‘Hey Alexa,what did I forget?’: Networked devices,Internet search and the delegation of human memory

Introduction

Networked devices with search capability, from smartphones to smart speakers, such as Google Home and Amazon’s Echo, are emerging as a key interface between everyday human activity and information networks. They significantly expand our capacity to find and retrieve digitally encoded information without the deliberation often required by the stand-alone computer. Information that might have been out of reach is now clearly within reach due to the omnipresence of these devices and the increase in the speed of wireless and mobile networks. Such technological changes could enhance our cognitive capacity by allowing a much more efficient delegation of information to external sources, thus freeing the mind for more difficult cognitive tasks. Following this logic, the greater the facility by which information can be exchanged, the more that ‘external memory’ on the device or network can achieve parity with human memory. In many respects, the delegation of memorisation to devices is a positive advance, for we already have, in many cases, abrogated the requirement to remember to diverse forms of technology from the computer to the book. On the one hand, networked devices could represent just another stage in the augmentation of human memory, on the other, they are representative of a general diminution of the human capacity to remember, and therefore could be subject to the same criticism that Socrates and Plato levelled at writing. To investigate how networked devices both diminish and augment human memory, the article first draws upon Daniel Wegner’s theory of ‘transactive memory’, in which the task of remembering information is shared across social and work groups and is, therefore, externalisable (1995). It then considers previous research on the ‘Google Effect’, in which the facility with which search engines access external knowledge can affect memorisation, as well as the role of human expertise. These arguments are expanded to account for changes in the interface of networked devices, in particular, the shift from keyboard and screen to voice activation. We argue that while new interfaces might afford more efficient and seamless access to information, they could diminish cognitive deliberation by pre-empting a user’s thoughts and actions in times faster than actually accessing human memory.

Networked memory

The theory of transactive memory is a worthwhile starting point for an examination of the relationship between networked devices and memorisation, because it outlines how information is distributed across members of a human group or network. The theory was developed by psychologist Daniel Wegner (1986) to explain how we share intellectual and cognitive resources – characterised in the theory as shared memories – in work and family groups. The distribution of information within a human transactive memory system is directly compared by Wegner to a computer network in which each computer has its own memory. Because memory is individualised in such a network, each person, or computer, has to engage with the others to find out what is stored in their memory and, in order to do so, they must have some general idea of what is remembered by the others, which Wegner describes as a directory system (1995: 323–324). For this system to work efficiently, the directories must be continually updated to properly reference what is stored on the hard drive or personal memory, and in the interest of efficiency new information should be distributed to the computer that has the most suitable directory system (Wegner, 1995: 325). In short, there are three main elements to efficiently sharing information on a computer and human network: ‘directory updating, information allocation, and retrieval coordination’ (1995: 320).

In human networks, directory updating involves obtaining knowledge about the expertise of others, possibly through the application of stereotypes or gleaning information through conversation or in everyday working practice (Wegner, 1995: 327). Information allocation describes who remembers what within a group, in which remembering is organised according to expertise rather than each person remembering everything (Wegner, 1995: 329). Retrieval coordination concerns developing an overarching directory that coordinates all the other directories in the transactive memory system and knowing the procedure for acquiring information outside of one’s individual knowledge. Often the individual can decide to retrieve information elsewhere without tackling a problem first or drawing on their ‘own memory’ (Wegner, 1995: 333), which is particularly important in the delegation of information retrieval to networked devices and search engines.

In short, a complete transactive memory in a group occurs when each member keeps current on who knows what, passes information on a topic to the group’s expert on the topic, and develops a relative sense of who is expert on what among all group members. (Wegner, 1995: 326)

Despite some similarity between externally stored information and internal memory, the retrieval of the contents of memory (retrieval coordination) depends on group communicative practices, for an individual does not have a full understanding of what is remembered by others in the group. Instead they use ‘labels’ – essentially a directory structure – that point to the location and give a general idea of what has been remembered by others (Wegner, 1986: 188–189). Knowing how to retrieve information is based on the association of memory labels with particular individuals. The better the distribution of information, the better the group can function in response to a particular task. This use of labels greatly increases the breadth of information that an individual has access to because it is only a matter of remembering the location and not the contents:

In the course of a day, much of what we encounter is encoded internally, but probably much more is encoded externally. This is because the labels we encode for externally stored information can be very general, referring to hundreds or even millions of items. We surely cannot encode this much information internally in a short time, and it is for this reason that a large part of our internal storage capacity is devoted to location information that allows us to retrieve external items. (Wegner, 1986: 189)

Comparing memory to information shared on a network could be criticised for adopting what phenomenologist Don Ihde describes as the ‘otherness thesis’, in which human thought, perception and behaviour are explained through a familiar non-human metaphor – typically the behaviours of animals or the operation of machines – that does not sufficiently consider human intentionality (1983: 73–75). This would be a legitimate criticism if the model posited broad ‘mentalist’ claims. In fact it is quite modest, seeking only to explain the communicative processes that allow for memories (information) to be shared or accessed within a group (Wegner, 1986: 188–186). Despite its modest scope, the theory can broadly apply to most social groups from workplace teams to friendship and family groups. In each case, the delegation of memory contributes to the maintenance of the group by fostering communicative interdependence and sharing. Moreover, the metaphor of networked memory can also provide a basis for understanding how groups and individuals interact with devices that store and retrieve information from networks. It is only by first thinking through the metaphor of the network that we can come to an understanding of the limits through which human memory and cognition adapts to its computational environment.

To add one more caveat, the use of the term memory to describe information sharing is somewhat contentious because human memory, due to the interaction between conscious and nonconscious activities, cannot be compared easily to the storage of information in a network or database. Philosophers use a variety of terms in order to examine this relationship: occurrent memory, secondary memory, recollection, dispositional memory, habit memory, primary memory, and so on. Rather than work through each of these typologies, we will use Endel Tulving’s (1985) commonly adopted distinction between three memory systems based on types and levels of awareness, because it can be readily used to compare human and non-human aspects of information storage and recall. Procedural memory describes the type of memory that is implicit within action and perception and is usually understood in terms of capacities and skills. Semantic memory is propositional in form and describes representations of the world. Episodic memory combines some of the aspects of the other two but primarily refers to personal knowledge of events that remain situated within an individual’s past (Tulving, 1985: 2). In the transactive memory model, memory is mainly propositional and therefore linked with semantic memory, however, we argue that depending on when, where and how fast memories are retrieved, we could also be talking about procedural memory. Because networked memory is generalisable rather than rooted in personal experience, episodic memory is much less relevant to the discussion. Even when devices draw upon personal information to pre-empt a user’s intention, the information is mainly propositional.

Transactive memory and computational devices

The theory of external and transactive memory is not restricted to the analysis of human groups for it can also be used to analyse how individuals delegate memory to an external source, or ‘external memory’ (Wegner, 1986: 187), which may include, in addition to notebooks and personal files, computer databases and the Internet. Labels indicate where information can be found, for example, an individual knows that they can find information in a shared calendar about future meetings, even if they did not input the data, or consult an online dictionary for a definition, in a way that is similar to consulting a colleague about a field in which they have expertise. Following this line of enquiry, Sparrow et al. (2011) analysed the way that search engines and online databases are utilised in the delegation of memory and argue that because they ensure ‘future access to information’, individuals are less likely to go to the effort of remembering something (p. 776), which is generally referred to as the ‘Google Effect’. In a series of empirical studies that gauged reaction times to general knowledge questions, the researchers concluded that participants were prone to thinking about search engines when seeking answers to knowledge problems. Interestingly, this occurred when participants sought new information, that is, when confronted with difficult questions, but also when they were presented with easy questions whose solutions would likely be known by the participants and therefore retrieved from their own memory (Sparrow et al., 2011: 776). From what is essentially a small study, they demonstrated that Internet databases are readily sought when addressing questions of fact, that is, types of propositional knowledge commonly associated with semantic memory. This is hardly surprising and would agree with the common-sense assumption that the more we use online databases and search engines, the more they will be called on to address standard questions of knowledge. What is more interesting is the inclination to defer to these sources even when the information should be known, for based on the principle of efficiency, retrieving information from one’s own memory should be quicker. It might be the case that individuals turn to the Internet to not only retrieve information but also verify it, for human semantic memory could be considered fallible in a way that a database is not, irrespective of other doubts about the veracity of online knowledge.

In this aspect of ‘retrieval coordination’, the Google Effect could be regarded as a mainly conscious process in which the individual defers to a database because it has a better directory system and a much larger store of knowledge. However, the study also examined how the mere presence of a computer can affect how the participants memorise – the process of ‘information allocation’ in a transactive memory system. Sparrow et al. conducted tests that focused on the degree to which someone would remember something and found that the participants were less likely to remember something when they knew that it could be accessed at a later date but were more likely to remember something if they thought that it would be erased from the computer’s memory (2011: 777). From the study, it is not evident to what degree the participants were conscious of this shift in the way they remember, for there is a significant difference between consciously deciding not to remember something and automatically not remembering – the latter would more likely fall into the province of procedural memory. Although the participants were less likely to remember the information, they still acquired an understanding of where it could be located if needed – in other words, they still developed labels in a directory (Sparrow et al., 2011: 778). What this suggests is that ready access to computer databases and search engines entails a vast expansion in accessible knowledge that is translated into knowing ‘where information is to be found rather than the information itself’ (Sparrow et al., 2011: 776), which leads to greater dependence on computers and remaining ‘plugged in’ (p. 778). This delegation of memory to external devices can expand and supplement an individual’s knowledge by freeing up memory for other tasks. However, it is important to note that the labels indicating where information can be found are increasingly produced by our devices rather than through a conscious process of directory organisation in which we have control on the epistemological framework. Writing notes, collecting information or creating folders to store documents is significantly different to automated processes of file management favoured by social media sites, search engines and content streaming services such as Spotify. The more directories are produced for us, the more we rely on the structure of labelling built into the networks without a corresponding understanding of the content.

Voice assistants, networked devices and the extended mind

The idea that the incorporation of external memory into human cognition can diminish cognition and create over-reliance on external knowledge, that is, the Google Effect, is critiqued by Richard Heersmink, who argues that Sparrow et al. do not experiment with people actually using Google and instead derive their claims from quite limited laboratory experiments (2016: 394). Moreover, the emphasis on trivia and general knowledge is not representative of the diversity of ways we use the Internet, from online maps to calendars that are incorporated into our daily activities (p. 395). Rather than focusing on the negative effects of the regular use of search technologies, Heersmink also argues that the research should examine how the Internet frees up our limited memory to allow us to better attend to other tasks (2016: 396). One of the problems is that research into the Google Effect has so far only examined semantic memory and does not attend to both procedural memory, which is particularly important in the use of applications, and episodic memory (Heersmink, 2016: 397). For example, using an online calendar with an automated set of reminders will certainly relieve a cognitive burden, and does not automatically entail a diminution of cognition or the capacity to remember. In the degree to which it coincides with everyday practice, which is certainly achievable with a smartphone that remains with the person, the more it disappears into procedural memory. This semi-conscious, habitual mode of remembering becomes a background to our activity, freeing up more space and time to attend to other matters at the level of both semantic and episodic memory. This critique is certainly valid, but it does not attend sufficiently to the Google Effect’s role in a broader transactive memory system, including the role of expertise in the formation of semantic memory and how personally developing a directory system (epistemological framework) differs from relying on sets of pre-existing labels.

Using external forms of storage within any system of shared memory could lead to a situation in which we delegate much of our propositional knowledge, usually the preserve of semantic memory, to external devices. This practice is not entirely new, for individual memory has long been supplemented externally through various forms of writing and inscription and it could just as easily be argued that we overly depend on dictionaries and books. If we know that the meaning of a word will be available in a dictionary, one could assume that we would be less inclined to remember the actual definition or indeed the word’s etymology. In general, writing removes the need to remember everything required in a cognitive task and in doing so expands our cognitive range, and the same applies to any system that allows both the storage and recall of information. Andy Clark and David Chalmers argue in their ‘extended mind’ hypothesis that external devices, for example, notebooks and computers, extend the mind beyond the confines of consciousness and the brain and, in doing so, increase the range of thinking, memorising and cognition (1998: 7). The mind forms a ‘coupled system’, which is most evident in the use of language:

Think of a group of people brainstorming around a table, or a philosopher who thinks best by writing, developing her ideas as she goes. It may be that language evolved, in part, to enable such extensions of our cognitive resources within actively coupled systems. (Clark and Chalmers, 1998: 11–12)

In a later article, Clark argues that the mind will be further extended through the use of digital technologies that are integrated into ‘cognitive and physical problem-solving routines’ and refers to Alex Pentland’s work on ‘memory glasses’, which insert subliminal clues to help the user remember faces (2005: 9). Indeed, since the beginning of the millennium, Pentland from the MIT media laboratory has advocated for devices that respond directly to the behaviours and thoughts of users due to their ‘perceptual intelligence, and capacity to learn independently’ (2000: 35–36). He envisioned ‘smart rooms’ that would perform a function similar to a butler, remaining invisible but always ready to assist, and ‘smart clothes’, which due to their portability operate like ‘personal assistants’ (p. 37). In general, Pentland and his researchers ‘imagine building a world where the distinction between inanimate and animate objects begins to blur, and the objects that surround us become more like helpful assistants or playful pets than insensible tools’ (2000: 44). The implication is that such devices should seamlessly assist us in achieving our tasks, or simply act as company or support, without in any way detracting from cognitive functioning. This vision has recently found commercial success in voice assistants, from Siri and Google Assistant to Amazon’s Alexa, which, unlike early smart rooms and clothes, are able to draw upon the vast wealth of data on the Internet to better support the user. The new version of Alexa employs a system called ‘Hunches’ that uses machine learning and neural networks to ‘intuit’ a user’s needs, even recognising when the user has forgotten something. It correlates information gleaned from the user’s Alexa-enabled devices with publically available information such as timetables, clocks and weather patterns to develop an understanding of human habits. In addition, to reminders, Alexa will also develop an understanding of general preferences and suggest things such as shopping items or playlists to accommodate potential needs and wants (Harris, 2018). The system is also strongly adaptive, for it will learn the best way to respond through the analysis of repeated speech, or variations in the nature of the request (Biggs, 2019). Like Pentland’s smart devices, the claim is that Alexa will operate in the background like a personal assistant, helping to free up mental space to attend to other activities. The Vice President of Amazon, Rohit Prasad, expresses the company’s aim of creating voice assistants that are ‘invisible’ and, therefore, do not produce the same ‘cognitive overload’ as a smartphone. Alexa could ‘whisper to you if she thinks you’re trying to be quiet, or ask you about a light you left on if it has a hunch you did so unintentionally’ (Biggs, 2019). The idea is that we need to be emancipated from remembering mundane procedural tasks in this push towards forms of networked computing that adapt to the individual and are seamlessly integrated into quotidian life. Having all information ready to hand, even before we know we need it, also removes the burden of negotiating with others, including forms of external memory, to acquire information within a transactive memory system.

Internet search, generality and differentiated expertise

The idea of the neutral voice assistant that will assist us in remembering, while removing a range of other cognitive distractions, should certainly be welcomed, and there is no doubt that technology has played a significant role in broadening the reach of cognition. The question is not strictly one of arguing for and against any new advance but properly understanding how the technologies actually shape cognition and contribute to the sharing of knowledge. In terms of a transactive memory system, the main issues concern the role of networked technologies and search engines in coordinating external and internal information allocation, the speed and ease of information recall and the role of expertise. When considering devices that are directly connected to databases, the speed and range of access to information can affect how expertise manifests. One of the characteristics of an expert is their awareness of where information can be found – in books, lists, indexes and catalogues – as well as an understanding of the particular practices required for its retrieval. In contrast, networked devices with search capability offer a means of retrieving information without specific knowledge of a domain or consideration of who else might have access to this information.

When smartphones or smart speakers are integrated into a transactive memory system, they do not necessarily assist in the development of structured knowledge for a search query box or voice prompt such as ‘Okay Google’ is a universal label that can be used without understanding the broader directory system. Hillis et al. (2013) argue the Google Search operates at a level of generality in which information is ‘decontextualized’ and only ordered according to the popularity algorithms of PageRank and its indexes (p. 73). Furthermore, Google’s positivist methodology and aim of instantly retrieving search results does not provide the necessary distance central to the cultivation of ‘reflexive knowledge’ (2013: 74). In a human network, distance is always ensured by the fact that we must negotiate with others to gain access to external memory and, through negotiation, develop an understanding of epistemological contexts and the role of expertise. Using a search engine such as Google or requesting information through a voice assistant, due to the fact knowledge appears in response to a general query, does not provide sufficient understanding of either the effort required to develop knowledge or the particular human contexts in which it is maintained. In a transactive memory system, each individual has a specific area of knowledge and a form of directory knowledge by which they request information from others in the network. If a networked device is introduced into this system, it can greatly expand the range of expertise through the direct use of a search engine or indirectly through its own machine learning procedures that draw on external databases. In a human transactive memory system, a label is affixed to specific absent knowledge – we do not know the detailed information but know where it can be located. The use of search engines reshapes the transactive memory dynamic – where once numerous labels were established to denote types of knowledge associated with transactive memory participants, now only one is required, the search query.

In addition to its generality, the Internet can also appear infallible because it is extremely likely that one of the search results will be correct, in which case, our role is to choose from a set of options offered by the search-enabled device rather than come to an understanding through reasoning and negotiation. Adrian Ward states that the Internet is not a normal ‘memory partner’ in which there is a proportionate distribution of knowledge and certain level of fallibility, for it appears to surpass all forms of expertise: ‘Accessing the Internet can be like tapping into a field of actual experts, as opposed to simply asking the individual in one’s transactive memory structure that has the highest level of relative expertise’ (2013: 343). Having liberal access to information on the Internet would remove the burden of remembering because information that can always be retrieved later is largely ‘redundant’. It would also remove the need to develop distributed transactive memory systems in which information is shared among a group, in a way that ensures the functioning of that group (Ward, 2013: 343). There is no longer equal participation in the sharing of memory for the search engine supplants the specific memory of others in the group due to its generality, but also claims greater precision, for digitally inscribed information can be recalled without loss. Why rely on vague internal memories, or the recollection of other members of the group, when automated processes of information storage are presumed to be infallible, even if what is stored or remembered is inaccurate or untrue. The use of GPS presents an example of this. Greg Milner (2016) describes the example of ‘death by GPS’, a term coined by US park rangers who noticed a tendency of drivers to rely on an ‘uncritical acceptance of turn-by-turn commands’ even when doing so takes the driver on routes that lead into the ocean, over cliffs or across rarely traversed desert paths. Nicholas Carr (2015) refers to this practice of placing too much emphasis on the veracity of automatically generated information as ‘automation bias’, in which a human user’s ‘trust in the software becomes so strong that they ignore or discount other sources of information, including their own senses’ (p. 69). Although we have mainly been discussing semantic memory, such an example demonstrates that memory borders on procedural memory depending on when and how it is interpolated in consciousness. We expect the automated information source to be accurate and rather than addressing each unit of information in turn, it becomes something that directs our attention. In accepting the facility of access to information – asking Google, talking to Alexa, following a GPS mapping device – we do not sufficiently attend to how knowledge is posited and used. It will certainly foster dependence on devices that provide ready access to these databases but could also lead to a diminution in understanding of the intrinsic structure of types of knowledge – including the spatial organisation of a national park.

Speed, personalisation and memorisation

The speed with which information is provided by networked devices and the increasing application of voice assistants raises further questions about how external memory operates that have not yet been addressed in Wegner’s theory. In many ways, transactive memory focuses on an individual’s awareness of the location of specific knowledge within a group or network at a conscious or semi-conscious level. An individual knows that they can delegate the task of remembering to their devices, or more generally to the Internet, which will continually and automatically store information that might otherwise have to be retained in human semantic memory. In some respects, we make a choice when and what to remember knowing that a computer, or networked device, could be at hand and used to recall the memory. However, this still remains a relatively slow procedure that somewhat resembles asking an individual within a transactive memory network for information or advice. If there is an increase in the speed of information recall or retrieval, this can affect the degree to which the device, with its inbuilt mechanisms of recall, can be differentiated from human cognition and memory. René König and Miriam Rasch (2014) state that our familiarity with search mechanisms, especially now that they are integrated into a range of devices, has rendered them invisible (p. 10). ‘Within a remarkably short time range we have familiarized ourselves with the search logic: type, select, click, and move on. The ever-increasing speed we use to search has created a collective “techno-unconsciousness” from which we have to wake up’ (König and Rasch, 2014: 10). One of the main problems with search technology, especially due to the minimal interface and the increase in personalisation and prediction, is that it does not reveal how it operates, and is effectively a ‘black box’ technology (König and Rasch, 2014: 11). Clark suggests that our networked devices are like prostheses that cannot be clearly distinguished from the individual when engaged in action, but, unlike König and Rasch, he argues that this is not necessarily a problem for our neurological structures will adapt to incorporate whatever tools we are using (2005: 8). Likewise, Katherine Hayles argues that such epigenetic adaptations might just create new ways of thinking that may operate on a nonconscious level, which is entirely appropriate as nonconscious thinking allows us to deal with a much larger amount of information than we can in conscious attention (Hayles, 2012: 94–95). This integration of devices into cognitive behaviour certainly could be developed to support conscious awareness, but what is at stake is deciding which aspects of our networked technology best do this, and to what degree do we want the technologies to supplant or support conscious decision-making?

The problem with engaging with technologies that operate just below or at the threshold of consciousness is that we tend to believe or accept judgements and information that come most easily to us, that is, judgements that do not require too much deliberation. Daniel Kahneman in the popular Thinking, Fast and Slow argues that much of our reasoning is based on cognitive ease, in which familiarity and the relative ease with which we come to an argument – based on a primary level of subconscious reasoning – means that it is more likely that we will accept its premises (Kahneman, 2012: 62). The speed and habitual use of networked devices that complement thinking could lead to greater reliance of such devices and a possible diminution in higher-order reasoning. A small study by Nathaniel Barr and others proposes that the use of search on smartphones increases the degree to which users operate as ‘cognitive misers’, a term derived from Kahneman to refer to a common tendency to use lower levels of thinking and intuitive responses if they are available rather than drawing upon deeper analytical processes of cognition that require a greater use of memory (Barr et al., 2015: 474). They found there was a significant increase in ‘cognitive miserliness’ with the high use of search engines on smartphones, which led the authors to argue that cognitive miserliness can be linked to ‘a reliance on outside informational sources’. The heavy users would prefer to turn to the phone rather than think analytically through a task, and this is linked to a more intuitive cognitive style (Barr et al., 2015: 478). This willingness to abrogate intellectual effort to mobile devices might only be a minor effect of the conjunction of search with smartphones, which could be extrapolated to voice assistants, however, it supports the general claim of transactive memory theory; that the increase in the range of available information does not necessarily lead to an improvement in cognition. Finding a ready solution through search for the cognitive miser could be just another means of substituting the directory for detailed information and the logical structures implicit in this information. This can be correlated with a study by Nestojko and Roediger (2013) arguing that problem-solving and flexible thinking are enhanced by the capacity to recall information stored in one’s own memory, which is better adapted for use in higher-level cognitive processes than externally stored information (p. 323). The issue, of course, relates to the familiarity we have with the remembered information and the degree to which it has already been consciously negotiated before it is memorised. Networked memory could aid our reasoning if it complements what we already know, rather than substituting for higher-level cognitive practices.

How we evaluate the technology depends on what role memory plays in cognition, is it something that is recalled to support a particular decision or is it something that pre-empts the process of decision-making. Of course, it performs both functions, for procedural memory underpins cognition through creating habits of thought, including habits of logical thought, and allowing us to work through a sequence of steps in an argument. What the current range of devices hope to do is recall or provide information within the time of our action, in which case, it operates within the temporality of procedural memory – it appears before we have decided how to use and evaluate the information. In this case, memory is recalled to anticipate action and judgement rather than to be interrogated within a conscious deliberative process. There are two main facets to this shift in the use of networked aide-memoires, one is the disappearance of the screen interface and the other is the development of more extensive pre-emptive technologies.

The screen and keyboard should be removed because they slow down and disrupt fluid interaction with the technology, as these interfaces work for us rather than seamlessly with us. Indeed, research points to a notable increase in the use of voice-enabled voice assistants with predictions of 8 billion voice assistants in use by 2023 (Juniper Research, 2018). The arrival of these voice assistants creates a new type of interface that in many ways resembles a person in a transactive memory network, although one that recedes into the background as a virtual butler or assistant. This idea of the invisible companion has long been touted, from the early days of Google Glasses (cf. Pentland), but has become more realisable due to the rise in the Internet of Things and the preponderance of mobile devices. Victor Luckerson (2016) explains how Google search has changed, and is set to develop further, to fully integrate with evolving devices. One important aspect is the move from the keyboard bar to voice activated applications that are embedded in mobile devices and smart speakers such as Google Home and wearables such as Apple’s AirPods. This is all part of a push to develop applications and devices that respond to requests within the normal time of speaking and acting. In this emphasis on continual voice interaction, Luckerson (2016) argues that the user interacts with Google in ‘a kind of ongoing conversation with an omniscient narrator, ready to step in and fulfil any request – even ones you haven’t thought about yet’. Usually we can think of speaking in terms of the request, which is certainly a key function of most current voice assistants which are still hailed or asked questions in ways that resemble existing search queries. However, the more that the devices can passively acquire information – information that is not consciously inputted into the device – the more they can respond to unstated requests, such as reminding the user that they may have left their keys behind.

Even without the emphasis on voice activation, search engines have been looking at ways of pre-empting the user’s queries through personalisation algorithms that reflect individual patterns of computer use. These personalisation algorithms speed up the process of recall because the user does not have to consider the conditions under which something is retrieved, as it is already tailored to their preferences. In the early form of the technology, such as the development of Google Suggest in 2004, the user was given suggestions for possible search options that also give some indication of what other people are searching for (Gibbs, 2004). Later, Google proposed Google Instant which produces results as the searcher types. According to Marissa Mayer (2010) writing in Google’s official blog, Google Instant is only a precursor to fully pre-emptive technology, for what ‘you really want search-before-you-type – that is, you want results for the most likely search given what you have already typed’. This speed is achieved through the refinement of the ‘relevance’ of the search parameters such that Google will eventually predict a user’s intention without a clumsy interface and almost before the user knows what they want (Hillis et al., 2013: 56). Google has long been interested in ‘instantaneity’ in the development of its search engines and the general belief that eventually individuals will not have to even formulate search queries, which should significantly reduce the time of searching (Hillis et al., 2013: 64).

One of the main issues with pre-emption is that the technology recommends how we behave or should behave. Kylie Jarrett (2014: 23–24) states that Google’s suggestions can shape what the searcher is looking for and therefore operate as a feedback loop, which will suggest and confirm the system’s predictions to create a technological normative:

The intention we articulate in search is thus previously shaped by the intentions we, and others, have already articulated into the database. In this recursive logic, the potential of futurity becomes limited by past resolutions. Embedding the output of search into the logics of search in this way can be understood as a form of control. (Jarrett, 2014: 24)

Although pre-emptive technologies may support higher levels of cognitive functioning, we should not assume that networked devices and Internet search necessarily seek to cultivate such thinking. The fact that a range of devices and apps are looking to understand human moods further demonstrates this. Simon Fussell (2018) states that Amazon has placed a patent that will eventually allow Alexa to recognise different emotions, Google has applied for a patent that specifically reads negative emotions and even IBM is working on mood recognition to assist search queries. Fussell argues that the overall aim of such AI is to target advertisements to users, which is already manifest in Spotify, which builds mood profiles based on the playlist, accompanied by advertisements that suit those moods. Mark Andrejevic (2013) argues that research in areas such as neuromarketing hopes to exploit those types of decision-making governed by emotion rather than reason. Feeling is also a way of dealing with large amounts of information by providing a form of ‘shorthand’ response invested in the body and characterised by thoughtlessness (pp. 102–103). To assist in decision-making, which would include mood recognition algorithms, is not the same as cultivating the capacity to make decisions. It just means that many decisions will remain unconscious and not subject to the transparency of negotiation and reasoning.

Conclusion

The theory of transactive memory provides a foundation for rethinking how devices and specific applications, such as voice assistants and Google search, supplement our memory but also create the conditions for bypassing or supplanting it. It is not just about the augmentation of memory or the extension of the mind but how cognition is changed with regard to how we access and record memories. Transactive memory has been applied to computational technologies, but does not – as it currently stands – sufficiently consider the specific conditions under which information is recalled or stored. In a transactive memory system, memory is consciously managed in terms of the specialisations and interests of others in the group. We are constantly required to think about how we relate to the members of a group in terms of our own specialisation. Generalised search and other such applications reduce some of the conscious aspects of transactive memory including the formation of individual and group directories, recall of information through negotiation with others and a broader awareness of how knowledge is structured. With the arrival of voice assistants and the refinement of pre-emption in search applications, the spatial model for organising knowledge in a transactive memory system somewhat gives way to procedural memory that conditions behaviour and has the capacity to change how we think, reason and create. On the threshold between semantic and procedural memory, how and when information is interpolated into human thinking is as important as the type of information we receive. The theory of transactive memory is a first step in understanding this shift, in presenting a model of recall in which what we remember is contingent on what others in the network are doing. It is a model of relative expertise and knowledge, which has begun to take into consideration the role of large information sources that usurp human expertise; however, what should be considered now is the role of devices that pre-empt human action and also pre-empt human recall. The increased automation of mobile devices in a transactive memory system simultaneously opens and contracts access to information in a way that extends beyond the boundaries of awareness.