Beyond the black box: Rethinking copyright in the age of AI training

The widespread adoption of artificial intelligence (AI) questions the foundational contours. A jurisprudence designed for human creativity has now to grapple with the extractive, opaque mechanics of machine learning. The debate is no longer about the originality of AI-generated output, but the legality of AI learning. At its core is whether the wholesale scraping of copyright works during AI training is itself copyright infringement.

Indian copyright law is based on the division of ideas and their expression. It protects the form of expression while allowing the use of the underlying idea. AI training has disrupted this simple model. These systems copy entire works, not to replicate them but to extract unprotectable elements, facts, syntactic structures and relational patterns. This changes the doctrinal foundation of copyright scrutiny from the output to the input.

This reconstruction is profound. Courts may have to assess the legality of data acquisition, rather than the AI-generated output. The fair-dealing exception under section 52 of the Copyright Act, 1957 (act), although generally expansive, remains human-centric and is ill-suited to defend broad-spectrum, commercial AI training. This is particularly so when substitutional harm to the creator is likely and foreseeable. The Supreme Court in Engineering Analysis Centre of Excellence Private Limited v Commissioner of Income Tax and Anor, held that using software is distinct from copying it. However, the use of AI involves a form of reproduction, blurring this distinction.

The more important clashes may not involve section 52, but rather sections 65A and 65B. They prevent the circumvention of technological protection measures (TPM) and tampering with rights management information (RMI). Meant to curb digital piracy, these provisions may stop unauthorised AI-driven data scraping.

Because technology evolves, these sections may become independent of section 52. Even if the AI output is transformative and non-substitutive, bypassing TPMs or manipulating metadata may still attract liability. Delhi High Court’s recent decision in Dassault Systemes and Ors v Advanced Engineering Solutions and Anor implicitly recognised server-based tracking as a valid TPM. Although the ruling was not directed at AI, it shows judicial acceptance of broader interpretations of protective digital architecture.

Despite increasing legal protection, uncertainty persists. For example, a standard protocol designed to communicate web crawling preferences, robots.txt, is unlikely to be a TPM under section 65A because it is a behavioural directive and not a technical access control. This lack of clarity leaves AI developers and rights holders in a legal grey area.

Global jurisprudence is also inconsistent. In eBay Inc v Bidder’s Edge Inc, the US court treated the disregard of robots.txt as a trespass to chattels, equating digital overreach with unlawful interference. Conversely, Field v Google held that failure to use robots.txt constituted an implied licence, weakening claims of unauthorised use. These cases show that while robots.txt may lack hard enforceability, ignoring it may attract ancillary liability or undercut fair use defences.

In India, the lack of statutory or judicial recognition of protocols such as robots.txt as TPMs adds to the ambiguity. Inferring a section 65A intention to infringe in training an AI model will be crucial in the future. Unless a clear technological barrier is breached or RMI is altered, its disregard may not attract liability. However, if metadata, such as author identifiers or usage rights, are stripped or altered during AI training, section 65B may still apply.

Regulators increasingly require clean hands conduct. When AI developers circumvent digital safeguards or alter metadata, fair dealing defences may be untenable. Following the US Copyright Office’s stance, unlawful data acquisition may disallow fair use claims. The EU’s Copyright in the Digital Single Market Directive allows commercial text and data mining if rights holders have not opted out, with the EU Intellectual Property Office extending this to natural language opt-outs. India lacks official direction for this subject, leaving it to the courts to decide whether individual training methods are infringing. Policy reform is essential to balance innovation with rights protection.

Pravin Anand is the managing partner, Siddhant Chamola is an associate partner and Alvin Antony is an associate at Anand and Anand. Ajai K Garg, director at Anand and Anand, also contributed to the column