The critical study of technology has produced a substantial body of work that addressed broad issues of educational policy, as well as the ways in which digital cultures influence knowledge production within and beyond the classroom (Buchholz et al 2014; Burnett 2016; Potter and McDougall 2017; Williamson 2016).  This includes recent work highlighting changing forms of digitally-mediated education governance (Landri 2018; Piattoeva 2016; Selwyn 2016; Souto-Otero and Beneito-Montagut 2016), and the ‘performative’ role of computational methods in creating their objects of interest (Perrotta and Williamson 2016). This paper extends these approaches to a new phenomenon in education: the appearance of general-purpose AI infrastructures, accompanied by a growing interest among higher education institutions for predictive modelling. Our paper is written to directly respond to the NW28 call for studies of 'datafication, and the algorithmic governance of education'.

While much in the field of ‘AIEd’ remains speculative and over-hyped, this paper focuses on one tangible, established application of computational methods to automate specific tasks of classification and prediction – the field of ‘machine learning’. Machine learning is concerned with the identification of patterns in datasets and the computation of predictions and classifications when new instances of the same data are encountered. This technique now drives developments in many area of society – such as online shopping, face recognition, self-driving cars and cancer diagnosis. In education, specific attention is being paid to the use of machine learning to model – and then guide - aspects of performance and behaviour in educational settings.  

This paper explores this latter educational application of AI in the context of European higher education. Theoretically, we position our work in the fields of digital sociology and Science and Technology Studies.  Specifically, we draw on concepts and approaches that are being developed by studying the economic and social ramifications of applied AI. Notable contributions to this discussion include Mackenzie’s work on machine learning and Crawford and Joler’s analysis of Alexa, Amazon’s proprietary AI system (Crawford and Joler 2018; Mackenzie 2017).

From this perspective, the present paper presents preliminary findings from an ongoing study of the application of AI infrastructures in higher education ‘learning management systems’. Our overarching argument is that the convenience and accessibility of ‘off-the-shelf’ AI frameworks is particularly problematic in education in the ways in which the underlying logics of these technologies (and the datasets for training them) are designed to be controlled by a few entities (Crawford and Joler, ibid). In one sense, then, the application of AI in universities’ ‘learning management systems’ signals a trend where technology corporations hold sway over forms of expertise and resourcing that remain opaque and intractable for most non-specialist actors. While this is a familiar argument in the critical commentary on data and algorithms, AI adds another layer of complexity as its operational aspects are informed by very specific principles of cognition and stochastic reasoning that shape understandings of what ‘learning’ and ‘knowledge’ are.

These contentions are developed through the exploration of the following research questions:

1. How is predictive AI being assembled as a sociotechnical phenomenon in education?
2. What relationships and tensions are emerging between computational methods, datasets, software/hardware resources, regulatory frameworks, human agency and educational cultures?  
3. What are the current (and future) impacts of predictive infrastructures on educational governance in Europe and internationally?

The paper addresses these questions through the critical examination of one case of an applied AI framework (Google’s ‘Tensorflow’) and its integration into the widely used ‘Moodle’ learning management system. In brief, TensorFlow is a library for deep learning computations written in the software language Python. It is designed to provide a ready-made, scalable toolkit to develop predictive models that can be applied to any kind of digital data. In this paper, we trace how Moodle – in its guise as the producer of an open digital platform - has made formal and informal efforts to explore the integration of TensorFlow in the platform’s ‘learning analytics’ system. This integration will enable universities to train and deploy their own predictive models focusing on various aspects of educational performance and behaviour, recorded as digital data while students interact daily with the LMS. Our method can be described as a digital ethnography based on two research approaches. First is a series of in-depth interviews with key actors involved in the development of Moodle’s Learning Analytics Application Programming Interface (API). For our interviews, we targeted a small number of strategic informants working in various technical and educational roles at the interface between Moodle and one specific large international university. These include: • The data scientist who laid the groundwork for the Moodle learning analytics API and the TensorFlow integration; • The main educational researcher acting as an intermediary between the computer scientists and the educational communities in Moodle as well as in partner institutions. • Moodle’s Business Development Manager who coordinates the links with senior administrative staff in universities; • four individuals working in leadership and operational capacities in the ‘Education Innovation’ centre in a large international university. In addition to this interview data - we have also collected and analysed multiple ‘digital traces’ of the Moodle-Tensorflow relationship. These traces are discoverable on the internet due to the open-source nature of both Moodle and TensorFlow. The traces analysed for this paper comprise sections of digital code, data architecture diagrams and technical specification documents. This combined human/code methodology allows us to attend, simultaneously, to the technical and the social dimensions of how the tenets of predictive AI were ‘translated’, to use STS terminology, into the education contexts of Moodle and one specific university.

Our case study sheds light on where the impetus for adopting machine learning into education is originating, which actors are involved in supporting this agenda, the technical nature of these processes, and the institutional consequences of what is finally produced. By attending to this trajectory we raise three broader observations about the fast-changing relationship between algorithmic prediction and institutional governance: a) In a marketised HE sector, the interest of university leaders in predictive analytics appears less fuelled by a desire to improve pedagogy than by the promise of control over key performance indicators: throughput, dropout, satisfaction, failure and so forth. b) Our ethnographic engagement suggests that the way in which a particular instance of AIEd ‘comes together’ is a nuanced sociotechnical process where it is possible to detect conflicting views of education, technology and agency. Not all of these views are aligned with reductive notions of surveillance and metrification. c) As we analyse the emergence of these systems, striving to document their nuances, we must keep in mind that, for the time being, they remain marginal. There is a tendency for digital technology in education to occupy a symbolic space, where pockets of innovation fuel ‘policy imaginaries’ that have little bearing on the daily business of education. AI’s connotations as the latest ‘hype’ in education is a reflection of an established historical pattern. As a conclusion, we raise the challenge of how we can ‘think otherwise’. As education systems move towards the development of predictive infrastructures, can sociological research do more than shine a light on the uncertainty and sociality that pervade these systems? How can critical enquiry such as ours suggest ways to open up the ‘infrastructural imagination’ (Gray at al 2018) about how higher education in the digital age may be organised differently, or not organised at all?

Buchholz, B., Shively, K., Peppler, K. and Wohlwend, K., 2014. Hands on, hands off: Gendered access in crafting and electronics practices. Mind, Culture, and Activity, 21(4), pp.278-297. Burnett, C., 2016. Being together in classrooms at the interface of the physical and virtual: implications for collaboration in on/off-screen sites. Learning, Media and Technology, 41(4), pp.566-589. Crawford, K. and Joler, V., 2018. Anatomy of an AI System: The Amazon Echo As An Anatomical Map of Human Labor, Data and Planetary Resources. AI Now Institute and Share Lab. Available here: https://anatomyof.ai/ Gray, J., Gerlitz, C. and Bounegru, L., 2018. Data infrastructure literacy. Big Data & Society, 5(2), p.2053951718786316. Landri, P., 2018. Digital Governance of Education: Technology, Standards and Europeanization of Education. Bloomsbury Publishing. Mackenzie, A., 2017. Machine learners: archaeology of a data practice. MIT press. Perrotta, C. and Williamson, B., 2018. The social life of Learning Analytics: cluster analysis and the ‘performance’ of algorithmic education. Learning, Media and Technology, 43(1), pp.3-16. Piattoeva, N., 2016. The imperative to protect data and the rise of surveillance cameras in administering national testing in Russia. European Educational Research Journal, 15(1), pp.82-98. Potter, J. and McDougall, J., 2017. Digital media, culture and education: theorising third space literacies. Springer. Souto-Otero, M. and Beneito-Montagut, R., 2016. From governing through data to governmentality through data: Artefacts, strategies and the digital turn. European Educational Research Journal, 15(1), pp.14-33. Selwyn, N., 2016. ‘There’s so much data’: Exploring the realities of data-based school governance. European Educational Research Journal, 15(1), pp.54-68. Williamson, B., 2016. Digital education governance: An introduction. European Educational Research, Journal. Vol 15: 1.