Speaking at the India International Trade Fair in New Delhi on 14 November 2026, SEBI Chairman Tuhin Kanta Pandey said the regulator will "soon" issue detailed guidelines for the responsible use of AI in capital markets. According to MediaNama's report of the speech, citing PTI, the guidelines will cover disclosures, accountability of regulated entities, and a "human-in-the-loop" approach with human oversight over critical decisions. No timeline was given.
The direction is sensible. The design details will decide whether the guidelines help investors or only add paperwork.
The strongest case for the rules
The case for intervention is real. Securities markets run on trust and speed, and AI can fail in both ways at once. A model that gives retail investors biased or hallucinated advice can harm thousands of people before anyone notices. Correlated models can also herd, which amplifies volatility. Investors usually cannot tell whether a recommendation came from a person or a model, and cannot inspect the model if they could tell. Disclosure addresses that gap directly. Human oversight gives a regulator a named person to hold responsible.
SEBI's own record shows it is not starting from zero. Its board approved rules in December 2024 making regulated entities responsible for the outputs of AI tools they build or buy, and for the privacy and security of investor data those tools touch. SEBI's December 2024 board papers contain that approval. The amendments were notified in early 2025, according to Taxmann's summary. The principle is one we support: a firm cannot outsource its liability to a vendor or an algorithm.
What the 2025 consultation already told us
The coming guidelines build on SEBI's June 2025 Consultation Paper on Guiding Principles for Responsible usage of AI/ML in Securities Markets. We could not extract the PDF's text, so the details below come from MediaNama's coverage of it. As MediaNama reports, the paper proposed these measures:
- Client-facing AI uses, such as algorithmic trading, portfolio management and advisory, would require plain-language disclosure of purpose, risks, limitations, accuracy and fees.
- Third-party AI services would be treated as provided by the market participant, which stays responsible for compliance.
- Senior management would oversee models, independent audits would be required, and testing would run in an environment separate from live systems.
- Input and output data would be retained for at least five years.
- Internal uses such as compliance, surveillance and cybersecurity would face lighter requirements.
The last point is the most important. A risk-tiered approach, which regulates AI where it touches investors and leaves back-office tools alone, is what proportionate regulation looks like. If the final guidelines keep that tiering, they will be better than many AI regimes elsewhere.
Where the design can go wrong
"Human in the loop" is not a safety property by itself. Pandey acknowledged this in September 2026 at Global Fintech Fest, according to MediaNama. He said humans "can also overtrust machines" and that oversight "must be competent, empowered and accountable." He also said "an AI-generated alert is not a finding." That is the right diagnosis. A compliance officer who clicks "approve" on hundreds of model outputs a day adds liability without adding safety. If the rules require a human at every step, firms will build rubber-stamp processes that satisfy an auditor and protect no one.
The better test is outcome-based. SEBI should specify which decisions are critical, such as order execution limits, client suitability determinations and fraud flags. For those decisions it should require a reviewer with real authority, the information needed to override the model, and a record of the cases where they did. Applying one oversight standard to every AI use would raise costs for small brokers and advisers more than for large firms.
Disclosure must be useful, not boilerplate. A page of model caveats that every client scrolls past is no protection. Disclosure should say when AI materially shaped a recommendation and what the firm does when it errs. A standard format would also let investors compare firms.
Concentration risk needs a light touch. The consultation flagged dependence on a few AI vendors. Vendor reporting is reasonable. Rules that effectively bar frontier models from Indian markets would push firms toward weaker, local tools and leave them less secure. Accountability should stay with the regulated entity, as the 2024 rules already provide.
The wider policy context
This approach fits India's national direction. MeitY's India AI Governance Guidelines, released in November 2025, say a new horizontal AI law is not required at this stage. They leave enforcement to sectoral regulators such as SEBI, RBI and TRAI. That is a defensible model. A financial regulator understands market structure far better than a general AI authority would, and sector rules can be amended faster than a statute.
The trade-off is fragmentation. Banks, brokers, insurers and fintechs often use the same models under different regulators. SEBI, RBI and IRDAI should align their definitions of "critical decision" and their disclosure formats, or compliance costs will multiply.
What to watch
AI is also a tool for SEBI. According to MediaNama, its SUDARSAN system, which scans public social media for investment fraud, has flagged more than 20,000 instances since November 2025. A regulator that uses AI to catch fraud understands that blanket restriction would be self-defeating.
Three questions will show whether the guidelines work:
- Is oversight defined by the decision or by the technology?
- Are the rules issued as guidelines, which firms can adapt, or as binding regulations with penalties for technical lapses?
- Is there a consultation window on the final text?
SEBI has the right instincts: tiering, accountability for outcomes and a chair who admits the limits of human review. The final text should translate those instincts into rules that make oversight real and keep it affordable for firms of every size.