Programmed welfare: An ethnographic account of algorithmic practices in the public distribution system in India

Introduction

Algorithms have been traditionally perceived to be abstract and fixed sequence of computational procedures, a logic that Seaver (2014) argues relies on a rational pairing of certainties of mathematics with the objectivity of technology. Treating the computational properties of the algorithm as stable and fixed gives the illusion that the outcome can therefore be audited for biases, accountability and transparency (Lee, 2020). However, as many scholars in critical algorithm studies would contend, the actual nature of an algorithm can only be understood in its performative agency enacted within a network of human and non-human actors (Eubanks, 2018; Kotliar, 2020; Lee et al., 2019; Seaver, 2014; Ziewitz, 2017). This perspective of the algorithm as assemblage (e.g. Ananny, 2016; Lee, 2020) or as epistemic amalgam (Kotliar, 2020) implies that it is important to understand the types of alliances algorithms form with different actors that span the spectrum of agency, autonomy and opacity associated with an algorithm (Bechmann and Bowker, 2019; Kotliar, 2020; Lee, 2020). This by now common perspective among critical algorithm scholars brings focus to what algorithms do, rather than what they are, and how we might explore their outcomes as situated and emergent in practice (Dourish, 2016; Eubanks, 2018; Noble, 2018; O’Neil, 2016; Ziewitz, 2017). Hence, to debunk the real-life working of algorithmic governance, we need to look beyond the computational and material properties of algorithms and rather ask: how do algorithms work, how do other actors limit the potentials of algorithms, how do algorithms articulate and rearrange sociality among the constituent actors and how do humans perceive and interpret algorithmic systems? (Bucher, 2018; Lee, 2018). To ask these questions about algorithms implies observing the enactment of their agency in everyday practices.

In this article, I trace this agency through an ethnographic account of an algorithmic system of welfare distribution in India – the Aadhaar-enabled Public Distribution System (AePDS) – across three states of India over a period of 1 year. My enquiry operates at two levels. First, I focus on the situated practices through which AePDS governs the distribution of welfare, and second, I highlight the primary conditions that further shape its performance. Following Bucher (2018), I call AePDS a programmed welfare to show the dynamic and performative nature of the algorithm at work, the contextual embedding of an algorithmic system and its implications for governance of welfare, which includes a multitude of actors. I argue that an analysis through this lens will illuminate two important aspects of algorithms in context: (a) it will render visible the field of situated actions (in this case the welfare distribution) through which algorithms assert their agency and (b) it will present a more relational understanding of algorithms as emerging through existing networks of sociality in welfare, instead of thinking of algorithms as fixed computational resources curtailing human autonomy. At a broader level, the quotidian practices of AePDS reveal how an algorithm designed to de-duplicate the nation as a database (Cohen, 2019) reorganises relations of power for resource allocation while actually navigating the messiness of lived experiences of welfare distribution (Rao and Nair, 2019).

Programmed welfare: a conceptual framework for algorithms in context

Digital technologies, including data, algorithms and models, are increasingly used in public services for their perceived objectivity, efficiency and ability to eliminate human discretion and biases (Flügge et al., 2021). Eubanks (2018) describes the automated eligibility system, ranking algorithms and predictive scoring systems in public services as ‘digital security guards’ who collect information about us, monitor us, infer our behaviour and accordingly control our access to public resources. Digital biometrics, comprising of data, algorithms and models, has emerged as a popular digital tool in recent times (Rao and Nair, 2019), mostly for automating sorting function in targeted welfare programmes. Biometrics allows ‘measuring, analysing and processing the digital representations of unique biological data and behavioural traits such as fingerprints, eye retinas, irises, voice and facial patterns, body odours, hand geometry and so on’ to establish the unique identity of a person through one-to-many comparisons and to verify whether the person is who they claim to be through one-to-one comparison (Ajana, 2013: 3). Hence, it automates the process of both identification and verification in a sequential manner. Identification is automated through the process of enrolling, that is, by capturing digital representations of unique biological features through a sensor device which passes through an algorithmic operation to produce a ‘template’ and storing the ‘template’ on a database and/or on a chip card. Verification happens, first, through acquisition of a ‘live template’ using an algorithmic process and, finally, by matching the ‘live template’ to the ‘stored template’ (Ajana, 2013).

However, adopting digital biometrics for welfare management can have serious damaging effects on certain populations (those with physical or learning disabilities, the elderly and mentally ill, certain races and religions and the homeless) due to their lack of capacity to enrol and verify their biometric data (Wickins, 2007). Magnet (2011) argues that the efficiency orientation of technology design, which requires certain generalisations about the population and a standardised normative body, leads to cases of false rejects and hence the exclusion of certain groups. Furthermore, she argues, while, on the one hand, the introduction of biometric authentication in welfare opens up new market opportunities for private technology firms in public services, it, on the other hand, criminalises poverty by focusing on fraud detection among welfare beneficiaries (Magnet, 2011). Notwithstanding these early alarms, digital biometric identity systems continue to flourish in welfare management and reinforce these concerns as they form new alliances with artificial neural network–based machine learning algorithms of various kinds and more recently, with blockchain technology (Madianou, 2019). The algorithmic assemblages that these digital identity platforms generate include human (e.g. beneficiaries, private companies, policy-makers, public officials and last mile service providers) and non-human actors (e.g. data, algorithms, models, sensor devices), institutions and processes (e.g. welfare schemes, enrolment and disbursement processes, sorting eligibility, documentation), documents (e.g. proof of residence, welfare cards and muster rolls) and resources (financial, social, political, institutional, etc.). Their conditions and constitutive enactments vary as per the context and scope of welfare, such as in the case of humanitarian crisis (Madianou, 2019), ration disbursement (Hundal and Chaudhuri, 2020) and employment guarantee schemes (Dhorajiwala, 2018).

While the exclusionary impact of biometrics in welfare and specifically in AePDS is well-established (Drèze et al., 2017; Hundal and Chaudhuri, 2020; Magnet, 2011), I focus on how such impacts are generated through the contextual embedding of (de-duplication) algorithms in everyday practices of people in routinised yet flexible association with things, processes, documents and resources (Neyland and Möllers, 2017). In analysing these practices, I align with Bucher’s (2018) focus on ‘how and when different aspects of [the] algorithm. . .becomes available to specific actors, under what [institutional and social] circumstances, and who or what gets to be part of how algorithms are defined [and implemented]’ (p. 4). Bucher (2018) refers to the algorithmic power that emerges through specific assemblages to shape ways of knowing and acting as programmed sociality. While Bucher uses the term programmed to denote the assembling and organising capacity of the algorithms, underlying its dynamic and performative potential, by sociality she implies how different human and non-human entities come together to facilitate certain kind of interactions (Bucher, 2018). Analytically, it reveals ‘how actors are articulated in and through computational means of assembling and organizing, which always already embody certain norms and values about the social world’ (Bucher, 2018: 4). I apply this concept of programmed sociality in the specific context of welfare to analyse how computational means of sorting rightful beneficiaries are constituted and enacted through the heterogeneous and fluid networks of humans, databases, machines and institutional processes. In this sense, instead of treating the sorting capacity of the algorithm as its fixed and abstract potential, I ask how different actors embed this algorithm in their everyday practices around welfare and how they interpret, navigate, reinstate, resist and subvert the algorithmic authority. I unpack and examine the algorithmic assemblage of AePDS to demonstrate the actual nature of the de-duplication algorithm when enacted in the context of welfare management.

Methodology

I draw on three sets of fieldwork in Andhra Pradesh, Jharkhand and Karnataka, conducted respectively, in July 2018, March 2019, and May and November 2019. While the goal of the fieldwork was to understand the enactment of Aadhaar in various welfare schemes of the state, in this article, I present narratives of enactment of AePDS that provides subsidised food grains, particularly to people living below the poverty line. While the ultimate objective of AePDS as a pan-India scheme is to ensure food security for all, the operational details and day-to-day enforcement of the scheme are decided by the respective states. This means that in some states linking of Aadhaar to ration cards (RCs) was sufficient, while in others, Aadhaar-based biometric authentication (ABBA) was mandatory for disbursal of ration every month. The three states I covered during my fieldwork followed a targeted public distribution system and deployed both the Aadhaar-enabled PDS database and ABBA (fingerprint) for ration disbursal. As I was interested in the everyday enactment of Aadhaar in the welfare systems, I focused not only on the moments when beneficiaries accessed or received welfare, but also on moments of denial or failure, moments of coping with denial of failure and also moments of redressing grievances through formal and informal means.

My fieldwork consisted of observations of welfare disbursement practices, unstructured qualitative interviews and focus group discussions with beneficiaries of PDS (149), and semi-structured qualitative interviews with the front-level service providers and intermediaries (23), such as FPSOs, data entry operators and government personnel at the village-, block- and district-level offices (20). All interviews were conducted in the local language of the state. In Jharkhand and Karnataka, where access to the field was facilitated by a local non-governmental organisation (NGO), we also interviewed their community workers (10). In Andhra Pradesh, we followed the real-time monthly dashboard of the state civil supplies department and focussed on districts that showed higher rates of authentication failure. We covered two to three districts in every state. In all three states, research assistants and field interpreters played important roles in the data collection process.

We kept extensive notes on our observation data. Interviews conducted in vernacular languages were directly transcribed into English. We also collected pictures of RCs, application forms for linking of Aadhaar and RCs, Aadhaar update acknowledgement slips, government circulars and notifications and so on. These documents worked as supplements to the interview data. We identified main themes emerging from the field notes and interview transcriptions from the first phase of fieldwork and categorised data from all subsequent fieldwork along these themes. The three main themes that emerged across all data were the issues of access to welfare system, issues of exclusion post-access and issues of uneven allotment. I focus on all these three themes while analysing the implications of algorithmic sorting of rightful beneficiaries on the day-to-day experience of the PDS in India. Analytically, I focus on moments of conflict, friction and breakdowns to understand the messiness through which a real-world system of sorting come into being (Star and Bowker, 1999). To protect the identity of respondents, I use pseudonyms in narrating their accounts and mask the exact site locations.

Algorithms at work: AePDS system in India

The PDS is a key welfare programme of India where food grains are sold at subsidised prices in fair price shops (FPSs). One must have an RC, which is issued to a household with one or multiple members, to access the food grains. From the 1990s onwards, efforts were made to reduce the fiscal burden of the PDS by moving from a universal to a targeted system, improving efficiency and mitigating leakages to the non-target population. The introduction of information communications technologies (ICTs) in the PDS must be seen in the context of these systemic goals (Masiero, 2015; Masiero and Prakash, 2019) which continues with the integration of Aadhaar into the PDS.

Aadhaar is a digital identity platform where in all Indian residents are eligible to get a Unique Identity Number (a 12-digit random number) using their biometric and demographic details. Aadhaar captures fingerprints, iris scan data and facial photographs along with name, age, gender and address. Until 2020, 1.24 billion Aadhaar have been issued (UIDAI, 2020). At the core of this ID platform is a de-duplication algorithm that helps to authenticate the uniqueness of an individual against the existing database of the population, known as the Central ID Data Repository (CIDR). The de-duplication algorithm that completes the process of ‘templatisation’ is standardised and authorised by the Unique Identification Authority of India (UIDAI) (Venkatanarayanan, 2017). As a foundational ID, Aadhaar only provides a ‘purposeless’ number that guarantees identity alone, without any promise of benefits or entitlements. Hence, the performative agency of this algorithm emerges only in the contexts of its use.

The linking of Aadhaar with PDS started around November 2014 with the Government of India (GoI) pushing for digitisation of the RC database along with seeding of existing Aadhaar numbers (Government of India, 2014). During linking, the Aadhaar number is stored on the PDS database. This helps to de-duplicate the RCs database. A push to deeper integration came with the passing of the Aadhaar (Targeted Delivery of Financial and Other Subsidies, Benefits and Services) Act, 2016 (GoI, 2016: 2), which led to seeding of the Aadhaar database with their PDS database. This means that each member on a household’s RC will have to link their Aadhaar number to the RC. Post linking, they can authenticate their identity and eligibility at a ration shop every month through an electronic device that capture their biometrics. Any one member from the linked RC could perform this authentication on behalf of the entire family. Thus, Aadhaar as an algorithmic identity platform becomes a technology of governing welfare by objectively sorting who should get subsidised grain and who should not.

Aadhaar reportedly helps clean up bogus RCs, duplicate RCs and remove the names of beneficiaries who have wrongfully been entered into an RC. This sorting action is meant to remove any inclusion errors in the PDS database. The second level of sorting happens when beneficiaries with the Aadhaar-linked RCs are required to verify their identity biometrically through the electronic machines present at the FPS before receiving their entitlement. This is meant to ensure that only rightful beneficiaries mentioned on the RC get their entitlement and no one else can get it on their behalf. Besides these two direct algorithmic sorting functions, meant for efficient targeting of welfare beneficiaries, AePDS also claims to prevent leakages of food grains by FPSOs. Through Aadhaar linking, as soon as beneficiaries authenticate their fingerprints, the allotted quantity get reflected on the system (either a laptop screen or a screen of electronic point of sale machine). In many cases, the system is connected to the weighing machine via bluetooth which then displays the weight of the actual quantity of grains specified, before the grains are handed over to the beneficiaries. These additional steps are expected to work as deterrents for FPSOs from indulging in quantity fraud.

I focus on all these three intended fields of actions of the AePDS through three moments of beneficiaries’ everyday encounters with this algorithmic system of targeted welfare management that renders its underlying algorithmic assemblage visible and allows us to locate different configurations of agency and opacity with the system. It is important to mention here that beneficiaries of AePDS do not necessarily see this as an algorithmic system. They perceive it as a ‘government service’ in which Aadhaar manifested through a documentary proof and fingerprint authentication serves as a new administrative mandate. I leverage AePDS as a context to witness algorithms at work and hence describe the experience of AePDS as an algorithmic assemblage.

Algorithms, welfare and governance

I refer to AePDS as a programmed welfare to highlight its ability to (a) organise welfare allocation computationally, based on matching inputs from two databases, and (b) sort each transaction of welfare claim algorithmically, based on biometric data input from the beneficiaries. This lens underscores the ways in which the computational arrangements of the digital identity imbue the food distribution system with new actors, norms and alliances, in which both human and non-human actors come to shape each other’s performances. By invoking the lens of a programmed welfare, we can see it as a larger algorithmic assemblage, which more fully balances the algorithmic interest of sorting and targeting in welfare with different interests of multiple actors who embed the new algorithmic system in their everyday lives in different ways to coerce, subvert, resist, bypass and rework the algorithms’ intent (Kitchin, 2017). Below, I summarise how this balance occurs in all the three encounters with the algorithmic welfare system that I narrated above.

In the first moment of encounter, the algorithmic system decides who will stay on the food security programme and who will not by de-duplicating the PDS database. However, this agency of the algorithm is mediated and enabled through the human actors responsible for manually seeding the PDS database with Aadhaar numbers. Furthermore, those who are kept out will have to then identify human actors to renegotiate their entry into the system. At this point, AePDS changes the usual ways and the familiar actors through which beneficiaries were used to solve PDS-related problems. Instead of village heads and FPSOs, they now have to contact block offices or sometimes even district civil supply offices, or Aadhaar agents, to resolve their problems. As I have mentioned above, in some cases, FPSOs accompany them and guide them to do the paperwork because for most beneficiaries (and sometimes for the FPSOs as well), the new system remains a mystery, which they are unable to untangle, let alone solve. It became apparent through our interactions with the village- and block-level officials that a lot of the autonomy enjoyed by them has been now delegated to the algorithmic system, which is controlled in a more centralised manner at the state level. Therefore, they are unable to identify why a particular beneficiary was unable to get in and even if they do (e.g. in the case of data mismatch), they are no longer allowed to update the system. While curtailing the autonomy of local state officials, Aadhaar has introduced a whole new set of intermediaries such as enrolment agents, seeding agents and data entry operators in government offices (Chaudhuri, 2019; Khera, 2017); these intermediaries have now become crucial in getting entry into and in navigating the system.

In the second moment of encounter, beneficiaries had to rely on the algorithms’ ability to recognise their biometric existence and thereby their eligibility to get the welfare, even after successful seeding. However, the algorithm acted erratically, temporarily disabling their access and forcing multiple attempts of authentication with cleaner fingers or different fingers at each new attempt. Multiple attempts for authentications were so rampant that beneficiaries and FPSOs formed alliances to devise new strategies to beat the technological system. Many FPSOs in Andhra Pradesh and Karnataka performed fingerprint authentication and ration disbursal on separate days to avoid crowding at their shop (Chaudhuri, 2019; Muralidharan et al., 2020). Some FPSOs in Andhra Pradesh maintained manual records of allocated ration by issuing small yearly booklets for beneficiaries, in which they also kept notes on which fingers worked better in one particular month. While these strategies worked for partial authentication failure, in the case of complete authentication failure, beneficiaries had two options. They first tried to update their biometric details in the system, and if the problem persisted, they opted for alternative mechanisms (bypassing biometric authentication altogether) as mandated by their respective state authorities. Depending on what kind of legitimate alternatives were made available to people, ranging from mobile OTPs to being enlisted on the exempt list, beneficiaries had to forge alliances with FPSOs, village revenue officers or food officers at the block level. These alliances and their experiences varied from one place to another. For example, in rural Jharkhand, many households did not have mobile phones. FPSOs used their mobile phones to generate OTPs for them, but that meant too many OTPs were generated through a single phone. This raised serious suspicion for corruptions by FPSOs, and by the time of our field visit, this alternative channel was discontinued by the state government. In Karnataka, even though the use of mobile OTP was available as an alternative, many FPSOs did not enforce this at their shops, for reasons they did not specify during interviews.

In the third moment of encounter, some FPSOs leveraged the newly found opacity and uncertainty around the AePDS to retain their autonomy and discretion over the process. While on the one hand, this could be interpreted as a straightforward case of corruption, on the other hand, we found FPSOs openly requesting beneficiaries to spare a bit of their entitlements for elderly widows who are facing complete authentication failure and thereby exclusion from welfare. In all the three states, we interacted with elderly beneficiaries who were single card holders and were facing exclusion due to authentication failure. Many of them reported to have relied on the FPSOs to supply them with subsidised ration. As Malati in Jharkhand or Suramma in Andhra Pradesh said, ‘I do not know how he does it, but I get some subsidised food grain (to survive on) every month without authentication’. Here again, the opacity of the algorithmic sorting is rendered as a metaphor for resolving issues without the need to comprehend fully these mechanisms.

All three moments illustrate how AePDS serves to: (a) make welfare more economically viable by correcting inclusion errors in PDS (Drèze et al., 2017; Hundal and Chaudhuri, 2020); (b) ensure more effective digital surveillance of population and programmes (Jacobsen, 2012); and (c) integrate seamlessly private sector organisations, mainly technology firms, to enter the institutional arrangements of public service delivery, by creating quasi-state entities such as UIDAI (Dattani, 2019; Singh, 2019) and by bypassing local governance structures (Chaudhuri, 2020).

These normative orientations of AePDS, however, are not accomplished merely by the fixed properties of the de-duplication algorithms but through the co-constitution of algorithms with the institutional rearrangements of welfare that brings in new actors; rearranges power, resources and agency across different actors; and reworks or reinstates mechanisms of control and coercion. This algorithmic assemblage is, furthermore, situated within the larger political economy of algorithmic governance which aims to reduce welfare expenditure and make welfare inaccessible or unattractive for a larger section of the population (Eubanks, 2018) and encourages states to acts in a platformised manner to bring real-time citizen data for private sectors to innovate and provide services (Gillespie, 2010; O’Reilly, 2011; Singh, 2019; Sobbrio, 2018; Van Couvering, 2017) with the state tightening the centralised control over the platform for better surveillance and security (Jacobsen, 2012).

From my ethnographic analysis, I argue that considering the AePDS specifically as a programmed welfare system, render the underlying assemblage of human, non-human and institutional actors visible enough to analyse how the performative and contingent nature of algorithms manifest through articulations of their varied interests. It shows us how FPSOs, despite their curtailed autonomy, continue to forge alliances with beneficiaries to navigate the new system while local governance mechanisms at village and block levels are completely side-tracked from the new algorithmic assemblage; how beneficiaries who fall through the cracks of PDS and Aadhaar database seeding rely on new intermediaries such as data entry operators or Aadhaar agents to get back on the system; and how after a broken alliance with machines (in case of authentication failure), beneficiaries seek the assistance of actors at the district civil supplies or UIDAI offices with whom they hardly interacted before the introduction of AePDS. More importantly, while the algorithmic sorting adds to the opacity and uncertainty of targeting in welfare (as neither beneficiaries nor FPSOs understand when and how authentication work), it further consolidates the power and influence of the FPSOs, in the absence of any other recourse at local levels. Hence, the everyday enactment of AePDS as a programmed welfare system highlights how, instead of being stable technologies of governance, its algorithmic assemblages merely (re)arrange governance mechanisms which get enacted within a socio-material context (Bucher, 2018; Dourish, 2016; Lee, 2020) through the subjective realties of everyday experiences.

Conclusion

Algorithmic systems are often criticised for their opacity as they add to the existing complexities of bureaucratic institutions (Danaher et al., 2017) and make the logics and processes of governing more and more obscure for future scrutiny (Flügge et al., 2021). One way to address this opacity and build better models to govern how these systems are applied in particular use cases is to unpack the underlying socio-technical assemblages and examine how these assemblages emerge and perform in different contexts (Kitchin, 2017). In this article, I unravel how an algorithmic sorting of beneficiaries works in everyday practice of welfare in relation to new and old human actors, with old and new databases, machines and documents, in shifting institutional contexts. I show which actors gain or lose autonomy, which alliances are favoured and which are severed, how different users interpret and navigate the system, how spaces and rhythms of interacting within institutions change through algorithmic interventions and, finally, which kind of subversions become necessary to cope with the system. This nuanced ethnographic enquiry into algorithmic systems makes the process of their enactment more transparent and comprehensible, which will, first, inform the process of system design that often fails to account for the contextual embedding of algorithms and hence creates unintended impact and, second, bring in more visibility and legitimacy to the workings of algorithmic authority. As algorithmic systems are poised to become more sophisticated and ubiquitous, debunking their constitutive agency in specific field of actions remains imperative for better design and governance of these systems.