Introduction

In July 2025, SEBI passed an interim order against Jane Street, a US-based proprietary trading firm, freezing assets worth INR 4,843 crore on findings that the firm had manipulated the Bank Nifty index through coordinated algorithmic strategies executed across 18 expiry days. An algorithmic system could place, change, and cancel orders at a speed, accuracy, and volume that no human trader could match and, more importantly, that no human regulator could see in real time. Just a few months prior, 111 registered brokers settled proceedings with SEBI for associating with third-party algorithmic platforms that had promised retail investors guaranteed returns, further illustrating that the harms flowing from AI trading platforms reach well beyond institutional manipulation and into the savings of ordinary investors.

This piece discusses the core problem revealed by such incidents. There is a gap between emerging Artificial Intelligence-Powered Trading Systems (“AI platforms”) and the Indian Securities Market Regulations. Existing literature on this issue has focused on three distinct concerns – (i) the absence of a compensation mechanism for losses caused by algorithmic failures; (ii) the conflict-of-interest risks created by revenue-sharing arrangements between brokers and algorithmic service providers; and (iii) the governance ambiguity produced by SEBI’s principles-based consultation framework where the boards are told they are accountable but never told what accountability actually requires them to do, so companies are left to guess and regulators are left with nothing concrete to enforce against. This paper is narrower in scope. Each of those concerns presupposes a regulated ecosystem, but the ecosystem cannot be regulated if the entities generating the harm cannot first be classified. This piece therefore addresses that prior question, the definitional gap itself and treats the amendment of the classificatory framework as the necessary precondition for any downstream accountability mechanism.

The SEBI has consistently issued several circulars pertaining to algorithmic trading such as performance/ return claimed by unregulated platforms offering algorithmic strategies for trading, safer participation of retail investors in Algorithmic trading, and extension of timeline for implementation of SEBI Circular ‘Safer participation of retail investors in Algorithmic trading’.

However, this piece argues that, currently, SEBI regulations are ill-equipped to regulate AI platforms and therefore need to be amended to include AI investment platforms and automated decision-making. It identifies four necessary characteristics of such an amendment drawn from comparative analysis. For the purposes of this discussion, these entities may be described as Algorithmic Investment Service Providers (“AI platforms”), a term that captures both their automated nature and their role in shaping investor outcomes without prematurely classifying them as advisers, brokers, or an entirely distinct category.

The part-I of this piece starts by highlighting features of AI platforms that make them incompatible with existing SEBI regulations, then provides an overview of those regulations. Part-II  then provides the four necessary characteristics of the amendment inspired by the EU, Singapore and the USA. It concludes by warning that no such definition will be adequate without a proper allocation of responsibilities.

Why are AI platforms incompatible with the existing regulations? 

AI platforms differ structurally from existing trading entities. The following characteristics set AI platforms apart, making them incompatible with existing governance frameworks.

First, they operate at speeds and scales that remove any meaningful scope for human intervention at the point of decision, making the concepts of advice and execution, as understood in the SEBI (Investment Advisers) Regulations 2013, functionally inapplicable. Regulation 2(m) of the SEBI (Investment Advisers) Regulations, 2013 defines an investment adviser as “any person, who for consideration, is engaged in the business of providing investment advice to clients.” The definition is person-centric in three respects, each of which an AI platform structurally fails to satisfy and the failure runs deeper than mere technical non-compliance.

Second, their outputs are generated through probabilistic inference rather than rule-based reasoning, meaning that the same system, trained on the same data, may produce different outputs under novel market conditions that its training did not anticipate, as the Jane Street forensic reconstruction demonstrated.

Third, their decision logic is opaque (also known as the black-box problem), even to those who built them, which means accountability cannot be determined by examining what the system was told to do. Retaining chains of accountability is foundational to corporate governance frameworks because it provides the basis for identifying and assigning legal liability. A firm can be held answerable for a decision only when that decision can be attributed to it. Where the reasoning behind an output cannot be recovered, this chain of attribution collapses. Regulators, courts, and investors are then unable to determine whether the harm arose from a design flaw, inadequate data, a deployment decision, or emergent model behaviour.  It effectively shields the deploying entity from liability, since every potentially responsible actor, including the developer, the deploying broker, and the vendor, can rely on the opacity of the system to deny that any specific fault is attributable to them.

For example, under Section 27 of the SEBI Act, 1992, liability for an offence committed by a company attaches only once to a specific individual, whether a director, manager, or other officer is identified as having acted with consent, connivance, or neglect in relation to the default. The IA Regulations reinforce this by designating a named human individual as the “principal officer” under Regulation 2(1)(s), who is personally accountable for the non-individual investment adviser’s regulatory compliance under Regulation 20. Where an AI system generates a materially erroneous compliance output, such as a defective related-party transaction assessment or a misleading statutory disclosure, and the internal reasoning of that system cannot be recovered, no such identification is possible. The managing director, the compliance officer, and the technology vendor can each point to the opacity of the system to deny specific authorship of the default, and the statutory threshold for liability is never crossed.

Fourth, when multiple platforms deploy correlated models, their outputs under stress conditions reinforce rather than offset each other, producing systemic risks and it includes algorithmic herding, where platforms trained on the same data reach identical trading decisions at the same moment; self-reinforcing feedback loops, where the resulting price movement is read by those same platforms as a further signal to exit, deepening the fall; and liquidity evaporation, where the buyers needed to absorb simultaneous large-scale exits simply disappear. When training data and model architecture are shared across platforms, their outputs move in lockstep, turning what might have been an isolated market shock into a cascading crisis that spreads faster (pg.3, 4) than any concentration of human advisers acting on similar information could produce. that are qualitatively different from those generated by individual human advisers acting on similar information. When multiple AI trading platforms are trained on the same historical data and built on similar models, they tend to react to market conditions similarly. In normal markets, that mostly goes unnoticed. But during a shock, like a liquidity crunch or unexpected macro event, those same systems can start selling, rebalancing, or exiting assets at the same time.

Since trades occur at machine speed, there is little room for human intervention. The market fall itself then becomes fresh input for the models, reinforcing the same signal and accelerating the sell-off. The 2010 Flash Crash exposed the same dynamic. A single automated selling programme triggered rapid reactions across trading systems, causing the Dow Jones to crash nearly 1,000 points within minutes. Unlike human traders, correlated AI systems move almost instantly, leaving little time to stop the cascade.

These structural features of AI platforms produce systematic misclassification under SEBI’s existing categories. The following section demonstrates how this misfit operates within the current regulatory framework.

Existing Frameworks

The existing framework begins with Regulation 2(l) of the SEBI (Investment Advisers) Regulations 2013, which defines investment advice as advice relating to investing in, purchasing, selling or otherwise dealing in securities or investment products. Since it does not explicitly include AI platforms, its application to them is ambiguous. This ambiguity is exploited by platforms, which present themselves as technology products rather than investment advisers. While Regulation 2(l) is, in theory, broad enough to capture any personalised, compensation-driven, securities-related advice regardless of how a platform describes itself, the problem lies less in the law’s text than in its enforcement structure.

The IA Regulations operate primarily through registration (Regulation 3), and registration is self-initiated (regulation 5). Platforms presenting themselves as financial wellness apps, AI portfolio tools, or smart money ecosystems never approach SEBI as investment advisers, while SEBI lacks a proactive classification framework to independently identify them as such. Branding, therefore, performs important legal work. It does not alter the substance of the activity, but it determines whether regulatory scrutiny is triggered at all. SEBI’s actions against unregistered platforms have focused primarily on finfluencers rather than technology platforms.

SEBI’s 2012 circular on algorithmic trading directly targets the algorithmic system. It imposes risk controls and audit trail requirements on brokers offering direct market access. However, it was designed for institutional high-frequency trading, where pre-programmed rule-based algorithms (pg.77) execute orders at speed on behalf of sophisticated participants. Since the circular regulates the broker and the technical systems through which orders are routed and executed, rather than the platform that decides what to buy or sell and for whom, it cannot govern AI platforms that generate personalised portfolio recommendations and automate execution for retail investors. The circular asks whether the order was placed safely; it does not ask whether the decision to place it was appropriate for the client receiving it; cannot govern AI platforms that generate personalised portfolio recommendations and automate execution for retail investors.

Similarly, the SEBI (Prohibition of Fraudulent and Unfair Trade Practices) Regulations 2003 are broad enough to capture misleading AI-generated marketing, but they cannot be applied to algorithmic harm, as it requires proving that a specific representation was false, which is difficult when the harm flows from the aggregate, opaque behaviour of a machine learning model rather than from an identifiable human statement.

SEBI itself acknowledged these problems in its December 2024 SEBI Board Paper on Responsibility for AI Tools and its June 2025 Consultation Paper on the Responsible Use of AI and ML in Indian Securities Markets. These papers identify opacity, model risk, third-party dependency, and systemic risk as primary concerns, and propose principles of accountability, transparency, and explainability as the basis for a future framework.

However, these instruments are not binding legal documents and are merely advisory. Part II of this article proposes certain design principles, drawn from comparative analysis, that can be incorporated into amendments to these regulations.

The authors are Year IV students at the NALSAR University of Law.