How We Investigated Welfare Algorithms in India (Part II)

AI Accountability Fellow Kumar Sambhav Shrivastava and journalist grantee Tapasya investigated opaque welfare algorithms in India that wrongfully cut off benefits to thousands of its poorest citizens. In this piece, Tapasya describes their approach to accessing public records through India’s Right to Information Act, as well as other reporting methods they used to overcome records’ denials and government bureaucracy. Part I of their methodology can be read here. 

Allow me to start with an anecdote detailing my experience with India's transparency law, the 2005 Right to Information Act.

I filed Right to Information (RTI) requests with the Food Supply and Consumer Welfare department of Odisha, an eastern Indian state where, in 2015, public officials using IBM’s Master Data Management Tool rejected food security assistance to more than 5 million people . The RTIs aimed to access official records regarding the algorithms deployed by the department to identify beneficiaries of the food security scheme, which entitles poor citizens to free or subsidized food.

Initially, the information official rejected the RTI on a flimsy pretext, claiming that the requested information was not "specified." Upon interacting with several other individuals who had filed RTIs in the department, I found that it was their default response to reject every information request.

I filed an appeal to contest the denial of information. Despite requesting that the hearing be conducted online due to the considerable cost and time involved in physically attending the department's office in Odisha, I was asked to appear in person. Hearings can be conducted online; my request was not exceptional.

During the hearing, I vigorously challenged the denial of information, providing a detailed rebuttal to a team of officials who were reluctant to share the information. I cited provisions from the law, and the appellate authority seemed more impressed by my knowledge of the law than the genuine need for proactive and public disclosure of official decisions affecting millions of India's poor citizens.

Ultimately, the appellate authority ruled in my favor. I was once again summoned to the department's office in Odisha, this time to inspect official files. I felt thrilled and satisfied that my persistence had yielded results. However, my excitement was short-lived.

During the inspection, I was subjected to a series of probing questions by the information officer, such as my occupation, the specific information I sought, and my intentions. These questions were accompanied by extensive praise for the department's technological advancements and its improved ability to detect "ineligible" and "fraudulent" claimants, purportedly benefiting the most deserving citizens.

I was offered an interview with an official responsible for overseeing the department's technology initiatives. While he provided insights into the department's adoption of technology and its procurement process involving private companies, I sought official records to corroborate his statements and obtain comprehensive details beyond the optimistic narrative he presented.

After hours of waiting, I was finally permitted to inspect their files. With over half a dozen files placed on my designated table, I diligently reviewed and noted down file numbers and pertinent sections for which I intended to request copies. An official closely monitored my activities to prevent me from taking photographs of the files; I was only allowed to inspect them, with the promise of receiving requested information afterward.

Following the inspection, I submitted another application to the information officer, specifying the files from which I required information. However, upon returning to Delhi, I realized that I had been deceived by the information officer. He stopped responding to me. Despite sending emails and WhatsApp messages to remind officials of their obligation to provide me with information, I received no response. Even after the appellate authority intervened and instructed the information officer to furnish the files, I never received the files.

From anecdotes to systems 

We started off the project with instances and anecdotes of exclusions from welfare schemes that were previously reported by national and regional media. These reports often suggested the use of algorithmic systems in central and state schemes, leading to arbitrary exclusions. One example was the Odisha state government’s use of IBM’s Master Data Management tool during the implementation of the national food security program. Similarly, media reports from Telangana state shed light on the impact of the Samagra Vedika program for welfare delivery, which utilizes various official databases to determine a citizen's poverty status. There were reports of people protesting wrongful exclusions in various states across the country.

As an initial step towards investigating the use of algorithms in welfare delivery, we needed to narrow down the schemes for in-depth scrutiny. Our aim was to move beyond the stories of victims of algorithmic decision-making and explore the deployment, developers, and systemic flaws associated with these algorithms.

By collating and filtering information from media reports and government documents, I began creating a database of schemes using similar technologies/programs. The AI Observatory was very helpful during this exercise. This process yielded a list comprising three schemes of the union government and over six schemes implemented by four state governments.

We subsequently focused our investigation on selected schemes from this list.

Right to Information requests

We utilized India’s Right to Information Act to access non-public information concerning schemes using algorithms for beneficiary-identification. I initially filed RTIs with two central and six state government departments, requesting inspection of official files. To bolster our chances of obtaining information from reluctant governments, I simultaneously filed RTI applications with 16 district-level and 16 block-level (sub-district) offices, adopting a three-tiered approach for each scheme.

I also filed RTIs with multiple key authorities at the central and state levels, particularly focusing on dedicated IT departments overseeing algorithmic systems in various states. Persistent follow-ups and appeals were essential tactics, given the frequent initial refusals for information.

Concurrently, I conducted field reporting in the states, interacting with excluded beneficiaries. I was able to convince some of them to file RTIs seeking details on how their data was stored and processed in categorizing them as "ineligible." This proved beneficial. We got replies that stated clearly that the exclusions were a result of algorithmic decision-making, without verifying their accuracy.

As a team, we continuously reviewed responses to our information requests, refining our approach for subsequent filings. Our goal was to obtain comprehensive details on the algorithms, including source codes, databases used, and the number and reasons for beneficiary exclusions.

Given the limited available information on welfare-delivery algorithms, our RTIs were broad, aiming to inspect all records related to technology usage in a scheme. We hoped to at least uncover insights into the algorithmic deployment decisions and potential performance evaluations conducted by government departments. Key questions included the system's identification process, decision-making protocols of authorities, and evaluations of algorithmic effectiveness over time.

Challenges with RTI

When it came to RTIs, we faced an information blockade. Almost all the state and central departments either skirted our questions entirely or gave incomplete information that was of very little use when it came to understanding how the algorithms were used. In many cases, the RTIs were transferred from one government department to another, each claiming the other department had the information requested. Officials used tactics such as requesting our physical presence thousands of kilometers away for ‘hearings’ and ‘inspections’ to exhaust our resources, knowing journalists in India can rarely afford to be so persistent in the face of bureaucratic adversity.

Even after multiple visits, phone calls, and emails, officials never provided complete information as required under the transparency law.

Overcoming challenges and finding personal accounts

Since the authorities were uncooperative, the next step was to join the dots to make sense of different kinds of information that I was able to gather. To understand the harm that was caused by the use of algorithms, I combined information from the following sources:

The information from RTIs: Through RTIs, I received official memos and notifications regarding the exclusion criteria of various schemes, official letters to integrate welfare schemes with the algorithmic platforms that identify beneficiaries, standard operating procedures used by officials at different levels to use algorithmic decisions to make a final call about a claimant’s eligibility, and names of databases used by the algorithms. 

Some district, block, and municipality-level offices gave us lists of excluded beneficiaries. These lists also gave us information village-wise or municipality-wise on exclusions. But these lists consisted of thousands of people and there was no way to find out which of these exclusions were wrongful. However, within these smaller geographical and administrative boundaries, it became possible to track down certain individuals. This involved visiting the villages or municipalities, especially in Odisha, and going door to door in these areas and asking people if they or anyone they knew was on the list of excluded beneficiaries.

Interviews with officials: During the process of reporting from three states and meeting state and local-level government officials regarding RTIs, I was able to interview them on how they were using the software for welfare delivery. Some officials gave interviews on the record, while others talked to me off record to help me understand the process of beneficiary verification. The interviews made it clear that the officials went with the decisions of the algorithms displayed on the portals that they used for the scheme and did not cross-check before disqualifying someone from a welfare scheme. They took up specific cases of exclusion only when someone filed a grievance complaint with them. Local-level officials told me that if the algorithm, deployed at the state-level, finds someone ineligible, benefits are stopped at the state-level. Again, they would only look into it if a grievance was filed.

Information through activist groups: Before setting out for the field, I contacted local activists. I received significant help from  Right to Food Campaign activists in Telangana and Odisha. In Odisha, they pointed me to areas I could visit and helped me interview hundreds of people who spoke different regional languages that I was unable to grasp in full. In addition to that, I collected information on certain cases of exclusion filed in the High Court of Telangana and Odisha by people excluded from the states’ food-security and farmers’ welfare schemes, respectively. 

In Telangana, activists of a local NGO, ASEEM, put me in touch with several families whose food security cards were deactivated. Several of them had filed petitions in the High Court of Telangana, and the activist there, S.Q. Masood, had filed a case in the Supreme Court of India to challenge arbitrary exclusions from the food scheme without giving an opportunity to the claimants to present their side. Masood shared the Supreme Court case documents: orders, petitions, and affidavits submitted in the Supreme Court by the government. These documents gave us the overall figures of exclusion used in the Telangana story. The activists also informed me of their experiences while working on such cases of exclusions.

Cases of petitioners: The documents of cases filed by excluded beneficiaries of the food program in Telangana and the farmers’ scheme in Odisha helped in understanding specific case studies. In Telangana, the activists provided these documents. In Odisha, I received a list of petitions filed on exclusions from an RTI, then contacted lawyers’ networks and was able to track two lawyers working on two different cases. One of them provided all the case documents that showed how a small-scale farmer in the state was excluded because he was misidentified as a large-scale farmer by the software that the state is using.

Talking to affected individuals: The petitions of exclusion from Telangana and Odisha were a gateway to understanding the effects of algorithmic welfare exclusions in different states. The activists and lawyers representing the petitioners put me in touch with them. I was able to meet them in person and talk to several of them on the phone. This helped me to understand their reality and the indifference of the authorities who kept turning their back on them. 

Their struggle with the authorities proved how algorithmic decision-making had assumed the front seat, how officials had relegated their responsibility to machines, and how in many cases they had shifted the reasons for their exclusion after realizing that the algorithms were wrong. 

Tips for journalists

When undertaking long-term investigative projects, it is imperative to recognize the fast-paced nature of life in a developing economy. People's priorities and challenges evolve rapidly, presenting a challenge in aligning investigative focus with the dynamic nature of their lives. Navigating these dynamics requires persistent follow-ups and readily investing time in focused engagement, with patience and respect for the private lives of the individuals one is writing on.

Furthermore, stakeholders in the story may possess varying levels of understanding about their own circumstances. It is most essential for a journalist to verify everything from multiple sources and rely on official documents and recorded correspondence, especially when reporting on governments. 

When working in collaborative teams, it is essential to distribute work, commitment, time, and responsibilities with care and thoughtfulness. Half the job is to become a team. The other half is to work like one. Keep talking to each other. Be candid about how it’s working out and regroup if initial ideas don’t pan out the way the team imagined. Hold yourself accountable to each other and take care of each other. A team working together on a long-gestation investigation can go through mental as well as physical challenges. A healthy team produces the best results.