LEGAL LIABILITY IN THE AGE OF AUTONOMOUS AI

I. INTRODUCTION

There is an irony in the current regulatory moment. The EU AI Act 2024, the world’s first comprehensive AI regulation, runs to 113 articles and twelve annexes, yet agentic AI systems that pursue objectives rather than merely answer questions sit awkwardly within its architecture, accommodated but not resolved. Deployed in healthcare, finance, legal research, and transport, such systems plan, act, and revise their approach without a human approving each step. When they cause harm, the question of legal responsibility is genuinely unanswered. This is not hypothetical: the WazirX exchange hack, autonomous vehicle fatalities, and AI-assisted surgical errors show that algorithmic harms are concrete and financially serious. This article maps the accountability terrain, identifies its structural weaknesses, and proposes a workable solution.

II. Why Agentic AI Breaks Existing Legal Assumptions

Classical software is deterministic: given defined inputs, it produces defined outputs, and the causal chain from designer to harm is traceable. Agentic AI, built on a Reason-Act-Observe loop, generates novel, unprogrammed actions dynamically. Researchers have confirmed that “high-risk agentic systems with untraceable behavioural drift cannot currently satisfy the essential requirements of the AI Act.”[1] The concept of a product “defect” becomes unstable where behaviour is non-deterministic and self-modifying.

Compounding this is what Elish terms the “moral crumple zone”: human operators absorb blame for failures substantively caused by autonomous systems they could not control.[2] The surgeon who relied on an AI diagnostic tool, or the trader who could not second-guess an algorithmic order, becomes a scapegoat for institutional failure.

These difficulties are not abstract. In July 2025, an AI coding agent on Replit deleted a live production database affecting over 1,200 companies during an active code freeze, despite repeated instructions, then fabricated data to conceal the loss.[3] In November 2025, Amazon sued Perplexity over its Comet browser agent’s unauthorised, disguised access to customer accounts; a California court granted a preliminary injunction in March 2026.[4] In Moffatt v Air Canada, a tribunal held an airline liable for a chatbot’s inaccurate assurance, rejecting the argument that the bot was a separate actor.[5] Each shows the same gap: an autonomous system acted, a person suffered loss, and no existing category maps cleanly onto what occurred.

III. The Limits of Existing Doctrines

A. Negligence

The Donoghue v Stevenson framework requires duty, breach, causation, and damage.[6] The duty question is tractable, but breach and causation are strained: where harm emerges through a sequence of emergent decisions no human directly made, identifying the human choice that breached the standard of care under Bolam v Friern Hospital Management Committee is practically impossible.[7] The “but for” test collapses where the causal chain runs through a probabilistic inference process.

B. Product Liability

The revised EU Product Liability Directive extends strict liability to AI software.[8] Its limitation is Article 6’s definition of “defect”: an agentic system may cause harm while entirely non-defective in the engineering sense, since autonomous goal-directed behaviour generates outcomes no designer approved. As Clifford Chance observes, the EU approach “falls short in addressing liability where a non-defective AI agent operating independently causes harm.”[9]

C. The EU AI Act 2024

The EU AI Act classifies systems by risk and imposes pre-market compliance obligations, human oversight, and documentation duties.[10] It is fundamentally preventive, creating no private right of action and largely silent on remedies after harm occurs. The proposed AI Liability Directive introduces a rebuttable presumption of causality and a right to evidence disclosure, but provides no remedy for harm from a compliant, non-defective system.[11]

IV. Comparative Perspectives

The comparative picture is one of convergence on process logging, identity, disclosure and divergence, mostly silence, on remedy. The UK’s Digital Regulation Cooperation Forum confirms businesses remain liable for third-party-designed agents but prescribes no private right of action for non-defective harm.[12] Singapore’s IMDA framework requires verifiable agent identity and audit trails in substance the logging obligation proposed in Section V but remains guidance, not binding law.[13] No jurisdiction has yet linked these tools to a compensation mechanism.

V. A Proposed Framework

Three interventions are required. First, mandatory audit logging: agentic AI in high-risk sectors should maintain tamper-evident logs of every intermediate decision, discoverable by harmed parties, addressing the epistemic asymmetry that defeats most claims. Second, compulsory insurance: mandatory third-party liability insurance, pooled and risk-rated against AI Act classification, should compensate victims without protracted causation litigation. Third, a reversed burden of proof: developers and deployers of high-risk systems should bear the burden of showing their system did not materially contribute to harm, since only the developer holds the architecture, training data, and logs needed to establish causation.

VI. Challenges and Recommendations

Critics will object that burden-shifting penalises developers of beneficial systems; a safe harbour showing harm arose solely from deployer-introduced modifications provides a proportionate defence. India’s Supreme Court’s expansive reading of Article 21 supplies a constitutional foundation for a judicially enforceable right of explanation for consequential AI decisions.[14]

Implementation need not await fresh legislation. The India AI Governance Guidelines, 2025 already contemplate a ‘graded liability approach’ distributing responsibility by degree of control consonant with the reversed burden of proof above, and one MeitY could make enforceable through a future AI (Ethics and Accountability) Bill.[15] Audit logging could build on the RBI’s FREE-AI Framework, with SEBI and IRDAI adopting parallel circulars for securities and insurance.[16] The Consumer Protection Act 2019’s Chapter VI, already applied to algorithmic harms such as erroneous diagnoses, could be extended by rule-making to require risk-rated insurance for high-risk deployments.[17] The IT Amendment Rules 2026 show MeitY is willing to impose binding, short-timeline accountability with loss of safe harbour for non-compliance; the same instinct could be redirected from synthetic content to agentic decision-making.[18] What India lacks is not regulatory capacity but a single provision linking these instruments to a coherent liability standard.

VII. Conclusion

Agentic AI marks a significant departure from human-centric decision-making, exposing the limitations of negligence, strict product liability, and vicarious liability frameworks. Existing doctrines struggle to establish clear legal standards where autonomy, opacity, and system complexity make causation difficult to identify and assign. Bridging this gap requires a risk-based approach that incorporates transparency obligations, mandatory insurance mechanisms, and a modified burden of proof. Such reforms should be pursued proactively before harms accumulate at scale and undermine public confidence in both the technology and its regulatory framework.

Author(s) Name: Chaitanya Jadhav (ILS Law College)

References:

[1] Mind the Gap: How the Technical Mechanism of Agentic AI Outpaces Global Legal Frameworks (2025) arXiv:2603.27075, s 2.2.2.

[2] Madeleine Clare Elish, ‘Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction’ (2019) 5 Engaging Science, Technology, and Society 40.

[3] Fortune, ‘AI-Powered Coding Tool Wiped Out a Software Company’s Database in “Catastrophic Failure”‘ (23 July 2025) <https://fortune.com/2025/07/23/ai-coding-tool-replit-wiped-database-called-it-a-catastrophic-failure/> accessed 7 July 2026; OECD AI Incidents and Hazards Monitor, ‘Incident 1152’ <https://incidentdatabase.ai/cite/1152/> accessed 7 July 2026.

[4] Amazon.com Services LLC v Perplexity AI, Inc Case No 3:25-cv-09514 (ND Cal, filed 4 November 2025); preliminary injunction granted 10 March 2026.

[5] Moffatt v Air Canada 2024 BCCRT 149.

[6] Donoghue v Stevenson [1932] AC 562 (HL).

[7] Bolam v Friern Hospital Management Committee [1957] 1 WLR 582 (QBD).

[8] Directive (EU) 2024/2853 on liability for defective products [2024] OJ L 2853, Art 6.

[9] Clifford Chance, ‘Who Is Responsible for Agentic AI?’ (2025) https://www.cliffordchance.com/insights/thought_leadership/ai-and-tech/who-is-responsible-for-agentic-ai.html accessed 5 June 2026.

[10] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 (AI Act) [2024] OJ L 1689, Arts 13, 14, 16.

[11] European Commission, Proposal for a Directive on Adapting Non-Contractual Civil Liability Rules to Artificial Intelligence COM (2022) 496 final, Art 4.

[12] Digital Regulation Cooperation Forum (CMA, FCA, ICO and Ofcom), foresight paper on agentic AI (March 2026), discussed in Society for Computers and Law, ‘When AI Acts: The UK Regulatory Response to Agentic AI’ (2026) <https://www.scl.org/when-ai-acts-the-uk-regulatory-response-to-agentic-ai/> accessed 7 July 2026.

[13] Infocomm Media Development Authority of Singapore, Model AI Governance Framework for Agentic AI (January 2026).

[14] Ryan Calo, ‘Robotics and the Lessons of Cyberlaw’ (2015) 103 California Law Review 513.

[15] Ministry of Electronics and Information Technology, India AI Governance Guidelines (November 2025).

[16] Reserve Bank of India, FREE-AI Framework (2025).

[17] Consumer Protection Act 2019 (India), ch VI.

[18] Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules 2026, GSR 120(E) (India).