Illustration by Somnath Bhatt

AI-based policing: a veneer of neutrality to India’s casteist criminal justice system

A guest post by Nikita Sonavane & Srujana Bej. Nikita Sonavane is a lawyer and co-founder of the Criminal Justice and Police Accountability (CPA) Project. Srujana Bej is a lawyer and human rights researcher. The authors would like to thank Kanishka Singh and Sumit Ray for their research assistance with this essay.

This essay is part of our ongoing “AI Lexicon” project, a call for contributions to generate alternate narratives, positionalities, and understandings to the better known and widely circulated ways of talking about AI.

Police records in India often categorize people as ‘habitual’ or ‘dangerous’ criminals based solely off of colonial constructions of caste. With the Indian State now digitising all police records for ‘efficient’ technology-forward policing, these measures will only entrench and legitimise casteist oppression and caste-based notions of ‘criminality.’ Digitization obscures the caste bias on which such criminality is constructed and sows the seeds for AI-based ‘predictive’ policing.

In this essay, we explore the historical origins of caste, its legacy in policing, and its relevance to AI in the context of the criminal justice system. We argue that any new AI-based predictive policing system will likely only perpetuate the legacies of caste discrimination and the unjust criminalisation and surveillance of marginalised caste communities — since the very nature of historical police data is casteist. We highlight that any analysis of the framework of predictive policing requires engagement with the complexities within the institution of caste itself. Finally, we call attention to the risks of ‘efficient’ and ‘neutral’ AI-based technologies enforcing a digital-driven caste system.

Less than a fortnight away from the Parliamentary elections in 2019, scores of ‘habitual offenders,’ primarily identified from Vimukta communities,¹ were rounded up by the police and taken in for questioning. The lengthy interrogation included questions about their friends, family, assets owned, and an otherwise routine recording of fingerprints. This habitual offender ‘drive’ had been ordered from the ‘top’ (higher rungs of the police department) in the interest of maintaining law and order. Registers on habitual offenders, containing extensive amounts of personal information, are now being digitised.

(As told to authors by those detained in police custody, May, 2019)

Caste, Coloniality, and Policing

Caste is a ‘closed system’ of social stratification which assigns occupations by birth and enforces a strict behavioural code to segregate (and deprive) ‘impure’ and ‘polluted’ individuals who lie in the lower rungs of the caste hierarchy. The caste system’s hierarchy is organised within the varna system (the four-fold occupational hierarchy prescribed by Hindu religion and spearheaded by the priestly Brahmin caste²), and comprises hundreds of jaati (caste groups) organised within each varna. Adivasis (indigenous communities), nomadic, and semi-nomadic tribes are marginalised by this system despite not ascribing to caste and lying outside its fold.

British colonial authorities embraced the caste system in India because it provided a fertile ground for determining subjects of control. In order to establish the facade of ‘order’ with limited resources, colonial authorities selectively policed populations whose behaviours they branded as criminal, threatening, or unnatural. They were particularly threatened by nomadic and semi-nomadic tribes who challenged the colonial State’s monopoly over forests and did not pay State tax due to their unregulated, mobile economies. The caste system’s rigid hereditary occupational hierarchy inspired colonial construction of thievery, dacoity, and robbery as ‘occupational’ criminalities. Furthermore, nomadic and semi-nomadic tribes’ mobility and lack of caste capital contributed to the colonial fiction of their ‘dangerous’ and ‘hereditary’ criminal disposition.

Nomadic and semi-nomadic tribes were ascribed “criminality by birth” and considered as ‘hereditary criminals addicted to systematic commission of non-bailable offences’ under the colonial Criminal Tribes Act (CTA), 1871. These communities were liable to extensive police surveillance and control. The CTA became a primary basis for determining who ‘deserves’ to be policed, and it achieved its objective by employing and appealing to the logics of the caste system. Nearly simultaneously, colonial authorities completed formalizing and centralizing India’s modern police force. The CTA thus became ingrained in policing institutions and structure, since it developed alongside the modern Indian police.

The CTA was repealed in 1952, following the adoption of the Indian Constitution, and the communities criminalised were ‘denotified.’ They now use the term ‘Vimukta’ (‘released/freed’) to self-assert their identity. However, many Vimukta communities continue to face the colonial taint of criminality because caste prejudice is the logic of Indian society and the CTA is ingrained in police origins. Several Indian states have also adopted legal provisions concerning ‘habitual offenders’ (HOs) and maintained the surveillance systems designed under the CTA. The transfer of power from the colonial state to the post-colonial resulted in the concentration of power in the hands of oppressor caste elites. This has allowed for the colonial architecture of criminality to endure and thrive even after independence through the category of HOs.

Police stations across India maintain registers of HOs — also called ‘history-sheeters’ — in their jurisdictions, with extensive details of their lives and daily movements. While their identification may not explicitly be based on caste, collective police action overwhelmingly identifies members of the Vimukta communities as HOs. These registers record their demographic details such as place of residence and caste, personal information such as age and identifying marks on the body, and ‘evidence’ of criminality: details of their habits, their method of committing crimes, their property, particulars of their associates, and places they frequent. The caste-identified hereditary criminal of the past is now categorised in the seemingly caste-neutral administrative category of the HO. The branding as a HO makes one perpetually vulnerable to police surveillance, suspicion, indiscriminate arrests for petty or imagined offences, police extortion, and violence. The extent of the disparate criminalisation of Vimukta communities remains undocumented and invisiblised.

Caste as Administrative Data

There is no official data on the total population of Vimukta communities. The multiplicity of jatis and communities which do not ascribe to the caste system are subsumed within three major constitutional categories: Scheduled Castes (SC), Scheduled Tribes (ST), and Other Backward Classes (OBCs). The primary reason for the creation of these administrative caste categories is to recognise the historical marginalisation of groups by the caste system and provide affirmative action to remedy some of the structural oppression by the caste system. However, these administrative categories, merely constituted for paternal purposes, are devoid of historical contexts. The rationale for SC and ST categories were derived from characteristics of untouchability and primitiveness assigned respectively to communities that assert themselves as Dalits and Adivasis. For OBCs, the Mandal Commission formulated indicators along the axes of social, economic, and educational backwardness. However, the persecution faced by Vimukta communities did not translate into administrative recognition despite demands, due to objections from Constitution drafters. Vimukta communities have been administratively subsumed into multiple other caste categories, despite their distinct cultures and unique historical marginalisation.

Despite not being considered by the State as a distinct caste category, the Vimukta communities’ caste is encoded in administrative data. Colonial police practiced blanket surveillance of Vimukta communities to cultivate thorough knowledge of their ‘criminal habits’ and deter crime. This system of detailing the lives, habits, ancestry, and movements of individuals who are ‘habitually’ suspected of committing crimes continues through the administrative category of HOs in police registers. Narratives of genetic criminality have been replaced with narratives of economic need, group immorality, compulsive indiscipline, or addiction to narcotic substances. However, these narrative and discursive ideologies rationalising ‘habitual offenders’ mirror colonial discourses that constructed the ‘hereditary criminal tribe,’ sans pseudo-scientific justifications and explicit caste associations. In some cities, the police overwhelmingly identify ‘habitual offenders’ from Vimukta communities. At least in some police jurisdictions, habitual offender lists operate on explicit casteist principles and the assignment of inherent criminality to certain caste groups, particularly Vimuktas.

Situating Caste in AI-based Policing

Until the last decade, Indian police stations maintained HO registers and all other case documents in paper. However, the State seeks to capitalise on digital dividends to ‘empower’ and ‘modernise’ the police to tackle crime. With the government’s institution of the Crime and Criminal Tracking Network & Systems (CCTNS), an opaque digital repository, all police records including HO registers are being uploaded and stored in a centralised digital database. The CCTNS can be used to develop crime-mapping, analytics, and predictive policing systems. Some states have employed former Big Tech executives to help incorporate big data into policing.

HO registers, the products of casteist governmentality and policing, are being fed as data for potential AI-assisted predictive policing. That is, ‘objective’ and ‘neutral’ algorithms will use these historical police records and registers to identify which individuals are more likely to commit crime. We are assured that such identification will hinge not only on caste but on other ‘facts’ such as one’s ancestry, profession, residence or past criminal records. However, all these criteria are reproductions of caste. Given the casteist nature of these police records, the outcome of who gets identified as an object of surveillance will likely hinge on invisibilized caste identities. Algorithmic oppression — i.e. the unjust oppression of social groups through automated, data-driven processes — is being manifested in the ‘research and design’ process itself by feeding caste-biased data into predictive algorithmic models.

It is important to note that most police records, and not merely habitual offenders registers, are significantly encoded in caste bias and are reflective of caste. The police may fail to register a crime or may under-play the nature of crime allegedly committed when the accused are members of dominant caste groups and the victims belong to a marginalised caste group. For example, in the Dhankaur atrocity case, the Dalit family assaulted by the police and imprisoned for grave offences such as “rioting” and “attempt to murder,” was protesting against the police’s refusal to register a case against a dominant caste individual who had allegedly encroached their lands. In the Khairlanji atrocity case, similarly, a Dalit family’s attempt to register a criminal case against a dominant caste individual for land encroachment was first thwarted and then minimised by the police. When the family was later brutally killed by the dominant caste group for ‘transgressing’ caste, the police failed to register commensurate cases against the dominant caste accused. Thus police records may erase or invisibilise the criminal acts and propensities of dominant caste communities, overwhelmingly portraying them as victims. Police records also significantly underestimate the number of times marginalised caste communities are victims, and choose to overwhelmingly portray these communities as offenders.

The Khairlanji and Dhankaur atrocities provide an additional insight — the perpetrators belonged to dominant marginalised caste groups and not the so-called upper caste groups. Thus, they help us understand the heterogeneity of caste oppression and underscore the nature of localised caste relations and oppressions. Therefore, in critiques of caste, technology, and policing, it is important to account for the varying and overlapping dynamic implicit within an evolving system. It is also necessary to move away from the now commonplace flat comparisons between caste and race, in order to attend to the specific historically and geographically situated multitudes of caste kinship and enforcement.

Could public regulatory accountability measures — such as transparency and impact assessments — allow for unbiased predictive policing? Put another way, is it possible to perform predictive policing on non-casteist ‘objective’ data variables? Annette Zimmerman, Elena Di Rosa, and Hochan Kim write that algorithmic fairness (i.e. applying impartial rules for algorithmic decision outcomes) must reckon with the “issue of whether neutrality constitutes fairness in background conditions of pervasive inequality and structural injustice.”

Caste is fundamental to the structure of polity. It dictates one’s access to employment, education, land, and social capital. The institution of caste is also constantly evolving and increasingly veering away from “visible” markers of its existence — Particularly for oppressor caste people who have converted this into modern forms of capital like property, higher educational credentials, and strongholds in lucrative professions. Therefore, even if predictive policing relied on socio-demographic data, such as one’s income, educational qualifications, assets, or residence — instead of police records — caste oppression would continue. Indeed, it continues to influence the police’s internal organisation. The police practice caste through segregated eating, using caste slurs, collecting caste details even when unnecessary, and failing to fulfil the constitutional mandate of affirmative action in their composition.

AI based policing will then provide a veneer of neutrality to India’s casteist policing, and entrench the criminalities inscribed on Vimukta and other marginalised caste communities. It will also likely disproportionately curtail the liberties, agencies, dignities, and freedoms of caste minorities, and lend support to arguments that these communities’ lifestyles, rather than police practices, must be reformed. Renowned legal philosopher Ronald Dworkin argued in his essay ‘Artificial Intelligence: What’s to Fear?’ that the privileged would use AI technologies to cause harm and then escape legal responsibility by pinning causality of harm on the inanimate technology rather than piercing the veil to study the human agents behind the institution and use of such harmful technology. Will AI policing only provide further impunity to the police and the State to continue pursuing casteist policing?

Given the historically casteist nature of police records, the lack of ‘objective’ data for predictive policing, and the risk of entrenching impunity for casteist policing through AI, we must ask — should AI-based predictive policing even exist when it cannot produce or embody fairness, objectivity, and our fundamental principles of justice?

Footnotes

[1] Vimukta communities are nomadic and semi-nomadic tribes formerly notified as ‘criminal tribes’ by colonial authorities under the Criminal Tribes Act, 1871. These communities were criminalised, branded as hereditary criminals by birth and subject to extensive surveillance, marginalization and confinement. British colonial authorities considered their mobile lifestyles deviant and beyond the realm of imperial, European civilising order. The Criminal Tribes Act was repealed in 1952 and replaced by several legislations concerning ‘habitual offenders’. Vimukta communities continue to face the stigma of criminality and their traditional occupations remain criminalised through excise laws, wildlife conservation laws, and cattle slaughter prohibition laws. They continue to be targeted as criminals by the police, are regularly externed and are often required to sign bonds for “good behaviour.”

[2] The varna system comprises the broad occupational categories of Brahmin (priests), Kshatriya (warriors), Vaishya (merchants), and Shudra (labourers and craftspeople)