Knowledge machines

This work is supported under SOCIAM: The Theory and Practice of Social Machines. The SOCIAM Project is funded by the UK Engineering and Physical Sciences Research Council (EPSRC) under grant number EP/J017728/1 and comprises the Universities of Southampton, Oxford and Edinburgh.

The mechanistic view and the content creation view are not the only views of social machines to have been discussed within the philosophical and scientific literature. Palermos (2017) for example, countenances a cognitive systems view of social machines, which casts social machines as distributed cognitive systems—that is, as socio-technical systems that perform cognitive tasks. A not altogether unrelated view is proposed by Hooper et al. (2016). They suggest that social machines should be regarded as problem-solving organizations (this is what we might call the problem-solving view). Finally, there are a number of views that highlight the role of the human social environment in shaping the capabilities and performances of AI systems. These include the sociable machines view (Hendler & Mulvehill., 2016) and the socially situated machines view (Smart & Madaan., 2017).

More precisely, the phenomena of interest are what Kaiser and Krickel (2017) refer to as object-involving occurrents. These are said to consist of ‘an object (or system) that is engaged in a certain occurrent’ (Kaiser & Krickel., 2017: 24), where the term ‘occurrent’ is simply a shorthand way of referring to events, states, and processes.

Given the status of social machines as online socio-technical systems, Internet communication protocols are likely to play a crucial role in supporting the flow of information between the components of a socio-technical mechanism.

Note that the status of some system as a social machine is likely to require some form of empirical analysis. This is because the status of some system as a social machine depends on the mechanisms that are deemed to be responsible for system-level phenomena. For the most part, such mechanisms will be identified in the same way that mechanisms are identified in other scientific disciplines, for example, via the use of observational, experimental, and computer simulation techniques. Such forms of analysis may not be required in all cases, however. This is because the science of social machines is, at least in part, an engineering discipline and socio-technical mechanisms may be built from the ground up to realize a particular function. In this case, the material constituents of the mechanism will be relatively obvious, at least to those who are involved in the effort to build social machines.

Historical precursors to the knowledge machine concept can also be found in literature of a more technologically oriented nature. Writing in 1992, for example, David Gelernter presents the notion of a knowledge plant, which is a computational system designed to extract benefit from network-mediated data streams. ‘Instead of venting…data into the info-smog’, Gelernter writes, ‘we could treat our data sources as plunging waterfalls waiting to drive software powerplants that convert data into knowledge’ (1992: 112). This quotation helps to reveal a crucial difference between Gelernter’s notion of a knowledge plant and the concept of a knowledge machine. A knowledge plant is, as Gelernter himself notes, a system that is ‘realized by a software ensemble’ (1992: 112). This contrasts with a knowledge machine, which is a system whose epistemic processes are realized by a socio-technical mechanism.

This is particularly evident when it comes to systems that aim to generate epistemic resources (e.g. computational ontologies) for the Semantic Web.

The idea of a distributed approach to knowledge engineering is based on the notion of distributed cognition, as discussed in the cognitive science literature (Hutchins, 1995). The core idea is that socio-technical systems are able to implement some of the activities that we typically associate with knowledge engineering, for example, the attempt to elicit, acquire, model, and exploit human knowledge.

As noted by an anonymous reviewer, another potential omission relates to systems that support collective forms of debate, reason, and argumentation (e.g. http://www.debatepedia.org/). Although such systems are commonly devoted to the expression and reconciliation of conflicting opinions, their outputs may, on occasion, qualify as epistemic outputs. In such cases, there seems little reason to discount the status of these systems as fully paid up members of the class of social machines.

Watson and Floridi (2018) go on to note the virtues of the online environment with regard to issues of social scale. ‘Only online platforms’, they suggest, ‘offer the kind of scalability required to host hundreds of thousands of volunteers for any given project, and only at these volumes does the data processing power of untrained amateurs begin to compete with (or exceed) that of experts using traditional observation methods’ (Watson & Floridi, 2018: 753).

For example, eBird preserves the provenance of user contributions, enabling specific users to receive public recognition for important observations (e.g. the first sighting of a bird species in a particular geographic area).

It is estimated that there are hundreds of millions of gamers worldwide who collectively spend more than 3 billion hours per week playing video games (see McGonigal, 2011).

Despite the fact that citizen science systems and GWAPs are discussed separately in the present paper, there is nothing to prevent GWAPs (of either the goal-transparent or goal-opaque varieties) functioning as citizen science systems. In other words, there is no reason to assume that GWAPs and citizen science systems are disjoint categories of knowledge machine. There are clearly many cases where a game-oriented system can be used for serious scientific purposes, and, conversely, many forms of intellectual endeavour may benefit from the inclusion of ludic elements (e.g. Laszlo, 2004).

Neither is the remit of GWAPs necessarily restricted to the scientific domain. One area of recent attention in the knowledge engineering community is the use of GWAPs to support the development of Semantic Web resources (Siorpaes & Hepp, 2008; Simperl et al., 2013).

See http://www.seaheroquest.com/en/

The social status of these systems is not something that should be in doubt. For even GWAPs that fail to support direct player-to-player interactions still require a large number of participants in order to fulfil their epistemic purpose. In other words, important forms of knowledge discovery are often predicated on the contributions of multiple individuals. We see evidence of this in both the Genes in Space and Sea Hero Quest games. In the case of Genes In Space, multiple (independent) contributions are required in order to ensure the reliability of analytic outcomes. According to comments posted on the Cancer Research UK website, for instance, Genes In Space has been used to analyze the ‘entire genomes of 1980 patients, each checked 50 times for accuracy’ (see http://www.cancerresearchuk.org/support-us/citizen-science/the-projects#citizenscience1). A somewhat different role for multiple human contributions is apparent in the case of Sea Hero Quest. Here ‘social’ participation is a prerequisite for the success of the larger scientific effort, that is, the assembly of a normative dataset for the purposes of diagnostic testing.

Such claims resonate with the idea that social machines serve as part of the realization base for social phenomena. In this respect, work into what are called Web Observatories is of particular interest, especially since such efforts often seek to observe and monitor the behaviour of social machines on the Web (Tiropanis et al., 2013; Tinati et al., 2015).

This does not mean that DBpedia should, itself, be viewed as a knowledge machine. What is crucial to the status of some system as a knowledge machine, at least from the standpoint of the mechanistic view, is the idea of some knowledge-relevant process being realized by a socio-technical mechanism. The output of such processes will typically be some form of epistemic product (e.g. a computational ontology), but this does not make the product, itself, a knowledge machine.

The notion of a ‘local environment’ is typically understood in terms of spatial criteria. In the case of the Web, however, the local environment means the set of online resources (e.g. Wikipedia articles) that are accessed by multiple individuals.

Such claims establish a useful point of contact with work in the cognitive sciences, especially work that highlights the role of just-in-time action as a means of exploiting the extra-organismic environment for cognitively relevant purposes (Clark, 2008; Myin & O’Regan, 2009).

According to Goldman, epistemic systems are ‘social system[s] that [house] social practices, procedures, institutions, and/or patterns of interpersonal influence that affect the epistemic outcomes of its members’ (2011: 18). This is a concept that is broadly compatible with the idea of a knowledge machine. In particular, Goldman sees the epistemic standing of a social system as tied to the operation of one or more social mechanisms that are housed within the system. This much is clear from the emphasis that Goldman places on the role of organizational structure and patterns of inter-agent communication in the generation of epistemic outcomes.

See Weld et al. (2015) and Steyvers and Miller (2015) for a useful overview of some of the mechanisms that can be used to enhance the reliability of socio-computational systems. Watson and Floridi (2018) also provide a useful overview of some of the quality control procedures adopted by the Galaxy Zoo system.

It should be noted that there is a potentially important parallel here with the notion of ecological assembly in the cognitive sciences. The focus in a cognitive scientific context is typically on the mechanisms that enable a particular cognitive agent to select and assemble a set of extra-organismic resources into some larger problem-solving whole. Clark (2008) provides a useful characterization of the idea in the form of the Principle of Ecological Assembly. According to this principle, ‘the canny cognizer tends to recruit, on the spot, whatever mix of problem-solving resources will yield an acceptable result with a minimum of effort’ (Clark, 2008: 13).

Palermos (2017) raises a similar point in respect of the Wikipedia system. In this case, Palermos suggests that the relevant form of reliability is ‘an emergent, distributed property that belongs to Wikipedia as an overall, integrated system and cannot be accounted for in terms of the reliability of its underlying mechanistic and organismic components. Its reliability arises, instead, out of the actual and potential synergetic interactions of all components—organismic and mechanistic ones alike—at the same time’ (2017: 972–973).

This highlights the way in which one kind of mechanism (i.e. a socio-technical mechanism) can assist with the epistemic analysis of other mechanisms (e.g. the mechanisms responsible for species distribution). There is, of course, no reason why a socio-technical mechanism cannot be used to further our understanding of other socio-technical mechanisms. Indeed, one of the objectives of a knowledge machine may very well be to improve our understanding of the mechanisms that realize the epistemic performances of other knowledge machines.

See https://fsapps.nwcg.gov/afm/