(Play the video in full screen mode)
Answer:
AI-based credit assessment can potentially expand formal credit by using
alternative data and identifying borrowers who may not have a conventional
credit history. That is particularly relevant for MSMEs and underserved
segments. However, the quality of the outcome depends on the quality and
representativeness of the underlying data. If historical data contains
socioeconomic or demographic bias, an AI model may reproduce or amplify that
bias. Therefore, I would see the regulatory objective as not preventing
AI-based credit assessment, but ensuring that its deployment is fair,
explainable, accountable and subject to appropriate model governance. The
institution should know what data is being used, why it is relevant, how the
model is validated, how bias is tested, and how adverse decisions can be
explained to the customer. I would also distinguish between the model making a
recommendation and the regulated entity taking the final responsibility. AI
should enhance credit judgement, not create an accountability vacuum.
Q2. Suppose almost all major banks use similar AI models for credit and
trading. What systemic risk could emerge?
Answer:
The risk is that diversification at the institutional level may not
translate into diversification at the model level.
If multiple institutions rely on similar datasets, vendors, models or
assumptions, they may respond to changing conditions in similar ways. In credit
markets this could produce correlated lending decisions. In financial markets,
similar trading models could generate herding and amplify volatility.
There could also be third-party concentration risk, particularly
if critical AI infrastructure is provided by a small number of cloud or model
providers.
Therefore, supervision needs to move beyond asking whether an individual
bank's model is sound. It should also examine common exposures, common
providers, model concentration and correlated behaviour across the financial
system.
This is where micro prudential supervision and macroprudential
surveillance need to intersect.
Q3. If AI makes an erroneous loan decision, who should be responsible,
the model developer, the bank, or the human officer?
Answer:
From a regulatory perspective, responsibility should not disappear
merely because technology is involved.
The regulated entity taking the financial decision must retain
accountability for the outcome. It cannot outsource regulatory responsibility
merely by outsourcing the technology.
That does not mean the technology provider has no responsibility.
Contractual, legal and supervisory frameworks can allocate responsibilities
appropriately. But from RBI's perspective, the regulated entity must have
adequate governance, validation, monitoring and oversight over the models it
uses.
The principle should be:
Technology can be outsourced, accountability cannot.
This becomes particularly important for material decisions such as
credit, fraud detection, customer access and risk management.
Q4. Why is "human-in-the-loop" important in financial AI?
Could excessive human intervention reduce the benefits of AI?
Answer:
Human-in-the-loop is important because financial decisions have
consequences involving trust, fairness and accountability. AI can identify
patterns much faster than humans, but there may be circumstances where
contextual judgement is necessary.
At the same time, human intervention should not become a ceremonial
approval mechanism. If an officer simply approves every AI recommendation
without understanding it, there is little meaningful human oversight.
The appropriate approach is therefore risk-based human oversight.
Routine low-risk decisions may be highly automated, whereas high-impact
decisions should have stronger human review, escalation mechanisms and audit
trails.
The objective should be to combine machine capability with human
accountability rather than treating them as substitutes.
Q5. If you were advising RBI on AI regulation, what would you
prioritise?
Answer:
I would focus on five areas.
First, governance, including board-level responsibility and clear
accountability.
Second, model risk management, covering validation, monitoring,
documentation and periodic review.
Third, data governance, particularly quality, privacy, security
and representativeness.
Fourth, explainability and customer protection, especially where
AI materially affects access to financial services.
Fifth, systemic risk surveillance, including concentration in
models, data, cloud infrastructure and technology providers.
I would also favour proportionality. AI used for a low-risk internal
administrative process should not necessarily face the same governance
requirements as AI used for credit underwriting or market trading.
That would allow innovation while keeping the regulatory framework focused on material risks.
No comments:
Post a Comment