Devansh Awasthi*
Introduction
The objective of Jharkhand’s Draft AI Policy of 2026 is to use AI in governance instead of just for profit. The policy plans to use AI for welfare analytics, fraud detection, beneficiary analysis, grievance prioritisation and correction of errors in inclusion and exclusion under targeted schemes.
These AI uses can help a lot with administrative work like detecting fraudulent applications, checking whether the benefit in a particular case has been provided or providing speedy responses to complaints. There is no way to perform such a detailed analysis without the use of automated technology.
However, the question that arises is not whether Jharkhand should adopt AI for welfare purposes but whether a person can demand human intervention before being removed from the benefits he/she is entitled to based on automated algorithm-driven decisions.
From Beneficiary to Data Pattern
Normally, an administrative decision is taken by an identifiable authority, which can allow individuals to make an appeal against the decision. However, with algorithmic screening, it becomes difficult to point to the authority responsible for making the decision.
For instance, a fraud detection system would compare the data of the person for identification with his/her records in the identity database, bank records, household details, payment history and others and flag him/her as an anomaly on the basis of any one of the following: a misspelled word in the records, incorrect address in the records, sharing a mobile number or irregular payment history. The flag itself does not often have the capacity to eliminate individuals from consideration since it often results in actions such as verification drives or delaying disbursals at the very worst. Therefore, merely regulating ‘automated decisions’ neglects the full impact of action because the decision-making component of the algorithm operates far in advance of any final decision. For example, via risk scoring or predictive ranking, the algorithm determines which files will be reviewed first and subsequently flagged. By the time a human resolves the issue, the algorithm has already determined which file to select, what suspicion to write down, and what score to assign. It is thereby very easy for a human to simply adopt the algorithm’s assessment with no original thought, regardless of whether it was right. Hence, safeguards utilised only when a decision is entirely automated will rarely take effect, despite the actual influence of algorithms on the process. This idea is the basis of the protocol stated below: Algorithmic actions and not algorithmic results must be protected.
The Draft Policy acknowledges this issue regarding health and safety; however, it does not impose analogous duties when it comes to welfare eligibility, fraud warnings, complaint triage, and education risk assessment. The difference is obvious: an algorithm may not substitute for a physician, but it can replace the officer deciding on an individual’s pension.
The Digital Personal Data Protection Act, 2023 does provide minimal safeguards. The Act ensures that personal data remains correct and complete, as well as provides the procedure for grievance redressal.
Still, accuracy is not synonymous with fairness. Many legitimate beneficiaries may have the same mobile number lawfully because one mobile device is being used by the whole family or due to an intermediary connecting people to government IT systems; hence, an accurate representation of this data may lead the model to view it as duplication.
None of this detracts from the compelling nature of the case for government automation. The sheer scale of welfare delivery in Jharkhand makes manual verification of each claim logistically impossible, while there are considerable losses due to ghost beneficiaries and duplicates, which fact alone makes for compelling reasons calling for algorithmic screening, as only algorithmic inputs can detect patterns of fraud that no individual officer could; money recovered from dubious beneficiaries becomes readily available for legitimate claims. Hence, there is a compelling argument that supports automation, which involves fiscal and administrative reasons at once. The problem involved is that statistical proxies used for fraud detection work for that purpose but fail for others as well.
Use of the targeting system also brings forth both types of errors; there are exclusions as well as inclusions to the algorithm since the proxies involved are not likely to be perfect; thus, even a lone asset captured by the household income model could lead to perceiving a family as no longer vulnerable. The Comptroller and Auditor General has agreed that the identification criteria for the poor cause both types of mistakes.
Jharkhand has an even bigger challenge because the Draft Policy aims to serve the people who are primarily rural, with tribal and linguistic diversity. Algorithmic models developed using data based on urban, formalised, and electronic access to information will yield wrong results for the precise population that social welfare was intended to serve.
The inconsistencies that exist in various records and affect the benefit delivery system have been acknowledged by the Supreme Court; these inconsistencies have become entrenched into a pattern of deprivation of benefits when processed algorithmically for lakhs of people. Administrative law applies even when decisions are taken through statistical modelling rather than a manual process; these laws are applicable regardless of the manner used for using public authority. For instance, if a beneficiary’s pension is delayed because the model indicates duplication, then the beneficiary should be informed about this. If a grievance is given a low priority based on statistics, the person whose complaint is being addressed should know about the reason. If a benefit has been denied to a person based on an ‘anomaly’, the concerned officer should determine if the inconsistency does exist and not only on the basis of the model. Similarly, the Responsible AI principles set down by NITI Aayog call for ensuring transparency and accountability on issues that can impact the lives of people.
Protocol on the Welfare Due Process
The Government of Jharkhand must formulate the binding Welfare Due Process Protocol to ensure that whenever an Artificial Intelligence system is used to determine benefits of social welfare schemes, due process is followed. The protocol should consist of five essential principles.
First of all, welfare systems qualifying as at risk/high-impact should be defined strictly as such when they affect eligibility, level of benefits or continuity of benefits, priority in addressing grievances, and actions regarding fraud investigations. The definition should be based on the impact to the person and not merely on whether the algorithm itself generated the output.
Secondly, individuals should be made aware when decisions are made by means of algorithms. This knowledge entails being informed of the deployment of AI, its risks, and the agency making the decision. Saying only that the program failed to pass “system checks” does not sufficiently convey this information.
Thirdly, everyone who is impacted should be able to seek a human review of his/her situation. The reviewer must be able to disregard the algorithm’s direction, be able to assess circumstantial evidence, modify original records, and provide an explanation detailing how the conclusion was derived. When fraud is proved, benefits may be temporarily suspended. No one should be faced with starvation, lack of treatment, or financial insecurity while his/her issue is being resolved.
Fourthly, some form of a prior evaluation should take place before any algorithms are utilised. These evaluations should include the error rate, the efficiency of the training data, effects on affected tribal and rural populations, language access, disability-related concerns, and other issues associated with false positives. A periodic evaluation after the program implementation should be conducted by a disinterested party. This investigation will provide a public document on how the program functions, which department is responsible for it, which vendor is being supplied for the program, and what kind of process is used for decision-making, as well as the statistics.
Finally, any contracts made between the government and AI companies must offer public accountability. No agency can deny providing information to the public on the grounds that the model belongs to some private entity. The contract must grant the Government access to audits, logs of any changes made to the model, and dismissal logs, as well as allowing the Government to evaluate all applications on an individual basis.
AI Should Facilitate Welfare Processes, Not Determine Them
The State of Jharkhand is deserving of congratulations for having recognised concepts of privacy, fairness, explainability, auditability, and inclusivity within the Draft AI Policy. We also appreciate the ongoing commitment to using auditing and risk assessment to protect the rights of citizens. The process by which the promise shall be implemented remains to be seen, so that citizens can trust the system that has been put in place. Despite the necessity of dealing with the problem of welfare fraud, this does not warrant a complete abuse of power. Not every inconsistency in record-keeping is indicative of fraud or requires those on the lower end of the economic scale to justify themselves.
The essence of the above is that the constitutional issue comes into play. The algorithm is able to detect patterns and gauge risk; however, it cannot provide a rationale for its decisions or allow the persons involved to present evidence. Article 14 provides for protection against arbitrary treatment.
* The author is a third-year Law student at Dr. Ram Manohar Lohiya National Law University, Lucknow. The author may be contacted at devanshawasthi.rmlnlu@gmail.com.
This blog reflects the personal views of the author and does not necessarily represent the views of The Policy Chronicle.