AI And Its Impact On Legal Privilege

AI And Its Impact On Legal Privilege

A. Introduction

1. The rapid proliferation of artificial intelligence tools has generated pressing questions regarding the continued integrity of legal professional privilege.
2. Generative AI platforms such as ChatGPT, Gemini, Claude, etc. have become integral to the workflows of lawyers globally, generally employed for research, drafting, strategic analysis, and document review. Indian practitioners are no exception to this trend. With the increasing use of generative AI, the foundational doctrines of privilege and confidentiality have been placed under significant strain when legal advice is generated by a lawyer or a client through an AI platform.
3. The central question is whether disclosure of privileged information to an AI platform constitutes disclosure to a third party sufficient to waive legal professional privilege.
4. Indian law, deriving its framework of privilege under Sections 132 to 134 of the Bharatiya Sakshya Adhiniyam, 2023 (“BSA”), has not yet confronted these questions directly. In this article, we consider what some of these possible answers could be.

B. Defining Privilege in India

5. In its simplest meaning, legal privilege is “where legal advice of any kind is sought from a professional legal adviser in his capacity as such, the communications relating to that purpose, made in confidence by the client, are at his instance, permanently protected from disclosure by himself or by the legal adviser, except when the protection be waived.”¹
6. Section 132 of the BSA prohibits any advocate from disclosing, at any time, any communication made to him in the course and for the purpose of his employment.² The Section further extends this protection to interpreters and the clerks or servants of advocates. Section 134 confers upon the client the right to refuse disclosure of confidential communications with his legal professional adviser.
7. Notably, the BSA employs the term “advocate,” replacing the earlier references to “barrister, attorney, pleader or vakil” found in the repealed Indian Evidence Act, 1872. Accordingly, for the purposes of Sections 132-134 of the BSA, the term “advocate” refers only to persons enrolled under the Advocates Act, 1961 and privilege attaches exclusively to communications made between a client and such an enrolled advocate in the course of professional engagement.
8. Pertinently, the Supreme Court has now settled the position of in-house counsel. In the case In Re: Summoning Advocates who give legal opinion or represent parties during investigation of cases, 2025 INSC 1275, a three-Judge Bench held that in-house counsel are not entitled to the privilege under Section 132, since their salaried employment takes them outside the definition of an advocate practising in courts; they retain only the narrower protection of Section 134, which cannot be claimed for communications between the employer and the in-house counsel themselves. The judgment also holds that the Section 132 privilege belongs to the client, may be invoked by the advocate on the client’s behalf, extends to non-litigious and pre-litigation engagements, and is anchored constitutionally in Articles 19(1)(g), 20(3), 21 and 22(1) of the Constitution. The in-house holding has a consequence that deserves emphasis in the present context: decisions about deploying AI within an enterprise are made precisely by in-house legal teams, whose own communications with their employer now enjoy no statutory privilege at all. The privilege position of a general counsel’s office experimenting with AI tools is therefore categorically worse than that of a law firm doing the same.
9. In foreign jurisdictions, privilege is similarly conceptualized but structured differently. Under English law, legal professional privilege is split into (i) legal advice privilege - protecting confidential communications between a lawyer and client made for the dominant purpose of giving or receiving legal advice, and (ii) litigation privilege - which covers communications between the client, their lawyer, and third parties where the dominant purpose is litigation that is in progress, pending, or reasonably in prospect.

C. Impact of AI on the doctrine of legal privilege

10. The advent of generative AI introduces a fundamental tension into the architecture of privilege. When a lawyer inputs client information into a generative AI platform, that communication is no longer confined within the lawyer-client relationship. The AI platform acts as a “digital stranger” that is not regulated by any legal professional bodies, not bound by any contractual obligation to the client, and operating under terms of service that may permit the collection, retention, and onward disclosure of user inputs. This concern is most acute where publicly available AI platforms are used, as opposed to enterprise or closed-source systems operating under contractual confidentiality obligations. Disclosing information to a public AI platform may be no different from disclosure to an unregulated outsider, and so risks being treated as a waiver.³
11. The contractual terms on which AI platforms process user data now arrange themselves into a recognisable ladder, and the privilege analysis maps onto it rung by rung. At the bottom sit free consumer tools, whose terms typically permit retention of inputs, use for model training, and disclosure in defined circumstances. Above them sit consumer tiers offering training opt-outs; then enterprise and API offerings under negotiated terms, including contractual confidentiality, prohibitions on training with customer data, limited or zero retention, and audit rights; and at the top, private deployments within the firm’s or client’s own infrastructure. The higher the rung, the closer the arrangement stands to the outsourced service providers, translation agencies and e-discovery platforms whose routine handling of privileged material has never been thought to waive anything. A further feature of the technology sharpens the inquiry: a provider’s deletion promise is subject to forces external to the contract.⁴ Confidentiality, for the purposes of privilege, must therefore be assessed against the provider’s entire legal exposure, and not against its marketing.
12. Several distinct risks arise:
    1. Confidentiality: publicly available AI platforms typically reserve the right to collect user inputs, use them for training, and disclose them to third parties including government authorities;
    2. Identity of the adviser: under Section 132 of the BSA, privilege attaches to communications made to an “advocate.” An AI platform is not an advocate, is not enrolled under the Advocates Act, and expressly disclaims the provision of legal advice;
    3. Purpose: where a client communicates directly with an AI tool without the direction of counsel, the communication may not be made for the purpose of obtaining legal advice within the ingredients of privilege;
    4. Waiver: even if privileged material was previously communicated to counsel, sharing that material with a public AI platform may constitute a voluntary waiver, destroying the privilege that previously attached.
13. These risks, however, distribute very differently across two situations which must be kept distinct, because they attract opposite answers. The first is the client alone with the machine: a litigant who, without the direction of counsel, uses a public AI platform to analyse his position or prepare his strategy. Such communications satisfy none of the ingredients of Sections 132 or 134; there is no advocate in the loop, and Indian law recognises no privilege for a person’s self-help legal preparation. Heppner, discussed below, would be decided identically in India. The second situation is the advocate using AI in the service of the client, where the privileged communication already exists and the question is whether its transmission to the platform destroys confidentiality. Here the statutory text supplies the frame: Section 132 itself extends the obligation of non-disclosure to interpreters and to the clerks and servants of advocates, recognising that the privileged circle has always included the non-advocate instrumentalities through which the advocate’s function is performed. Modern practice has applied the same logic, without controversy, to the photocopying vendor, the translation agency and the e-discovery platform hosting privileged review sets. The question posed by generative AI is whether it is the newest member of this line, and the answer should turn on whether its terms of processing replicate what those instrumentalities have always offered: confidentiality, control, and use of the material for the client’s purpose alone.

D. Foreign Jurisprudence

14. The recent authority on this conundrum is United States v. Heppner (25 Cr. 503 JSR). The defendant, indicted on charges of securities fraud and wire fraud, had used the generative AI platform Claude to prepare 31 documents outlining his defense strategy after receiving a grand jury subpoena. The FBI seized these documents during execution of a search warrant. The court held that the AI-generated documents were not protected by attorney-client privilege or the work product doctrine for several reasons: (i) the communications were not between the defendant and his attorney since Claude was not an attorney; (ii) the communications were not confidential, as Anthropic’s privacy policy expressly permitted collection of user inputs and outputs, use of that data for training, and disclosure to third parties including government authorities; (iii) the documents did not become privileged merely because they were subsequently shared with his lawyer. The court further held that the work product doctrine did not apply as the documents were not prepared by or at the direction of counsel and were prepared on Heppner’s own volition.
15. By contrast, in Warner v. Gilbarco (2:24-cv-12333), the United States District Court for the Eastern District of Michigan upheld work-product protection over a pro se litigant’s AI-assisted litigation materials. The plaintiff had used generative AI to analyse evidence and prepare litigation-related materials after the close of discovery. The Court held that those materials reflected the plaintiff’s own mental impressions and litigation strategy and were therefore protected under the work-product doctrine, notwithstanding that they had been prepared with the assistance of AI. Rejecting the argument that disclosure to a public AI platform necessarily waived such protection, the Court concluded that waiver occurs only where the disclosure substantially increases the likelihood that the protected material will come into the hands of an adversary. As the use of the AI platform did not materially increase that risk on the facts of the case, the work-product protection was held not to have been waived. The Court also observed that generative AI programs are merely tools, rather than persons capable of independently waiving privilege or work-product protection.  
16. The position was reaffirmed by the Supreme Court of New York in Assini v. Hayward, 2026 NY Slip Op 26086. The dispute arose from a subpoena issued to OpenAI seeking production of a litigant’s ChatGPT prompts, uploaded documents, outputs, legal research, and litigation-preparation materials. While the plaintiffs relied upon United States v. Heppner to contend that AI interactions were neither privileged nor protected, the Court expressly declined to follow Heppner and instead adopted the reasoning in Warner. The Court held that materials prepared through AI for the purposes of litigation continued to enjoy work-product protection and that the mere fact that public AI platform collect and retain user data did not, by itself, extinguish all reasonable expectations of confidentiality or amount to a waiver of protection. Accordingly, the subpoena directed to OpenAI was quashed, although the Court simultaneously cautioned that the use of AI remains subject to applicable court rules governing responsible deployment of artificial intelligence in litigation.
17. These authorities demonstrate that the treatment of AI-assisted legal work remains unsettled, with courts differing principally on whether disclosure to an AI provider undermines the confidentiality necessary to sustain privilege or work-product protection.
18. A note of caution is warranted before these authorities are imported. Warner and Assini turn substantially on the work-product doctrine, and Indian law has no free-standing equivalent; the nearest analogue is litigation privilege protecting materials prepared in contemplation of litigation.
19. In UK v Secretary of State for the Home Department[2026] UKUT 81, the English Upper Tribunal Immigration and Asylum Chamber considered the case of TMF Immigration Lawyers, whose solicitor had uploaded client correspondence and Home Office decision letters to a generative AI tool for the purpose of making summaries. Although the proceedings principally concerned AI hallucinations and lawyers’ professional duties, the Tribunal observed (obiter) that uploading confidential client documents to an open-source AI platform would place them in the public domain (at Paragraph 21), thereby breaching confidentiality and waiving legal privilege. The Tribunal distinguished closed-source AI platform operating under appropriate safeguards.

E. India’s Next Steps Forward

20. The interplay of artificial intelligence and legal privilege presents challenges that Indian law must address proactively. The precedents set by foreign jurisdictions, particularly the cautionary lessons of Heppner, Warner, and Assini, offer a roadmap. Regulators and judges in those jurisdictions have been clear: the responsibility does not shift with the technology and stays with the advocate. The profession’s fundamental duties of confidentiality and privilege must remain inviolate, regardless of the sophistication of the tools employed in their service.
21. Indian courts have consistently treated confidentiality as the foundation of privilege. Consequently, if information is voluntarily disclosed to an AI provider whose terms permit retention, review, or onward disclosure, a court may conclude that the confidentiality underpinning privilege has been compromised. Whether such disclosure constitutes waiver is likely to depend upon the nature of the AI platform, its contractual safeguards, and the degree of control retained by the user.
22. Indian law presently contains no specific legislation, judicial guidance, or regulatory pronouncement addressing AI and privilege. The Bar Council of India has not issued any guidance on the use of AI by advocates or how it will govern the rules of privilege. Accordingly, several reforms and clarifications are warranted. For instance, the BSA’s use of the phrase “communication made to him” under Section 132 presupposes a human recipient. A plausible argument may therefore be made that privilege should extend to AI-mediated communications, provided the AI platform merely functions as a technological intermediary and maintains contractual confidentiality. The stronger textual answer, however, lies within Section 132 itself: the extension to interpreters, clerks and servants shows that the section’s logic does not turn on the humanity of the instrument but on its subordination to the retainer. The clerk is within the privileged circle not because he is human but because he serves the advocate’s function under duties of confidence; an AI platform processing material under binding confidentiality obligations, for the client’s purpose alone, stands on the same footing.
23. A similar legislative policy is reflected in Section 227 of the Companies Act, 2013, which provides that nothing shall require the disclosure of information in respect of which a legal adviser would, in the ordinary course, be entitled to claim legal professional privilege. Although the provision operates within the specific context of investigations under the Companies Act, it evidences Parliament’s recognition that legal privilege survives even against otherwise extensive statutory powers of inquiry. This legislative affirmation strengthens the argument that, while privilege remains a substantive protection under Indian law, the use of artificial intelligence must be structured in a manner that preserves the confidentiality upon which that protection depends.
24. Further, firms must implement accountability frameworks, data governance protocols, and structured training to ensure that AI use remains necessary, proportionate, and consistent with the duty to preserve privilege and confidentiality.
25. Concretely, four moves are available, in ascending order of formality. First, the courts can act interpretively: reading the clerk-and-interpreter extension purposively, and adopting in substance the Warner test for waiver; the constitutional register in which In Re: Summoning Advocates places the privilege reinforces the purposive reading, for a protection anchored in Articles 19(1)(g), 21 and 22(1) should not be construed so as to evaporate upon the profession’s adoption of the ordinary tools of its time. Secondly, Parliament can place the matter beyond argument by an Explanation to Section 132, to the effect that processing of a communication by an electronic system under binding obligations of confidentiality, without human review by the provider and without use of the communication for training or any other purpose of the provider, does not constitute disclosure. Thirdly, the Bar Council of India can state minimum standards: no client material on tools whose terms permit training use or unrestricted retention; contractual confidentiality, retention limits and breach notification as conditions of any tool approved for client work; and informed client consent where matter sensitivity warrants it. Fourthly, firms need not await the regulator: matters should be classified by sensitivity, approved tools mapped to that classification, consumer-tier tools prohibited for client-identifiable material, and training delivered accordingly, with the transmission of client material to an AI platform treated with the same discipline as its transmission to any other outside vendor.

Footnotes

^{1 John Henry Wigmore, Treatise on Evidence in Trials at Common Law, 3rd Edn, Little, Brown & Co, Boston, 1940, p. 2292.}

^{2 Section 132, Bharatiya Sakshya Adhiniyam, 2023. This section consolidates the earlier Sections 126 and 127 of the Indian Evidence Act, 1872.}

^{3 Goraya, Makhani and Nigam, “Digital Strangers in Litigation: Does Sharing with AI Breach Privilege?” International Bar Association, Litigation Committee (29 October 2025).}

^{4 In Assini v. Hayward, 2026 NY Slip Op 26086, a preservation order required OpenAI to retain even conversations that users had deleted.}

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.