The Evolution of Liability and Transparency Frameworks for Autonomous Artificial Intelligence: A Comparative Analysis of the EU AI Act and Emerging Global Regimes

Authored By: Sefali Jaiswal

Mahatma Gandhi Kashi Vidyapeeth, Varanasi

Introduction

The rise of autonomous artificial intelligence (AI) agents has forced legal systems worldwide to confront questions that were once confined to science fiction. Unlike the deterministic software of the late twentieth century, today’s AI systems are dynamic, adaptive, and capable of producing outputs that even their developers cannot fully anticipate.[1] This unpredictability challenges the very foundations of tort law and statutory liability, which traditionally rely on a clear causal chain linking human action to harm.[2]

Consider the healthcare sector. Imagine an AI diagnostic tool misclassifying a malignant tumor as benign because its training data contained subtle biases.[3] The physician who relied on the tool may argue that they acted reasonably by using a certified medical device, while the developer may claim that the system’s emergent behavior was unforeseeable. Courts are left in a quandary: should liability attach to the physician, the developer, or both?[4] Similar dilemmas arise in financial markets, where autonomous trading algorithms can trigger cascading losses across jurisdictions in mere milliseconds.[5]

These examples illustrate a broader systemic dilemma: how can responsibility be allocated in a way that protects victims without stifling innovation? The stakes are high. If liability rules are too rigid, they risk discouraging investment in AI research and deployment. If they are too lax, they leave individuals and institutions vulnerable to harms with no clear path to remedy.[6]

This article undertakes a comparative analysis of contemporary legal mechanisms designed to regulate autonomous AI harms. Special emphasis is placed on the European Union’s regulatory shifts, particularly the enforcement of Article 50 transparency requirements of the EU AI Act on August 2, 2026.[7] These provisions mark a decisive move toward risk‑tiered, ex‑ante regulation, contrasting sharply with the sector‑specific, ex‑post models emerging in common law jurisdictions such as the United States and India.[8]

Ultimately, this paper argues that addressing the “black box” accountability gap requires a hybrid regulatory architecture — one that couples strict transparency mandates with robust, specialized ex‑post liability frameworks.[9] Transparency alone cannot solve the attribution crisis; nor can liability rules that fail to account for the evolving nature of autonomous code. A balanced approach, combining upfront disclosure with backend accountability, offers the clearest path forward for harmonizing global AI governance.

The Transparency Mandate: Deconstructing Article 50 of the EU AI Act

On August 2, 2026, the global digital economy reached a turning point. The European Union’s Artificial Intelligence Act, long debated and refined, finally brought its core transparency provisions — Article 50 — into full force.[10] While the broader enforcement timelines for “high‑risk” AI use cases were deferred until December 2, 2027 under the Digital Omnibus Regulation,[11] Article 50’s transparency rules apply immediately, regardless of whether a system is classified as high‑risk. This reflects the EU’s conviction that transparency is not a luxury but a baseline requirement for trust in digital systems.

Operational Arbitrage: Providers vs. Deployers

Article 50 draws a sharp line between providers (entities that design and commercialize AI systems) and deployers (entities that operate those systems under their organizational authority).[12] This distinction matters because it allocates responsibility differently depending on whether one builds the system or merely uses it.

For providers, obligations are proactive. They must ensure that interactive AI systems disclose their non‑human nature at the very start of engagement.[13] In practice, this means that when a consumer begins chatting with a customer service bot, the system must immediately and clearly state that it is software, not a human agent. The disclosure must be distinct and accessible.

Providers of generative systems face an even stricter requirement: they must embed synthetic provenance tracking into all AI‑generated content.[14] This involves inserting persistent, machine‑readable watermarks and metadata into text, images, audio, or video produced by the system. The goal is to make synthetic content traceable, reducing the risk of deception, misinformation, or manipulation.

Deployers, by contrast, bear obligations tied to how they publish or disseminate AI outputs. If they release text or audiovisual media that resembles real persons, objects, or historical events — in other words, deepfakes — they must conspicuously label the content as artificially manipulated.[15]This requirement directly addresses the growing concern that deepfakes can distort public discourse, undermine elections, or damage reputations.

Public Interest and Deepfake Carve‑Outs

The EU’s approach is not absolute. Article 50 introduces a carve‑out for content that has undergone “substantive human editorial review.”[16] This exemption raises thorny questions: how much human intervention is enough to transform algorithmic output into protected speech? If a journalist edits an AI‑generated draft before publication, does that count as substantive review? Or must the human contribution fundamentally alter the AI’s output?

This carve‑out reflects the EU’s attempt to balance transparency with freedom of expression. By allowing human‑reviewed content to escape mandatory labeling, regulators acknowledge that not all AI‑assisted media is deceptive. Yet the ambiguity of “substantive” review may invite litigation, as courts and regulators struggle to define the threshold.

Transparency as a Prerequisite for Trust

Underlying these rules is a broader philosophy: transparency is the foundation of trust in AI. By requiring disclosure and provenance tracking, the EU hopes to mitigate risks of manipulation while empowering citizens to make informed choices.[17] The transparency mandate is not merely about compliance; it is about reshaping the digital public sphere to ensure that individuals know when they are engaging with machines and when content has been artificially generated.

In this sense, Article 50 represents more than a regulatory requirement. It is a cultural statement: in the age of autonomous AI, trust must be earned through openness.

III. Comparative Paradigms: The Common Law Approach vs. Civil Risk‑Tiers

The European Union’s approach to AI regulation is rooted in a centralized philosophy: precaution first, innovation second.[18] By enforcing transparency obligations extraterritorially, the EU seeks to reshape global supply chains and ensure that trust in AI is not contingent on geography. This stands in sharp contrast to the fragmented frameworks of common law jurisdictions such as the United States and India, where regulation is piecemeal, sector‑specific, and often reactive.

The United States: Sectoral Decentralization

The United States has resisted adopting a singular federal AI statute. Instead, it relies on a patchwork of state‑level data laws and sector‑specific federal oversight.[19]For example, the Food and Drug Administration (FDA) monitors autonomous diagnostic software using traditional medical device parameters,[20] while the Federal Trade Commission (FTC) addresses deceptive AI practices under consumer protection statutes. Civil injuries caused by algorithmic errors, however, are litigated through state tort doctrines.[21]

This decentralized model creates hurdles for plaintiffs. Under American products liability doctrine, victims must show proximate causation between a developer’s code and an autonomous algorithmic choice.[22]Yet modern AI systems are designed to adapt dynamically, making it nearly impossible to trace a single line of code to a harmful outcome. Victims of AI harm often face a labyrinth of technical architecture and legal ambiguity, leaving them without a clear path to remedy.

The result is a system where accountability depends heavily on the sector in which harm occurs.

India: Navigating the Accountability Gap

India’s regulatory approach sits at a crossroads. The country relies on statutes such as the Information Technology Act and principles of vicarious corporate liability.[23] Yet civil courts lack specialized expertise to adjudicate disputes involving machine learning systems.[24]Judges trained in traditional doctrines of negligence and product liability often struggle to apply these frameworks to autonomous code.

Legal scholars in India have proposed the creation of a dedicated AI Liability Tribunal, equipped with technical auditors and specialized judges.[25] Such a tribunal would combine upfront risk assessments with backend liability rules, ensuring accountability across the software lifecycle. Importantly, this model avoids granting autonomous code independent legal personhood — a step that many scholars warn could undermine human accountability.

India’s approach reflects both caution and pragmatism. By considering specialized tribunals, the country acknowledges that traditional courts may not be equipped to handle the complexity of AI disputes. At the same time, it resists the temptation to anthropomorphize AI systems, keeping liability firmly within the human and corporate domain.

Civil Risk‑Tiers vs. Common Law Fragmentation

The contrast between the EU and common law jurisdictions is stark. The EU’s risk‑tiered model imposes uniform obligations across sectors, ensuring that transparency and accountability are consistent regardless of context.[26] The United States and India, by contrast, rely on sector‑specific or tribunal‑based approaches that may leave victims vulnerable depending on where harm occurs.

This divergence highlights a broader philosophical divide: the EU views AI regulation as a matter of public trust requiring centralized safeguards, while common law jurisdictions treat AI harms as extensions of existing doctrines, to be addressed case by case. Neither approach is perfect. The EU risks overregulation that could stifle innovation, while the United States and India risk under‑regulation that leaves victims without remedies.

The Attribution Crisis: Deconstructing Tort Theory in the Era of Autonomy

The most profound legal challenge posed by autonomous AI agents is the breakdown of traditional tort categories. For centuries, courts have distinguished between product defects and conduct errors.[27] A defective product — say, a faulty brake pad — triggers strict liability for the manufacturer. A negligent act — such as a driver failing to stop at a red light — invokes fault‑based liability. But autonomous AI systems blur this line, producing harms that cannot easily be classified as either product or conduct.

Failure of Negligence

Negligence doctrine relies heavily on foreseeability.[28] A defendant is liable if they failed to anticipate a risk that a reasonable person would have foreseen. Yet machine learning models often generate outputs that are not only unforeseen but arguably unforeseeable. Consider the phenomenon of “hallucinations” in generative AI, where systems invent facts or produce nonsensical outputs without any clear trigger.[29] If programmers could not reasonably predict such behavior, fault localization becomes nearly impossible. Courts are left asking: who is negligent when no human actor could have foreseen the harm?

This failure of negligence doctrine exposes a deeper problem: tort law was built for human actors, not autonomous systems. When machines act in ways that defy human foresight, negligence loses its explanatory power.

Product Liability Mismatch

The European Union’s updated Product Liability Directive treats software as a tangible product.[30] This allows victims to sue developers under strict liability, without proving negligence. Yet strict liability struggles when applied to systems that evolve after deployment.[31]A self‑learning algorithm may adapt its behavior in ways the developer never coded. Is the harm a defect in the original product, or a consequence of autonomous adaptation?

This mismatch reveals the limits of treating AI as a static product. Unlike a brake pad, which either works or fails, AI systems are dynamic. They change over time, sometimes in ways that even their creators cannot anticipate. Strict liability, designed for fixed products, falters when confronted with evolving code.

The Attribution Gap

Together, the failure of negligence and the mismatch of product liability create what scholars call the attribution gap.[32] Victims of AI harm struggle to identify a responsible party, while developers and deployers deflect blame by pointing to the system’s autonomy. Courts, meanwhile, lack doctrinal tools to bridge the gap.

This attribution crisis underscores the need for new legal frameworks. Traditional tort theory cannot fully account for autonomous systems that act without direct human intervention. Without reform, victims will continue to face barriers to remedy, and accountability will remain elusive.

Strategic Legal Blueprints for Global Harmonization

As autonomous AI systems proliferate across borders, the risk of regulatory fragmentation grows. A medical diagnostic tool developed in Europe may be deployed in India, while a trading algorithm coded in California may trigger losses in Tokyo. Without harmonized frameworks, victims face inconsistent remedies, and developers confront conflicting compliance obligations. To bridge this divide, scholars and policymakers have proposed several strategic blueprints for global harmonization.

Mandatory Strict Liability for High‑Risk Operators

One proposal is to impose mandatory strict liability on entities deploying high‑risk AI systems.[33] Under this model, deployers would bear responsibility for harms caused by their systems, regardless of fault. To balance the burden, mandatory insurance schemes would ensure that victims receive compensation without protracted litigation.[34]

This approach mirrors existing regimes in areas such as aviation and nuclear energy, where operators are held strictly liable due to the catastrophic risks involved. Applying similar principles to high‑risk AI — such as autonomous medical diagnostics or financial trading systems — ensures that victims are not left without remedies simply because causation is difficult to prove.

Evidentiary Presumptions and Digital Black Boxes

Another blueprint involves strengthening evidentiary rules through digital black boxes.[35] Regulators could require AI systems to maintain tamper‑proof operational logs, recording inputs, outputs, and decision pathways. If providers fail to produce clear audits, courts could apply rebuttable presumptions against them, shifting the burden of proof.[36]

This evidentiary innovation addresses the attribution gap by ensuring that victims are not disadvantaged by the opacity of AI systems. Just as aviation relies on flight recorders to reconstruct accidents, AI governance could rely on digital black boxes to reconstruct algorithmic decisions.

Cross‑Border Regulatory Sandboxes

A third proposal emphasizes cross‑border regulatory sandboxes.[37] These sandboxes would allow developers to test AI systems under harmonized conditions across multiple jurisdictions before scaling globally. By proving compliance in a sandbox, developers could streamline approval processes and reduce the risk of conflicting regulations.

Sandboxes also foster collaboration among regulators, enabling them to share best practices and align standards. For example, a sandbox jointly operated by the EU, the United States, and India could establish common protocols for transparency, liability, and risk assessment. This would not only reduce compliance costs but also promote trust in AI systems across borders.

Toward a Hybrid Framework

Taken together, these blueprints suggest a hybrid framework: strict liability for high‑risk operators, evidentiary presumptions supported by digital black boxes, and cross‑border sandboxes for harmonization.[38] Such a framework balances victim protection with innovation, ensuring that accountability does not come at the expense of progress.

The challenge lies in implementation. Strict liability requires careful calibration to avoid discouraging deployment. Digital black boxes demand technical standards that prevent tampering while respecting privacy. Sandboxes require unprecedented cooperation among regulators with divergent philosophies. Yet despite these hurdles, the pursuit of harmonization is essential. Without it, AI governance risks becoming a patchwork of conflicting rules that undermine both justice and innovation.

Conclusion

The enforcement of the EU AI Act’s transparency rules in August 2026 marked a decisive shift in the global governance of artificial intelligence.[39] For the first time, transparency was treated not as an optional safeguard but as a mandatory baseline for trust in digital systems. By requiring disclosures, provenance tracking, and deepfake labeling, the EU signaled that accountability must begin at the point of interaction, not after harm has occurred.

Yet transparency alone cannot resolve the deeper attribution crisis. As Sections IV and V demonstrated, traditional tort doctrines falter when applied to autonomous systems that act unpredictably. Negligence fails when harms are unforeseeable, and strict liability struggles when products evolve after deployment.[40]Victims of AI harm are left navigating a legal landscape that was never designed for machines capable of independent decision‑making.

This gap underscores the need for hybrid frameworks. Transparency provides the ex‑ante safeguard, ensuring that individuals know when they are engaging with AI and when content has been artificially generated. Liability rules, meanwhile, provide the ex‑post remedy, ensuring that victims are compensated when harms occur.[41]Together, these mechanisms form a balanced architecture: openness at the front end, accountability at the back end.

Global harmonization is the next frontier. Without it, AI governance risks becoming a patchwork of conflicting rules that undermine both justice and innovation. Mandatory strict liability for high‑risk operators, evidentiary presumptions supported by digital black boxes, and cross‑border regulatory sandboxes offer promising blueprints.[42] But implementation will require unprecedented cooperation among jurisdictions with divergent philosophies. The EU’s centralized model must find common ground with the United States’ sectoral approach and India’s tribunal‑based proposals.

Ultimately, the challenge is not merely legal but cultural. AI systems are reshaping the boundaries of human agency, forcing societies to reconsider what accountability means in an age of autonomy. If law fails to adapt, the gap between technological capability and legal remedy will widen, leaving victims unprotected and innovation unchecked.[43]

The clearest path forward lies in embracing hybridity: transparency plus liability, precaution plus remedy, local safeguards plus global harmonization. By coupling these elements, regulators can ensure that AI development proceeds responsibly, victims are not left without recourse, and trust in digital systems is preserved.[44]

Bibliography (Bluebook Style)

Statutes and Regulations

Regulation (EU) 2024/1689, art. 50, 2024 O.J. (L 2024/1689).

Digital Omnibus Regulation, 2026 O.J. (L 194).

Council Directive 2024/2853, 2024 O.J. (L 2853).

Restatement (Third) of Torts: Products Liability § 2 (Am. Law Inst. 1998).

Government Documents

European Commission, Guidelines on Transparency Under the Artificial Intelligence Act, COM (2026) 450 final (July 20, 2026).

Secondary Sources

Formulating Autonomous Agency Liability: Comparative Approaches to Algorithmic Harm, 39 Harv. J.L. & Tech. 204 (2026).

¹ Formulating Autonomous Agency Liability: Comparative Approaches to Algorithmic Harm, 39 Harv. J.L. & Tech. 204, 215–18 (2026).

[2] Id.