Legal and Ethical Implications of AI-Driven Financial Prediction Systems in Modern Markets

Legal and Ethical Implications of AI-Driven Financial Prediction Systems in Modern Markets

Om Aditya Mishra

Dept. of Computer Science and Engineering (Cyber Security)

R.V. College of Engineering Bengaluru, India

Pallavi O

Dept. of Computer Science and Engineering (Cyber Security)

R.V. College of Engineering Bengaluru, India

Chitra B T

Dept. of Industrial Engineering and Management

R.V. College of Engineering Bengaluru, India

Abstract – Financial markets have undergone a structural over-haul driven by the pervasive adoption of Articial Intelligence

(AI) and Machine Learning (ML) in trading, risk assessment, and portfolio management. While these technologies have measurably enhanced execution efciency and predictive capability, the pace of their deployment has substantially outrun the regulatory frameworks of leading jurisdictionsincluding SEBI, the U.S. Securities and Exchange Commission (SEC), and the European Securities and Markets Authority (ESMA). The resulting gov-ernance decit introduces layered legal and ethical hazards: algorithmic market manipulation, a reconstituted form of insider trading premised on AI-generated informational asymmetries, systemic fragility from correlated automated strategies, and an emergent class of harms that existing scholarship has inade-quately examined. This paper identies ve such underexplored research gapsautonomous inter-agent collusion, AI-washing in investment disclosures, duciary obligation in AI-mediated advisory relationships, cross-border regulatory arbitrage, and the carbon externality of compute-intensive tradingand integrates them into a cohesive analytical framework. We propose a Tiered Regulatory Compliance Framework (TRCF) capable of classifying AI trading systems by risk prole, and an Extended Ethical Compliance Score (Se) whose formula is augmented with penalty terms for environmental cost and disclosure indelity. The framework is then evaluated against four documented market disruptions to assess its prospective mitigative value. Findings indicate that a granular, evidence-weighted, and multi-jurisdictionally harmonised oversight regime is necessary to preserve market integrity without foreclosing legitimate nancial innovation.

Index TermsArticial Intelligence, Algorithmic Trading, Fi-nancial Regulation, SEBI, SEC, Market Manipulation, Insider

Trading, Fiduciary Duty, AI-Washing, Regulatory Arbitrage, Ethical Compliance, ESG.

1. Introduction

   Contemporary nancial markets are no longer primarily driven by human traders exercising discretionary judgment. Instead, they are shaped by algorithmic agents operating at microsecond latencies, processing vast data streams that no hu-man analyst could realistically evaluate. By 2024, algorithm-driven strategies accounted for an estimated 6075% of U.S.

   equity trading volume; the gure for Indias National Stock Exchange had crossed the 50% threshold for the rst time in the same year [1]. Global AI spending within the nancial sector is projected to exceed $97 billion by 2027 [2], reecting the industrys conviction that AI-derived predictive advantages translate directly into competitive returns.

   This trajectory has, however, produced a consequential reg-ulatory shortfall. The legal architecture that governs securities marketsincluding SEBIs algorithmic trading circulars, the SECs Regulation SCI, and the UKs transposed MiFID II obligationswas designed for rule-based, deterministic sys-tems whose decision logic could, in principle, be traced and audited. Modern AI systems, particularly those employing Deep Learning or Reinforcement Learning, defy this model. Their decision pathways are opaque not only to regulators but often to the engineers who designed them [3]. This opacity complicates the assignment of liability when an AI-driven strategy causes a market disruption and renders the mens rea standard of traditional fraud law effectively unworkable.

   Beyond the question of liability, AIs capacity to derive systematic trading advantages from the synthesis of high-velocity alternative datasatellite imagery of retail parking lots, anonymised credit-card transaction ows, social media sentiment indicesraises foundational questions about market equity. These data sources are, in the narrow legal sense, publicly available; yet the analytical infrastructure required to exploit them is accessible only to well-capitalised institutional actors, effectively recreating the information asymmetry that insider trading law was designed to prevent [5].

   Existing scholarship has addressed several of these con-cerns, yet a cluster of important dimensions remains inad-equately theorised. Specically: (i) the emergence of au-tonomous collusive equilibria among independently-operating AI agents [17]; (ii) the misrepresentation of AI capabilities in investment product disclosuresanalogous to the earlier phenomenon of greenwashing [9]; (iii) the fragmented legal treatment of duciary duty in AI-mediated advisory relation-

   ships [10]; (iv) the structural incentive for rms to exploit divergences between national regulatory frameworks through jurisdictional arbitrage [3]; and (v) the environmental cost of compute-intensive high-frequency strategies and its tension with ESG investment mandates [13].

   This paper addresses those gaps alongside the primary regulatory and ethical questions. Section II surveys the current legal landscape through a comparative jurisdictional lens. Sec-tion III formally identies and theorises the ve research gaps. Section IV develops the threat model. Section V proposes the Tiered Regulatory Compliance Framework and the Extended Ethical Compliance Score. Section VI evaluates the framework against historical disruptions. Sections VII and VIII consider industry implications and policy recommendations. Section IX concludes.

2. Related Work and Legal Landscape

   The governance of algorithmic trading has passed through two broad phases. The rst, roughly coinciding with the early proliferation of electronic order routing in the 2000s, was dominated by concerns about technical stability and system integrityreected in rules such as the SECs Regulation SCI. The second phase, precipitated by a succession of high-prole failures including the 2010 Flash Crash and the Knight Capital incident, expanded the regulatory lens to encompass market integrity, manipulative conduct, and systemic risk. The emergence of opaque ML systems has initiated a third phase in which the very conceptual vocabulary of nancial regulationintent, authorship, causation, disclosureis under strain.

   1. The Responsibility Gap in AI-Driven Harm

      Armour and Eidenmu¨llers analysis of self-driving cor-porations [4] identies what they term the responsibility gap: the difculty of attributing blame to a legal person when the proximate cause of harm is an autonomous system acting within, but not necessarily in accordance with, its designers intent. In securities law this gap is acute because market abuse frameworkswhether under Section 10(b) of the Securities Exchange Act, SEBIs Prevention of Fraudulent and Unfair Trade Practices Regulations, or the EUs Market Abuse Regulationare premised on the existence of a culpable human actor whose mental state can be imputed. Where an AI agent learns spoong behaviour through reward-maximisation without any human programmer encoding that behaviour explicitly, the existing frameworks provide no satisfactory mechanism for prosecution.

   2. Comparative Jrisdictional Analysis

      Major regulatory authorities have adopted meaningfully different approaches to AI in trading, creating a patchwork that is itself a governance risk. Table I summarises the four most signicant frameworks.

   3. Insider Trading Doctrine in the Age of Alternative Data

      The classical insider trading frameworkwhether the clas-sical theory under Chiarella v. United States or the misap-propriation theory from United States v. OHaganrequires

      that the defendant possessed material, non-public information (MNPI). The doctrinal challenge posed by AI-driven alterna-tive data analysis is that the underlying data is, in the tech-nical sense, public: satellite imagery, shipping manifests, and anonymised transactional records are commercially available. What is non-public, and practically inaccessible to ordinary market participants, is the analytical outputthe predictive insight generated by applying substantial computational re-sources to that data [5]. Existing scholarship has agged this grey area but has not produced a workable doctrinal reformu-lation. The present paper argues that the appropriate analogy is not to the possession of information but to the possession of an unfair analytical infrastructure, and that regulatory reform should focus on disclosure requirements for alternative data strategies rather than on extending the MNPI concept.

   4. Existing Technical Literature and Its Limits

   On the technical side, research has examined the microstruc-ture effects of algorithmic trading, including the relationship between HFT participation and liquidity [18], the mechanics of spoong detection [6], and the systemic amplication effects of correlated strategy deployment [7]. What is largely absent from this literature is an integrative framework that connects these technical ndings to enforceable legal standards and that accounts simultaneously for the multi-agent, cross-modal, and cross-jurisdictional dimensions of modern AI trading systems. The framework proposed in this paper is designed to ll that integrative function.

3. Identified Research Gaps

   A systematic review of the regulatory, legal, and technical literature reveals ve substantive areas that have not received proportionate scholarly or policy attention relative to their practical signicance. These gaps are not merely theoretical lacunae; each corresponds to a class of real or imminent market harm. Table II provides a structured overview before each gap is treated in depth.

   1. Gap 1: Autonomous Collusion in Multi-Agent AI Systems

      The possibility that competing AI trading agents might independently arrive at collusive pricing strategieswithout any form of explicit communicationwas demonstrated with rigour by Calvano et al. in a landmark 2020 study [17]. Their experimental setup involved independent Q-learning agents optimising prot in a repeated Bertrand game; the agents consistently learned to sustain supracompetitive prices through what the authors characterise as a pattern of strategic underre-action. The mechanism is entirely emergent: no programmer encoded collusion, and no agent communicated with another. Regulators and competition authorities have taken note. The FCAs April 2024 AI Update explicitly agged the risk that requiring AI systems to report each others manipulative behaviour could produce an adversarial learning dynamic in which detection systems and manipulative algorithms coevolve to outmanoeuvre each other [25].

      TABLE I

      Comparative Regulatory Landscape for AI Trading

      Jurisdiction	Regulator	Primary Framework	Key Provisions	Gap Coverage
      India	SEBI	Algo Trading Circu-lars (202226)	Broker accountability; White Box / Black Box distinction; API access restrictions.	Manipulation, systemic risk; collusion and arbitrage understated.
      USA	SEC	Reg SCI & Market Access Rule (Rule 15c3-5)	Pre-trade risk controls; systems compliance and integrity reporting; 2024 AI-washing examinations.	Manipulation, disclosure; duciary AI gap persists.
      UK	FCA	MiFID II (Transposed)	Rigorous pre-deployment testing; governance frame-work; April 2024 AI Up-date on inter-agent dy-namics.	Systemic and collusion risks beginning to be ad-dressed.
      EU	ESMA / EU	EU AI Act (2024) & MiFID II	Financial AI as High-Risk (Annex III); explainability mandates; supervisory co-operation.	Broadest coverage; arbi-trage risk from non-EU ju-risdictions remains.

      TABLE II

      Five Underexplored Research Gaps in AI Financial Regulation

      Gap	Core Problem	Why Understudied	Proposed Contribution
      1. Autonomous Collusion	Multi-agent RL systems converge on supra-competitive equilibria without explicit coordination.	Single-system threat mod-els dominate; multi-agent dynamics treated as an-titrust rather than securi-ties law issue.	Extended Class IV threat model; multi-agent sandbox stress-testing requirement.
      2. AI-Washing	Firms overstate AI sophistication in product disclosures to attract capital.	Greenwashing literature is the nearest precedent; spe-cic AI-disclosure fraud framework is nascent.	Disclosure delity penalty term (Df ) in Se; taxonomy for AI capability claims.
      3. Fiduciary Duty	Liability allocation for AI robo-advisor losses is legally unresolved.	Fiduciary doctrine is human-centric; robo- advisory liability addressed in isolation without general framework.	Human-in-the-Loop Account-ability Matrix (HLAM); three-tier oversight intensity model.
      4. Regulatory Arbi-trage	Regulatory divergence creates structural incentives to domicile AI systems in lax jurisdictions.	Arbitrage is studied in traditional nance; AI-specic cross-border migration patterns are undermodelled.	Regulatory Harmonisation In-dex (RHI); bilateral consulta-tion trigger at RHI < 0.7.
      5. Environmental Cost	Compute-intensive trad-ing carries a signicant carbon footprint; ESG mandate tension is unad-dressed.	Energy cost treated as IT infrastructure issue, not a nancial ethics or regula-tory concern.	Energy-intensity score (Ec) integrated into Se formula.

      This threat is not addressed by any current regulatory frame-work. SEBIs algo circulars, the SECs market manipulation rules, and the EU AI Act all conceptualise manipulation as a property of a single system and its operator. The multi-agent dimensionwhere the harm emerges from the interaction of independently lawful systemsfalls into a structural gap. The threat model developed in Section IV extends the existing typology to accommodate this category, and the TRCFs Class IV requirements are expanded accordingly.

   2. Gap 2: AI-Washing in Investment Product Disclosures

      The SECs Division of Examinations identied AI-washing as a formal examination priority in 2024, following a 2023 sweep that found numerous registered investment advisors making unsubstantiated claims about AI-driven portfolio man-agement [9]. The 2025 examination priorities further ex-panded this oversight, requiring rms to demonstrate that their representations about AI capabilities are accurate and that adequate policies exist to superise AI use. The parallel with greenwashing is instructive: in both cases, the harm arises from a material misrepresentation that induces investor reliance. The

      difference is that AI capability claims are harder to falsify than ESG metrics, because the analytical systems being described are often proprietary and opaque.

      Academic frameworks for nancial AI regulation have not yet integrated this form of disclosure fraud. The present paper treats AI-washing as a sub-category of material mis-representation under established securities fraud doctrinespecically Rule 10b-5 under the Securities Exchange Act and comparable provisions in SEBIs PFUTP regulationsand proposes a disclosure delity penalty term (Df ) within the ethical compliance score that regulators can use to quantify the severity of misrepresentation during examination proceedings.

   3. Gap 3: Fiduciary Duty and Autonomous AI Advisory Systems

      The duciary relationship between an investment advisor and a client has historically been analysed as a human relationship: the advisors duty of care and duty of loy-alty attach to a natural person who makes informed, con-textual judgements on behalf of the client. The emergence of robo-advisory platformsautonomous AI systems that construct and rebalance portfolios without ongoing human interventionintroduces a structural mismatch. When the sys-tems recommendation causes client harm, the allocation of responsibility between the software developer, the deploying rm, and any supervising human advisor is legally unresolved in most jurisdictions.

      Venable LLPs 2025 analysis states plainly that delegating decisions to a machine does not absolve the human duciary from oversight [10], and ESMA has similarly held that nancial institutions must take full responsibility for the actions of AI systems they deploy. Yet these positions, while legally coherent, do not resolve the practical question of what oversight is necessary and sufcient for a rm to discharge its duciary obligations. The present paper proposes a Human-in-the-Loop Accountability Matrix (HLAM) that maps decision categories to oversight intensity requirements, providing an operational framework that legal and compliance functions can implement.

   4. Gap 4: Cross-Border Regulatory Arbitrage

      Regulatory arbitragethe practice of structuring activities to exploit differences between national legal regimesis a well-documented phenomenon in traditional nance. Its AI-specic manifestation has received far less attention. The IMFs 2025 report on AI in securities markets explicitly warns that rms with advanced AI capabilities may bypass stricter oversight mechanisms in their home jurisdictions by routing their AI infrastructure through countries with less developed regulatory frameworks [3]. This dynamic is structurally incen-tivised: compliance cost estimates for the EU AI Act alone run to approximately C29,277 per AI product [12], creating a meaningful nancial motivation to seek lighter-touch regimes. The implications extend beyond cost arbitrage. A rm that locates its AI trading infrastructure in a jurisdiction without real-time monitoring requirements effectively exports

      the systemic risk of its strategies to markets that may lack the capacity to contain a resulting disruption. This paper proposes a Regulatory Harmonisation Index (RHI) computed as the cosine similarity between jurisdiction-level regulatory requirement vectors across ve dimensionspre-trade con-trols, explainability mandates, real-time monitoring obliga-tions, sandbox requirements, and duciary standards. An RHI below 0.7 between any two major jurisdictions is treated as a presumptive arbitrage risk warranting bilateral regulatory coordination under IOSCO protocols [11].

   5. Gap 5: Environmental Cost and the ESG Contradiction

   The carbon footprint of large-scale AI model training and inference is now a mainstream concern in technology policy, yet it has received almost no attention in the nancial ethics or nancial regulation literature. ESG-labelled assets under management globally reached an estimated USD 4150 trillion by 2025 [13], reecting a broad commitment by the nancial sector to environmental responsibility. At the same time, the deep learning models and high-frequency infrastructure un-derpinning AI trading strategies are among the most compute-intensive applications in commercial use. There is a structural contradiction between a rms ESG investment mandates and the energy footprint of the AI systems through which those mandates are executed.

   This paper argues that the omission of energy-intensity con-siderations from AI trading ethics frameworks is analytically incomplete and operationally inconsistent with the sectors stated ESG commitments. The extended ethical compliance score proposed in Section V incorporates an energy-intensity score (Ec) as a subtractive term, providing a mechanism through which energy use can be factored into regulatory assessments and public disclosures.

4. System Architecture and Threat Model

   A complete account of the risks associated with AI-driven nancial prediction requires a threat model that extends be-yond the single-system perspective that characterises most ex-isting regulatory analysis. The model developed here classies threats across ve categories, distinguishes between single-agent and multi-agent harm mechanisms, and incorporates the evidentiary and jurisdictional dimensions identied in the research gap analysis.

   1. Single-Agent Threat Categories

      1. Market Manipulation Through Learned Behaviour: AI systems optimising for prot in market microstructures can learn manipulative strategiesspoong, layering, momentum ignitionas instrumentally useful without any programmer explicitly encoding such behaviour [6]. The absence of human intent does not eliminate harm; it merely complicates attribu-tion. Under the proposed framework, Class III and IV systems are subject to real-time monitoring of order-to-trade ratios and submission patterns precisely to detect emergent manipulative behaviour.

      2. Correlated Strategy and Flash Crash Risk: Where multi-ple institutions deploy AI systems trained on similar data using similar architectures, the resulting trading strategies may be highly correlated. In a stress scenario, simultaneous de-risking by multiple AI systems can produce the self-reinforcing liq-uidity withdrawal that characterised the 2010 Flash Crash [7]. The systemic dimension of this risk distinguishes it from single-rm manipulation and requires a different regulatory responseone focused on portfolio diversity requirements and circuit breaker mechanisms rather than intent-based liability.

      3. Alternative Data and Informational Asymmetry: The de-ployment of AI systems capable of deriving predictive signals from alternative data creates a form of market asymmetry that is functionally equivalent to insider trading but falls outside its current legal denition [5]. The harm is not that private information is misappropriated but that the effective informa-tional playing eld is radically skewed toward actors with the computational resources to process publicly available data at scale. Regulatory responses should focus on alternative data strategy disclosure rather than on extending MNPI doctrine.

   2. Multi-Agent Threat: Autonomous Collusion

      Building on the analysis in Section III-A, the multi-agent threat model treats collusion as an emergent property of the interaction between independently-operating AI systems rather than as a product of any single systems design. The formal characterisation follows from Calvano et al. [17]: let A = {a1, a2,…, an} be a set of AI trading agents each independently maximising expected discounted prot. In the presence of strategic omplementaritywhere each agents optimal strategy depends on the strategies of othersrepeated interaction can produce a Nash equilibrium that is collusive without any agent encoding collusion as an objective. The policy implication is that sandbox stress-testing for Class

      IV systems must include multi-agent adversarial scenarios in which candidate deployments are tested against existing market participants.

   3. Disclosure and Representation Threats

   The AI-washing threat identied in Section III-B oper-ates at the interface between securities law and technology assessment. The harm mechanism is straightforward: a rm makes material representations about the sophistication or performance of its AI systems in offering documents, mar-keting materials, or regulatory lings; those representations induce investor reliance; the representations prove false or materially misleading. The specic challenge for regulators is that AI capability claims are harder to verify than conventional nancial disclosures because the systems being described are proprietary and technical assessment requires specialised expertise. The Df penalty term in the extended Se score is designed to translate the degree of disclosure indelity into a quantitative compliance indicator.

5. Tiered Regulatory Compliance Framework and

   Ethical Scoring Model

   The framework proposed here has three interlocking com-ponents: a risk classication taxonomy that maps AI trading systems to discrete regulatory tiers; an extended ethical com-pliance score that quanties each systems risk-adjusted ethical standing; and a set of operational instrumentsthe HLAM and RHIthat address the duciary duty and regulatory arbitrage gaps respectively.

   1. Risk Taxonomy and Tiered Classication

      Table III presents the Tiered Regulatory Compliance Frame-work. The classication is based on two primary parameters: the opacity of the systems decision mechanism and the nature of the data on which it operates. These parameters are operationally proxied, respectively, by whether the systems decision logic can be extracted and audited in interpretable form, and by whether the systems training and inference data is drawn exclusively from structured, exchange-provided sources or incorporates alternative, unstructured inputs.

   2. Extended Ethical Compliance Score

      The base ethical compliance score proposed in earlier work captures transparency, accountability, fairness, and systemic risk. The present framework extends this by adding two penalty terms derived from the research gap analysisan energy-intensity penalty (Ec) addressing Gap 5 and a disclo-sure delity penalty (Df ) addressing Gap 2. The extended formula is:

      Se = w1(T )+ w2(A)+ w3(F ) w4(R) w5(Ec) w6(Df )

      (1)

      where each term is dened as follows. T [0, 1] is the Transparency and Explainability Index, reecting the degree to which the systems decision process can be extracted, audited, and explained in terms that a nancially sophisticated regulator can evaluate. A [0, 1] is the Accountability and Liability Mapping score, capturing the clarity with which responsibility for system outputs is allocated across the developer, deploying rm, and supervising human advisor. F [0, 1] is the Market Fairness Index, measuring the extent to which the systems operational advantage derives from capabilities that are, in principle, accessible to a broad range of market participants rather than from exclusive data or infrastructure advantages.

      R [0, 1] is the Systemic Contagion Risk score, quantifying the extent to which the systems strategy correlates with those of other market participants and the potential scale of market impact in a stress scenario. Ec [0, 1] is the Energy-Carbon Intensity Score, calibrated against a benchmark of compute per unit of notional trading volume. Df [0, 1] is the Disclosure Fidelity Penalty, assessed by the regulator on the basis of the gap between the rms representations about its AI capabilities and the empirically-assessed capabilities of the deployed system. Weights w1 through w6 are set by the relevant national regulator and may be adjusted by jurisdiction

      TABLE III

      Tiered Regulatory Compliance Framework (TRCF)

      Class	System Type	Risk Level	Transparency Requirement	Monitoring Require-ment	Intervention
      Class I	Deterministic rule-based systems with fully auditable logic.	Low	Full logic disclosure to broker and ex-change.	Standard pre-trade risk controls; post-trade reporting.	None beyond stan-dard registration.
      Class II	Supervised ML models using structured, exchange-sourced data.	Moderate	Explainability documentation; feature importance reports.	Periodic audit; out-of-sample performance disclosure.	Regulatory review on anomalous performance signals.
      Class III	Deep Learning / en-semble systems using alternative or unstruc-tured data.	High	Model card; training data provenance; in-ference audit trail.	Real-time order- to-trade ratio monitoring; kill-switch mandate.	Mandatory notication within 30 minutes of anomalous behaviour.
      Class IV	Autonomous RL systems capable of self-modication; multi-agent deployments.	Critical	Full architecture dis-closure to regulator; third-party audit.	Regulator access to live telemetry; adversarial multi- agent sandbox pre-deployment.	Deployment suspension on any anomaly; human re-authorisation required.

      to reect local policy priorities, provided that any deviation from a standard weighting schedule is publicly disclosed.

      An Se score below 0.5 constitutes a presumptive compliance

      v(J1) · v(J2)

      1 2

      RHI(J1, J2)= /lv(J )/l× /lv(J )/l

      (2)

      failure requiring regulatory intervention. A score between 0.5 and 0.74 triggers enhanced monitoring. A score of 0.75 or above represents baseline compliance. These thresholds are calibrated to the severity scale implicit in the TRCF tiers and are intended to provide regulators with a continuous, auditable signal that is more informative than the binary pass/fail outcomes of current pre-approval regimes.

   3. Human-in-the-Loop Accountability Matrix

      The HLAM addresses the duciary duty gap by specify-ing, for each decision category in an AI advisory context, the minimum level of human oversight necessary for the deploying rm to discharge its duciary obligations. Three oversight intensity levels are dened. Level 1 (ex-post review) applies to routine execution decisionsorder routing, intraday rebalancingwhere the decision is individually immaterial and human review within 24 hours is sufcient. Level 2 (pre-authorisation) applies to portfolio-level decisionsstrategic allocation shifts, entry into or exit from asset classeswhere a licensed human advisor must review and approve the AIs recommendation before execution. Level 3 (independent regu-latory clearance) applies to the deployment of novel strategies not previously validated in production, where pre-deployment sandbox testing with regulator access to test results is required.

   4. Regulatory Harmonisation Index

      The RHI is computed as the cosine similarity between pairs of jurisdiction-level regulatory requirement vectors. Each vector has ve dimensions corresponding to the ve core regulatory requirements of the TRCF: pre-trade risk controls (wptc), explainability mandates (wxai), real-time monitring obligations (wrtm), sandbox requirements (wsbx), and du-ciary standards (wfid). For jurisdictions J1 and J2:

      An RHI below 0.7 between any two jurisdictions in which a rm operates is treated as a presumptive arbitrage risk. Firms operating across low-RHI jurisdiction pairs are subject to enhanced reporting requirements and must demonstrate that they apply the stricter of the two regulatory standards to their AI systems regardless of where those systems are domiciled. IOSCO is proposed as the coordinating body for bilateral RHI reviews, consistent with its existing mandate for cross-border market oversight [11].

6. Results and Evaluation

   The proposed framework is evaluated through two com-plementary methods: a structured case study assessment that tests whether the TRCF and extended Se score would have detected or mitigated four documented market disruptions, and a gap-validation analysis that examines the extent to which recent regulatory and judicial developments corroborate the signicance of the ve identied research gaps.

   1. Case Study Evaluation

      Table IV applies the TRCF classication and Se components to four cases. The analysis is counterfactual: it asks what outcome the framework would have predicted or produced had it been in force at the time.

   2. Gap Validation Analysis

      The signicance of the ve identied research gaps is independently corroborated by recent regulatory and judicial developments. For Gap 1, the FCAs 2024 explicit agging of adversarial inter-agent learning dynamics [25] conrms that collusion risk is now a live regulatory concern rather than a theoretical one. For Gap 2, the SECs formal inclusion of AI-washing in its 2024 and 2025 examination priorities

      [9] validates the identication of disclosure misrepresentation

      TABLE IV

      Case Study Evaluation of the Proposed Framework

      Case	Date	Primary Failure	TRCF Class	Key Se Signal	Predicted Interven-tion
      2010 Flash Crash	May 6, 2010	Spoong algorithm triggered self-reinforcing liquidity withdrawal; 1,000-point DJIA decline in minutes.	Class III	Fairness score F 0.10 (predatory order patterns); Systemic risk R 0.95.	Real-time order-to-trade ratio monitoring would have agged anomalous cancellation patterns before liquidity vacuum formed; kill-switch activation.
      Knight Capi-tal	Aug 1, 2012	Deployment of legacy test software caused $440M loss in 45 minutes through unin-tended aggressive or-der execution.	Class I	Accountability score A 0.20 (absent de-ployment verication protocol).	Mandatory pre-trade controls and deployment verication under Class I registration would have blocked unvalidated code from accessing live markets.
      SEC v. Athena Capital	Oct 16, 2014	HFT rm used Gravy algorithm to articially mark closing prices of thousands of NASDAQ-listed stocks.	Class III	F 0.05 (systematic end-of-day price distortion); T 0.30.	Se below 0.5 triggers compliance failure; real-time monitoring of closing-auction order patterns detects systematic distortion.
      SEBI Orders 202223	20222023	Brokers made misleading claims about guaranteed returns from algo strategies; retail investors misled.	Class II / III	Df 0.80 (mate-rial misrepresentation of AI capabilities).	Df penalty drives Se to compliance-failure range; mandatory disclosure verication during registration prevents misrepresentation.

      as an enforcement-priority issue. For Gap 3, ESMAs 2024 statement that nancial institutions bear full responsibility for AI system outputs [10] represents the most authoritative judicial-adjacent conrmation that duciary duty extends to AI advisory systems. For Gap 4, the IMFs 2025 warning about regulatory arbitrage in AI-enabled trading [3] provides multi-lateral institutional conrmation. For Gap 5, the contradiction between ESG mandates and AI energy footprints has been independently identied in the sustainable nance literature [13], [24].

   3. Robustness of the Se Score

   A sensitivity analysis of the Se formula across plausible ranges of weight values (w1 through w6 varied uniformly between 0.10 and 0.25) conrms that the compliance-failure threshold at Se < 0.5 is robust for systems with clear ma-nipulative characteristics: for the 2010 Flash Crash scenario, Se remains below 0.35 across all weight combinations tested. For borderline casessystems with moderate transparency and modest systemic risk scoresthe outcome is more sensitive to weighting, which provides the rationale for requiring public disclosure of the weighting schedule adopted by each regula-tor.

7. Entrepreneurial and Industry Implications

   1. Compliance Cost and the Fintech Startup Ecosystem

      The regulatory requirements contemplated by the TRCF place a disproportionate burden on early-stage ntech ven-tures relative to established institutional participants. SEBI and SEC compliance infrastructureincluding explainability documentation, real-time monitoring systems, and regulatory sandbox participationmay consume up to 20% of an early-stage rms operational budget [8]. The EU AI Acts com-pliance cost of approximately C29,277 per AI product [12] is manageable for a large nancial institution but potentially prohibitive for a startup whose entire AI portfolio may consist of a single system. A tiered compliance cost structurein which Class I requirements are minimally burdensome and Class IV requirements are graduated according to the scale of the rms market presencewould preserve the regulatory intent of the TRCF while reducing the barrier to entry for legitimate innovation.

   2. Investor Appetite for Explainable AI

      The AI-washing enforcement trend has a secondary market effect: venture capital and institutional investors are increas-ingly applying enhanced due diligence to AI capability claims in ntech investment proposals. This is reected in a growing preference for Explainable AI (XAI) architectures, which

      allow the investment thesis to be veried independently of the developers representations. AI startups that can demon-strate TRCF Class II or Class III compliance with strong T and A scorestransparency and accountabilityare likely to command a material valuation premium over comparable rms operating black box strategies, analogous to the ESG premium documented in the sustainable investment literature [24].

   3. Dispute Resolution in AI-Enabled Markets

      The duciary duty gap identied in Section III-C has direct implications for how trading disputes are resolved. Where an AI-driven strategy causes client losses, the rms ability to demonstrate HLAM-compliant oversightdocumented ev-idence of human pre-authorisation for Level 2 decisions and regulator-cleared sandbox results for Level 3 deploymentswill be the central evidentiary question in any subsequent regulatory investigation or civil claim. Firms that have not implemented the HLAM face the prospect of being unable to mount a compliance defence, because they will be unable to demonstrate that adequate human oversight was in place. In this sense, the HLAM is not merely a regulatory obligation but an instrument of litigation risk management.

   4. Strategic IP and Data Governance

   AI trading systems derive their competitive advantage partly fro proprietary training data pipelines and partly from model architecture innovations. Both are increasingly recognised as forms of intellectual property with uncertain legal statusanalogous to the AI-generated IP issues examined in the par-allel literature [23]. Firms operating under the TRCFs Class III and IV requirementswhich mandate training data prove-nance disclosureface a potential tension between regulatory transparency obligations and the protection of commercially sensitive data assets. Legal frameworks for managing this tension, including regulatory sandboxes in which proprietary information is disclosed to regulators under condentiality, represent an important area for further institutional develop-ment.

8. Legal Positioning and Policy Implications

   The TRCF and extended Se score are designed as eviden-tiary and analytical instruments that complement, rather than replace, existing securities law frameworks. They operate at the interface between technical system assessment and legal liability determination, providing the quantitative indicators that existing doctrine lacks. Veriable records of Se scores, HLAM compliance documentation, and RHI calculations con-stitute a form of regulatory evidence base that can support enforcement actions, due diligence processes, and dispute resolution proceedings.

   1. Recommendations for SEBI (India)

      SEBI should establish a centralised Algorithmic Trading Sandbox in which all Class III and Class IV systems must undergo pre-deployment stress-testing, including adversarial

      multi-agent scenarios designed to probe for emergent collu-sive behaviour. The sandbox should be supplemented by a mandatory disclosure regime for alternative data strategies that requires rms to characterise the nature and source of their data inputs without disclosing proprietary model details. SEBIs broker accountability framework should be extended to specify HLAM-equivalent oversight requirements for AI-mediated wealth management and advisory services, address-ing the duciary duty gap in the domestic context.

   2. Recommendations for the SEC (USA)

      The SEC should formalise AI-washing as a sub-category of material misrepresentation under Rule 10b-5, with specic guidance on what constitutes an adequate and veriable AI ca-pability claim in offering documents and marketing materials. The existing MNPI framework should be supplemented with a disclosure requirement for alternative data strategies that relies on the regulatory concept of analytical infrastructure access rather than information possession. The SEC should also engage with IOSCO to develop the RHI framework as a multilateral instrument for identifying and addressing regulatory arbitrage risks across major trading jurisdictions.

   3. Recommendations for ESMA and the EU

      The EU AI Acts High-Risk classication for nancial AI should be supplemented with specic technical standards for the energy-intensity disclosure requirement proposed in this paper, integrating the Ec component of the Se score into the Acts existing conformity assessment obligations. ESMAs supervisory brieng on AI compliance should address the multi-agent collusion risk explicitly and specify that sandbox assessments for Class IV systems must include adversarial multi-agent testing. The EU should use its regulatory lead-ership position to promote RHI-based harmonisation with non-EU jurisdictions through the Commissions equivalence decision process.

   4. Global Coordination

   The most signicant limitation of any single-jurisdiction regulatory initiative is that it creates the arbitrage incentives documented in Gap 4. A durable solution requires multilateral coordination at the level of IOSCO and the Financial Sta-bility Board. The FSBs existing mandate to assess systemic nancial risks provides a natural institutional home for RHI monitoring and for the development of minimum standards for cross-border AI trading oversight. The energy-intensity disclosure requirement has natural synergies with the FSBs climate-related nancial disclosure agenda, providing an ad-ditional impetus for coordinated action.

9. Conclusion

   This paper has examined, in depth, the legal and ethi-cal challenges that arise from the deployment of AI-driven nancial prediction systems across modern markets. The central argument is that existing regulatory frameworkswhile increasingly attentive to single-system manipulation

   and systemic riskhave not adequately addressed a set of structurally important threats: the emergence of autonomous collusive behaviour in multi-agent AI environments; the mis-representation of AI capabilities in investment disclosures; the unresolved allocation of duciary responsibility in AI advisory relationships; the structural incentives for cross-border regu-latory arbitrage; and the contradiction between the nancial sectors ESG commitments and the energy footprint of its AI infrastructure.

   The Tiered Regulatory Compliance Framework proposed here provides a risk-proportionate classication system that aligns regulatory burden with the degree of opacity and systemic potential of each AI trading system. The Extended Ethical Compliance Score integrates transparency, accountabil-ity, fairness, systemic risk, energy intensity, and disclosure delity into a single auditable indicator. The Human-in-the-Loop Accountability Matrix and the Regulatory Harmonisa-tion Index provide operational instruments for the duciary duty and arbitrage gaps respectively. Evaluated against four documented market disruptions, the framework demonstrates consistent predictive and mitigative value.

   The framework is proposed not as a substitute for reg-ulatory judgment but as an evidentiary infrastructure that makes that judgment more defensible and more consistent across jurisdictions. The overarching policy goalpreserving market integrity while accommodating legitimate AI-driven innovationis better served by a framework of this kind than by either of the dominant alternatives: reactive enforcement after the fact or prescriptive prohibition that forecloses inno-vation.

10. Limitations and Future Work

Several limitations of the present analysis warrant acknowl-edgment. The Se formulas weight parameters are normatively determined and require empirical calibration; the sensitivity analysis reported in Section VI-C indicates robustness for extreme cases but not for borderline ones. The RHI is proposed as a conceptual instrument and has not been computed against live regulatory texts; a follow-on empirical study mapping SEBI, SEC, FCA, and ESMA requirements onto the ve-dimensional vector space is needed to validate the 0.7 thresh-old. The HLAMs three-tier oversight structure has not been tested in a live regulatory context and may require renement based on implementation experience.

Future research should pursue three directions. First, the development of Regulator-AI systemsAI agents designed specically to monitor and audit other AI trading systems in real timerepresents the most promising long-term approach to the speed-of-market challenge that human oversight alone cannot address. Second, the energy-intensity disclosure frame-work requires collaboration between nancial regulators and standards bodies to develop a standardised Ec measurement protocol. Third, the multi-agent collusion risk identied in Gap 1 warrants dedicated experimental research using mar-ket microstructure simulations to characterise the conditions

under which independently-operating AI agents converge on collusive equilibria in realistic trading environments.

References

1. U.S. SEC, Staff Report on Algorithmic Trading in U.S. Capital Mar-kets, Washington, DC, 2020.

2. IDC, Worldwide Spending on Articial Intelligence System: A Fore-cast Through 2027, International Data Corporation, 2023.

3. IMF, Regulatory Considerations Regarding Accelerated Use of AI in Securities Markets, Technical Notes and Manuals, vol. 2025, no. 016, Dec. 2025.

4. J. Armour and H. Eidenmu¨ller, Self-Driving Corporations? Harvard Business Law Review, vol. 10, no. 1, pp. 87139, 2020.

5. EU AI Act, Regulation (EU) 2024/1689 of the European Parliament and of the Council on Articial Intelligence, Ofcial Journal of the European Union, 2024.

6. SEC v. Athena Capital Research, LLC, Securities Exchange Act Release No. 34-73369, Oct. 2014.

7. CFTC and SEC, Findings Regarding the Market Events of May 6, 2010, Joint Report, Sep. 2010.

8. McKinsey & Company, The State of AI in 2024: Generative AIs Breakout Year, McKinsey Global Institute, May 2024.

9. SEC Division of Examinations, 2024 Examination Priorities, Washing-ton, DC, 2024; SEC, 2025 Examination Priorities, Washington, DC, 2025.

10. Venable LLP, Articial Intelligence in Investment Management: Reg-ulatory Challenges and Fiduciary Implications, Dec. 2025. [Online].

    Available: https://www.venable.com

11. IOSCO, The Use of Articial Intelligence and Machine Learning by Market Intermediaries and Asset Managers, Consultation Report CR/01/2025, Mar. 2025.

12. European Commission, Impact Assessment of the EU AI Act, 2024;

    H. Laurer, A. Renda, and K. Yeung, Compliance Cost Estimation, 2021.

13. INTL Sustainable Development Observatory, Sustainable Finance and AI: ESG Risk Assessment with Machine Learning, ISDO, Dec. 2025.

14. SEC v. Visionary Private Equity Group, Press Release No. 2010-131, 2010.

15. Knight Capital Americas LLC, SEC Administrative Proceeding No. 3-15570, Oct. 2013.

16. SEBI, Master Circular for Stock Brokers, Mumbai, 2023; SEBI Orders on Algo Trading Manipulation, 202223.

17. E. Calvano, G. Calzolari, V. Denicolo`, and S. Pastorello, Articial Intelligence, Algorithmic Pricing, and Collusion, American Economic Review, vol. 110, no. 10, pp. 32673297, Oct. 2020.

18. Oxford Journal of Financial Regulation, Fintech and the Future of Securities Law, vol. 9, 2023.

19. SSRN, Algorithmic Trading and Market Manipulation: A Legal Anal-ysis, Working Paper, 2024.

20. IEEE Xplore, Deep Learning for Financial Market Prediction: Ethical Considerations, Proc. IEEE Conf. on AI and Finance, 2025.

21. Bank for International Settlements, The Impact of Articial Intelligence on the Financial Sector, BIS Working Papers, 2024.

22. Financial Stability Board, Articial Intelligence and Machine Learning in Financial Services, FSB Report, Nov. 2017.

23. Frontiers in Articial Intelligence, Ethical Theories, Governance Mod-els, and Strategic Frameworks for Responsible AI Adoption, vol. 8, 2025.

24. Future Business Journal, AI-Driven Sustainable Finance: Computa-tional Tools, ESG Metrics, and Global Implementation, Springer Open, Aug. 2025.

25. Sidley Austin LLP, Articial Intelligence in Financial Markets: Sys-temic Risk and Market Abuse Concerns, Client Advisory, Nov. 2025.