conversely, social media also allows for the spreading of historical and cultural knowledge and practice that traditional communication approaches may sanitise [10, 55]. For example, Liebermann [55] identified peer-to-peer memory practices within digital networks - which therefore keep transactive memory between individuals - as a mechanism that defends the cultural histories and beliefs of minority groups against institutional hegemonic cultural histories and beliefs, allowing for culturally specific memories and beliefs to be communicated transnationally and transculturally. Researchers should consider including some measure of memory, whether that be a measure of the level to which participants rely on internet-supported transactive memory, or the level to which they have accepted or rejected revisionist histories.

Reductions in attentional capacity are of serious concern for social cognitive research, as it may affect complexity and length an experiment may have, before participants lose focus and/or drop out. Indeed, this may lead to those who excessively use the internet to produce more errors in their responses than those with low or regular internet use, as they may find it harder to keep their focus on the experiment. It follows, then, that many considerations related to attention may be addressed by controlling for internet use. However, as discussed earlier, it is possible that personality traits also play a role in attentional differences, increasing the likelihood of certain individuals to engage with social media, leading to a feedback loop where their attention is more affected and they become increasingly likely to engage with social media. As such, future research may explore the effect of the interplay between personality traits and attentional shifts over time in relation to social media use.

In line with the discussion of ecological validity and experimental control, setting up an internet-based experiment on a centralised platform may serve as a starting point for consideration. There are a variety of factors that must be considered, with three key sources of experimental noise: software, hardware, and experimental instrument [47, 72]. Software refers to not just the experimental software, but to all programs running during an experiment. Internet browser - and browser version - are easily detectable [3, 59], but background programs are harder to detect. Hardware refers to the specific make and model of computer/smart-phone/tablet, the participant is using to respond. These often differ between participants, but researchers may request the participants to provide their system information. This may, however, require that the researcher provides instructions to do so, with different instructions based on the hardware they are using. Experimental instrument(s) refers to the specific psychological instruments that data was collected within, with different instruments having different measurement accuracy and precision [47]. When designing and/or using a digital laboratory, one must therefore consider the degree to which the laboratory minimises these sources of noise. For example, a browser-based platform, such as Harvard’s Digital Laboratory for the Social Sciences (https://dlabss.harvard.edu/), increases experimental control in relation to the experimental instrument aspect; participants accessing the laboratory are shown a standardised formal template, with collection of survey data occurring via an embedded Qualtrics widget. A similar laboratory that wishes to allow for both survey and cognitive experimental data may look to embed instruments that allow for both types of testing, such as PsyToolkit [89], while one wishing to incorporate digital interviewing might look to embed a video conferencing option. A more complex alternative might be the creation of a partially or entirely stand-alone platform, where participants either download and run experiments locally on their computer, or download a client through which the experiments are run. This would remove experimental noise related to differences between browsers and browser versions [72]. Indeed, if participants provided the platform permissions, it could record what other software are running and their versions. However, the creation of a stand-alone platform would require significant time and resources to create and maintain.

Digital researcher presence, as a part of a digital laboratory or a stand-alone experiment, is another path forward for increasing experimental control. Researchers in social cognition have explored its efficacy across different experimental protocols and found promising results [9, 52]. For example, Leong et al. [52] examined whether Remote Guided Testing (RGT) - the digital presence of a researcher - improved data quality in internet-based response time experimentation. This was a between-participant experiment that built on previous research which determined that certain internet-based instruments do not entirely replicate response time data obtained in the laboratory [65]. Participants responded to experiments conducted in CANTAB, Inquisit, and i-ABC in either a laboratory (with a researcher physically present) or via the internet (with a researcher digitally present). Both groups were of equal age, with similar gender, ethnicity, education, and income distributions. Laboratory testing involved a 3.5 hour testing session where participants first completed demographic questionnaires, and then undertook the battery of cognitive tasks in the presence of the researcher. RGT was administered via a video conference platform, with participants first undertaking a 30 minute software task (installation, resolving technical issues, completing demographic questionnaires) followed by a 3.5 hour testing session involving the same cognitive tasks as the laboratory group, and with the researcher digitally present. Data quality was assessed using trial-level measures (missed trials, outlying and excluded responses, and response times) as well as participant task performance. Leong et al. [52] observed that participants in both samples produced statistically equivalent responses for most measures, however RGT participants’ verbal intelligence was significantly higher than that of participants in the lab. The authors interpreted this finding as an indication of RGT participants having higher ecological validity than laboratory-based participants. This suggests that digital researcher presence may be beneficial if incorporated into a digital laboratory platform, reducing noise related to the interaction between the browser and video platform used. However, digital researcher presence requires time effort of the researcher, removing that as a benefit of internet-based experimentation, and may not always be relevant or appropriate for a given experiment to implement. One potential solution is the use of AI-controlled avatars as the “present” agent [23, 91]. Recent research in telepresence has indicated that AI-controlled avatars provide the benefits of researcher presence while improving participants’ performance on cognitive relational-reasoning tasks, even beyond those obtained with a researcher digitally present [91]. Indeed, an AI agent is not required to be overly attentive as - regardless of whether the AI agent or a researcher was present - inattentive presence was found to improve accuracy, while attentive presence improved response times. While this point is based on just one study, it highlights promising methodological avenues for future research.