The Bias in Facial Recognition is Hindering Financial Inclusion: Why Every Financial Org Must Wake up to This Problem

The Supreme Court’s 2025 judgment in the acid attack survivor’s case directed authorities to make digital KYC more accessible, including by providing reasonable accommodations and alternative verification methods.
The Bias in Facial Recognition is Hindering Financial Inclusion: Why Every Financial Org Must Wake up to This Problem
Published on: 
5 min read

Authored by Vignesh Krishnakumar, CTO and Co-founder, HyperVerge

India has made tangible strides in moving financial services online. Opening a bank account, completing KYC and accessing financial services can now be done through a screen. However,  the experience is not uniform for everyone. As India is a land of diversity, differences in skin tone, facial features, language preferences and physical abilities can affect how people interact with digital verification systems. For some, the KYC journey is seamless, and for others, it can be a struggle when systems aren't designed with these differences in mind.

The Supreme Court’s 2025 judgment in the acid attack survivor’s case directed authorities to make digital KYC more accessible, including by providing reasonable accommodations and alternative verification methods. This August, a widely shared video from Nagaland showed a man's KYC eye scan failing repeatedly while people around him tried to help. Both raise a legitimate question: when a biometric check fails on a genuine person, is the system built to recognise that?

These incidents make it clear that digital KYC cannot be considered inclusive simply because it is automated. What makes inclusivity a reality is that verification systems work reliably in Indian realities

The technological flaw that goes unaddressed

A facial recognition system only performs as well as the data, design and conditions it encounters. If a model is trained on datasets that do not adequately represent the diversity of its intended users, its performance can vary across skin tones, facial features, ages and other demographic groups. But the model itself is only one part of the problem. The customer experience depends on the entire verification pipeline, from image capture and preprocessing to liveness detection, model inference and the final decision threshold.

For instance, a face verification system may successfully detect a face but fail during liveness detection. Uneven lighting can affect facial feature extraction. On top of that, differences in camera quality and image resolution can reduce the reliability of the biometric embedding generated by the model. A liveness check may also reject a genuine customer if their movement or depth cues do not match the conditions it was trained to recognise.

These failures should not automatically be treated as evidence of identity fraud. The system needs to distinguish between a genuine identity mismatch, an inconclusive model output and a failure caused by the capture environment or verification method. Each outcome should trigger an appropriate response, such as another capture attempt, an alternative verification route or human review, rather than sending every failed result through the same rejection workflow.

When bias-prone technologies are introduced into Indian realities like modest phone cameras or shaky networks, it’s like stacking one complexity over the other. Each factor alone might shave off a percentage point. If all the wrong combinations come together, they can turn a 99% accurate model into something less reliable for a specific slice of the population it’s meant to serve. Most likely, people who land on the wrong side of the gap tend to be first-time applicants, thin-file customers, older users, and people onboarding from Tier 2 and Tier 3 towns battling with inconsistent lighting and patchy connectivity. These are frequently the exact people with the fewest alternate paths into formal finance.

How does this impact BFSI organisations?

The RBI’s KYC Master Direction states that a financial institution’s customer acceptance policy should not result in denial of services to the general public, particularly people who are financially or socially disadvantaged. Since 2025 the clause explicitly includes persons with disabilities. The RBI has also clarified, in its FAQs on the Master Direction, that the decision to reject a KYC application cannot be automated and must be reviewed by an authorised official.

For banks and financial institutions, this means an automated verification failure cannot be treated as an automatic reason to deny access. The system needs to distinguish between a confirmed identity mismatch, an inconclusive result and a technical or accessibility-related failure. Each outcome should have a defined next step, whether that involves another capture attempt, an alternative verification method or human review.

There is also a business cost to getting this wrong. A false rejection can mean lost customer acquisition spending, additional manual review and a customer who may not return to complete the onboarding process. The challenge is to build verification systems that manage fraud risk without turning genuine customers away because the technology cannot reach a reliable decision.

Framing a fair game for both sides: What should financial institutions ask vendors?

Vendor evaluation needs to be more than a successful product demonstration. Financial institutions should examine how the technology performs across their actual customer base and customer conditions. 

Some questions worth asking include:

  • Does the training and evaluation data reflect the institution’s customer base? This includes differences in skin tone, facial features, age and other relevant demographic characteristics.

  • How does the system perform across customer segments? Vendors should provide disaggregated results rather than relying only on a blended accuracy figure.

  • What happens when the image capture fails? Institutions should understand whether the issue is linked to lighting, camera quality, face positioning, liveness detection or another part of the pipeline.

  • How are uncertain results handled? The system should distinguish between a potential fraud signal and a technical or accessibility-related failure.

  • What fallback options are available? Customers who cannot complete the default workflow should have access to a meaningful alternative rather than being repeatedly sent through the same failed process.

  • Is there a human review process? The institution should know when a case is escalated, who can review it and how the decision is recorded.

  • How is the system monitored after deployment? Model performance can change as the customer population, devices and fraud patterns change. Production monitoring should track failure rates and potential disparities across relevant customer groups.

Third-party assessments and benchmarks can provide an additional reference point when evaluating vendors. However, a benchmark result shouldn’t be an alternative to institution-specific testing. A system that performs well in one dataset or operating environment may not produce the same results across a different customer population, device mix or onboarding process.

Fairness also needs to be treated as an ongoing operational responsibility. Fraud patterns evolve, customer demographics change and models are updated. Regular testing, incident reviews and monitoring of false rejections should be part of the system’s lifecycle. 

Accurate on average is not the same as fair

The question for any institution is not whether its verification system is accurate. It almost certainly is, on average. The question is who sits in the gap between the average and the person in front of the camera, and what happens to them when the first attempt fails. Answer that, and digital KYC starts to do what it was built for.

𝐒𝐭𝐚𝐲 𝐢𝐧𝐟𝐨𝐫𝐦𝐞𝐝 𝐰𝐢𝐭𝐡 𝐨𝐮𝐫 𝐥𝐚𝐭𝐞𝐬𝐭 𝐮𝐩𝐝𝐚𝐭𝐞𝐬 𝐛𝐲 𝐣𝐨𝐢𝐧𝐢𝐧𝐠 𝐭𝐡𝐞 WhatsApp Channel now! 👈📲

𝑭𝒐𝒍𝒍𝒐𝒘 𝑶𝒖𝒓 𝑺𝒐𝒄𝒊𝒂𝒍 𝑴𝒆𝒅𝒊𝒂 𝑷𝒂𝒈𝒆𝐬 👉 Facebook, LinkedIn, Twitter, Instagram