Published On: May 25, 2026
Authored By: Chaitanya Saxena
Mahatma Jyotiba Phule Rohilkhand University
Abstract
The use of consumer-grade generative AI tools by employees without authorisation, termed “Shadow AI”,[1] has become a leading vector of corporate legal liability. This article explores how converging regulatory developments will impact the critical implementation phase of global AI governance, particularly the structural instability created by productivity-driven employees bypassing enterprise security controls. It examines how the interaction between the AI literacy mandates of the EU AI Act and India’s Digital Personal Data Protection (DPDP) Act (and its 2025 Rules) renders a passive approach to corporate policy legally untenable. The piece analyses the “triadic liability”[2] framework, encompassing employer, employee, and AI provider, finding that existing jurisprudence has moved towards an employer-responsibility model, with liability arising unless the employer has provided lawful, secure alternatives.
Introduction
In the international business arena, a critical tipping point has been reached. The early phase of enthusiasm about generative AI, marked by organisational tolerance and rapid adaptation, has collided head-on with an imposing regulatory framework. In the initial period of AI adoption, executives focused on developing “High-Level AI Principles” and “Ethics Charters.” Those high-minded instruments have proven inadequate to counter the covert, daily realities of Shadow AI.
Shadow AI refers to the pervasive use of consumer-grade, unsanctioned generative AI tools by employees to perform official duties, without the oversight, security vetting, or contractual protections that enterprise IT governance provides. Current industry benchmarks highlight a significant disconnect between policy and practice: while fewer than half of global firms have officially deployed hardened, enterprise-grade Large Language Models (LLMs),[3] a substantial majority of employees acknowledge using personal accounts to process sensitive company data. This statistical disparity demonstrates that the workforce is moving materially faster than corporate policy, creating an invisible, unmanaged data layer.
The central legal issue of this era is a fundamental shift in how regulators and courts understand the use of artificial intelligence. Where accidental breaches involving AI were once argued before courts as instances of “technical novelty”, Data Protection Authorities and courts have since re-evaluated that position. Shadow AI is now treated as a failure of corporate governance, not an isolated technical incident.
Unlike traditional “Shadow IT”, Shadow AI presents a distinctive legal hazard: recursive data exfiltration.[4] Every interaction with a consumer-grade AI model involves prompts that are frequently fed back into public training sets to refine future model iterations. When employees paste proprietary code, draft sensitive litigation strategies, or input Personal Data to summarise meetings, they are effectively transferring corporate intellectual property and protected personal data into the public domain. This exfiltration occurs in a governance vacuum: because the usage is unofficial, there are no Data Processing Agreements or audit trails to satisfy regulatory logging requirements.
This phenomenon effectively dissolves the corporate perimeter. Regulators now take the position that an organisation which permits Shadow AI to operate unchecked is not merely a victim of employee negligence, but a legally responsible entity that has abdicated its statutory duty to maintain possession and control over its data assets. The resulting governance vacuum has transformed what was once a productivity tool into a primary vector for legal liability and regulatory “dark breaches.”
Legal Analysis
1. The Liability Triangle: Apportioning Responsibility in a “Prompt Leak”
The central analytical problem is the triadic relationship between the enterprise, the individual employee, and the AI developer when a shadow prompt causes a significant data breach or loss of intellectual property.
Employer Liability
Recent decisions have demonstrated that “prohibitory policies”, that is, blanket prohibitions on AI use, are an ineffective legal defence for employers operating under the doctrine of respondeat superior. Judicial reasoning has aligned with the concept of “digital negligence.” Courts have held that where an employer demands rapid output without providing secure, enterprise-grade AI options, it has made a breach foreseeable, if not inevitable. That failure is increasingly interpreted as a breach of the duty to maintain reasonable security safeguards, a position compounded by the failure to implement technical controls or “sanitised gateways.” Employers with deep pockets have accordingly become the primary targets of class-action plaintiffs in data breach litigation.
Employee Error vs. Misconduct
In 2025, many employment contracts characterised unauthorised AI use as “gross misconduct.” The legal picture in 2026, however, is more nuanced. Labour courts are increasingly treating Shadow AI as a tool of operational necessity rather than an act of malice. While an employee may face an internal disciplinary process, imposing personal financial liability for a multimillion-dollar data breach is both legally contentious and practically difficult to enforce. Absent evidence of wilful sabotage, the concept of the “Governance Gap”, that employees resorted to unofficial tools because no compliant alternative was available, provides significant mitigation in proceedings against the individual.
AI Provider Defence
The vast majority of consumer-grade AI Terms of Service are carefully drafted to shield providers from liability. The enforceability of these take-it-or-leave-it arrangements has been upheld in a number of jurisdictions. Where an employee copies confidential data and inputs it into a public model, providers are expected to invoke the “voluntary disclosure” defence: arguing that the user voluntarily submitted confidential data to a public model and thereby forfeited confidentiality, since the Terms of Service permit training on user-submitted content. This leaves the organisation with very limited direct recourse against the AI developer, given that the provider’s contractual relationship was with the individual user, not the enterprise.
2. Conflict with the DPDP Act, 2023 and the 2025 Rules (India)
The notification of the Digital Personal Data Protection Rules, 2025 has elevated Shadow AI from a policy irritant to a statutory problem for Indian “Data Fiduciaries.”[5]
The Control-Liability Paradox
Section 8 of the DPDP Act imposes an obligation on Data Fiduciaries to protect personal data in their possession or control. A closer analysis reveals a structural contradiction: if an employee uses Shadow AI to process customer data, the enterprise loses operational control, the data resides on a third-party server without any Data Processing Agreement in place, yet retains vicarious liability for the resulting breach. This split between factual control and legal responsibility means that the fiduciary bears liability for a breach it could not detect and may not even have been aware of.
Dark Breaches and the Black Hole Effect
The 2025 Rules mandate strict, time-bound breach reporting and granular logging of data processing activities. Shadow AI creates what regulators have termed a “Dark Breach.” Because the processing occurs outside the corporate firewall, there are no logs, no API trails, and no metadata. When a leak occurs via a public LLM, the organisation cannot satisfy its obligations under the right to information provision in Section 11 of the Act, nor can it accurately notify the Data Protection Board of India (DPB) about the extent of the breach. This “Black Hole” effect triggers a secondary violation: failure to report a breach, which carries significantly higher penalties under the Act than the primary breach itself.
3. Comparative Compliance: The EU AI Act’s AI Literacy Mandate
The EU AI Act, whose enforcement provisions began taking effect from August 2025, has set an increasingly influential standard that renders deliberate regulatory ignorance, what might be called the “Ostrich Strategy”[6] , legally indefensible.
AI Literacy as a Legal Obligation
Article 4 of the EU AI Act introduces an AI literacy obligation that represents a significant shift in regulatory approach. It places the knowledge burden on the organisation rather than the individual employee: organisations now have a legal duty to ensure that anyone interacting with AI systems possesses a sufficient baseline understanding of AI-related risks and data privacy implications. This effectively removes the “ignorance defence” from employment disputes arising in early 2026. An organisation seeking to establish that an employee acted unilaterally must be able to demonstrate that it provided documented, substantive AI literacy training to that employee in advance.
From Security to Governance
Shadow AI has evolved from a firewall problem into a systemic governance failure. The United States frames the issue as one of “algorithmic accountability”; India frames it as “Data Fiduciary Responsibility”; the EU links both by requiring that automated decision-making remain subject to meaningful human oversight. The analytical conclusion for 2026 is that the mere presence of Shadow AI within an organisation constitutes prima facie evidence of a failure to meet the minimum international governance standard for managing high-risk AI activities. The corresponding legal development is the erosion of plausible deniability at the executive level.
Supporting Authority
DPDP Rules, 2025[7]
Rule 6 of the DPDP Rules, 2025 requires Data Fiduciaries to implement reasonable security safeguards, including access logs and monitoring mechanisms. Because Shadow AI activity by definition escapes these controls, any Shadow AI incident renders the fiduciary per se non-compliant with the Act.
EU AI Act, 2024[8]
Of particular relevance are the General Purpose AI (GPAI) transparency requirements and the AI literacy obligations under Article 4, the enforcement of which commenced in 2025.
Case Law and Precedents (2025–2026)[9]
Although many 2025 cases remain pending in the appellate process, an emerging “Preservation Order Trend”, illustrated by proceedings in The New York Times Co. v. OpenAI extending into 2025, demonstrates that courts are prepared to compel the production of extensive AI output logs. In a Shadow AI context, organisations that cannot produce these logs in discovery are at risk of adverse rulings based on spoliation of evidence.
Regulatory Policy[10]
The United States Federal Trade Commission’s 2025 enforcement activity targeting “AI-washing”, where organisations publicly disclaim AI-related risk while tolerating Shadow AI internally, and the Department of Justice’s 2026 Guidance on Ephemeral Messaging and AI together make clear that corporations must actively maintain control over their data flows. An organisation that claims a commitment to data protection while permitting a culture of Shadow AI to persist may be exposed to claims of deceptive trade practice.
Conclusion
The legal vacuum surrounding Shadow AI is rapidly closing, and not in the organisation’s favour. Developments in 2025 and 2026 demonstrate that regulators, particularly in India and the EU, treat unauthorised AI use not primarily as individual misconduct but as systemic governance failure on the part of the enterprise.
The triadic liability framework examined in this article points consistently in one direction: towards organisational accountability. Employers who provide no compliant alternative will bear the weight of respondeat superior liability. Employees who fill the resulting governance gap through Shadow AI use will receive increasing judicial sympathy. And AI providers will continue to shelter behind contractual terms of service designed specifically to insulate them from liability.
The anticipated emergence of “Safe Harbour” provisions, whether through certified enterprise AI gateways attracting liability exemptions, or through enforcement action against organisations that prohibit AI without providing compliant alternatives, signals a fundamental shift in the regulatory posture. The transition from “Do Not Use AI” to “Use Only This Approved AI” is no longer a governance recommendation. It is an emerging legal requirement.
References
[1] Shadow AI refers to instances where employees use personal AI tools (such as free-tier ChatGPT) for work purposes without the knowledge of their employer, in the process “leaking” company information into the AI system’s memory or training data.
[2] The “triadic liability” framework describes the three-party responsibility structure: the employer (who failed to provide a secure tool), the employee (who used an unsanctioned tool), and the AI provider (whose platform received the data).
[3] A Large Language Model (LLM) is an AI system trained on large volumes of text data to understand, process, and generate human language.
[4] Recursive data exfiltration describes a cycle in which data leaked to an AI system is subsequently used to train that system, making the disclosure permanent and self-reinforcing.
[5] A Data Fiduciary is a person or organisation that determines the purpose and means of processing personal data, as defined under the Digital Personal Data Protection Act, 2023.
[6] The “Ostrich Strategy” refers to the practice of deliberately ignoring a known risk rather than addressing it, named after the popular belief that ostriches bury their heads in the sand.
[7] Press Information Bureau, Govt. of India, Government Notifies Digital Personal Data Protection Rules, 2025 (Nov. 14, 2025), https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2190014; Ministry of Electronics & Information Technology, Digital Personal Data Protection Rules, 2025 (Nov. 14, 2025), https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc20251117695301.pdf; India’s New Digital Personal Data Protection Rules, 2025: A Detailed Reading, The Leaflet (Nov. 15, 2025), https://theleaflet.in/explainer/indias-new-digital-personal-data-protection-rules-2025-a-detailed-reading.
[8] Karen Gallagher, Isabel Humburg & Peter Watts, EU AI Act and AI Literacy, IBEC (Aug. 5, 2025), https://www.ibec.ie/connect-and-learn/insights/insights/2025/08/05/ibec-networks-eu-ai-act-and-ai-literacy; AI Literacy, Punter Southall Law, https://puntersouthall.law/insights/ai-literacy/.
[9] Law Offices of Salar Atri Zadeh, The Year in AI Law: 2025’s Biggest Legal Cases and What They Mean for 2026, Internet Lawyer Blog (Dec. 29, 2025), https://www.internetlawyer-blog.com/the-year-in-ai-law-2025s-biggest-legal-cases-and-what-they-mean-for-2026/.
[10] Rachel Anderson, Shadow AI: The Governance Gap That Business Can’t Afford to Ignore, CMS (Feb. 17, 2026), https://cms.law/en/gbr/publication/shadow-ai-the-governance-gap-that-businesses-can-t-afford-to-ignore.
