[By Shivam Tiwari]
The author is a student of Gujarat National Law University, Gandhinagar.
Introduction
Once a patentable drug compound is produced through an artificial intelligence system on its own, the conventional concepts of ownership and value creation begin to collapse. An example may be a cross-border structure in which the Indian research and development subsidiary offers data curation services to its Luxembourg parent on a cost-plus basis, such that the subsidiary is paid its costs plus a predetermined markup. The parent company’s AI system, which has access to both publicly available molecular databases and the proprietary patient data of the subsidiary, creates a commercially viable compound without the involvement of its human counterparts.
Under the OECD functional analysis, residual profits are generally allocated to the entity that performs the economically significant functions, assumes the relevant risks, and makes the principal contributions to value creation. AI-generated intangibles, however, complicate this analysis. Where an AI system autonomously develops a commercially viable drug compound using both publicly available molecular databases and proprietary datasets supplied by an Indian research and development subsidiary, it becomes difficult to determine whether the subsidiary merely rendered routine data-curation services or made a substantive contribution to the creation of the intangible. This uncertainty also raises a broader question: which entity should be recognised as the developer of the AI-generated intangible for transfer pricing purposes? These questions are no longer merely hypothetical.
Intangible assets now account for 82% of the total value of S&P 500 companies, according to Brand Finance’s Global Intangible Finance Tracker (GIFT) 2025. As Jonathan Haskel and Stian Westlake observe in Capitalism Without Capital, economies that are becoming more reliant on intangible assets pose special measurement and allocation problems. When non-human systems create or refine those resources, the problem becomes structural. Thus, the question is not whether AI can fit perfectly into the existing doctrine but whether it remains conceptually clear.
The OECD Framework and India’s Transfer Pricing Regime
The OECD Transfer Pricing Guidelines (Guidelines) are based on the arm’s length principle, which stipulates that related business enterprises should price transactions in a manner that is similar to how unrelated businesses would do so in the same market conditions. Chapter VI applies the DEMPE model of Development, Enhancement, Maintenance, Protection, and Exploitation to intangibles. This model identifies the party entitled to residual profits because it performs economically significant functions, assumes economically significant risks, and contributes to value creation. Legal ownership alone is insufficient.
The Guidelines also deal with hard-to-value intangibles, or assets whose capabilities to generate future income cannot be properly estimated at the moment of transfer. In such cases, tax authorities can consider the ex post performance to determine whether arm’s length conditions were incorporated in the original pricing. Recognised methods of valuation are functional analysis, comparability studies, discounted cash flow techniques and contingent arrangements such as royalties or milestone payments.
To a great extent, this structure is reflected in India’s transfer pricing regime under Sections 161 to 174 of the Income Tax Act, 2025 (IT Act), which govern the computation of income arising from international transactions and specified domestic transactions between associated enterprises in accordance with the arm’s length principle. Section 165 recognises six methods for determining the arm’s length price, while Rule 79 of the Income-tax Rules, 2025 (IT Rules) sets out the manner in which those methods are to be applied. The Cost-Plus Method compensates a service provider by marking up the expenses. The Profit Split Method is used to allocate total profits to related businesses based on their related contributions. The Transactional Net Margin Method helps to compare net margins of similar independent businesses.
In 2012, India introduced Advance Pricing Agreements (APA) in order to enhance certainty. An Advance Pricing Agreement is a pre-transaction agreement between a taxpayer and tax authorities stating in advance the way transfer pricing rules will apply to specific transactions (typically over a number of years). Of the over 500 unilateral and bilateral APAs that the Central Board of Direct Taxes (CBDT) had finalised by March 2023, approximately 18 percent were of intangibles. To resolve cases of double taxation, tax authorities in different jurisdictions invoke the Mutual Agreement Procedure (MAP) as a post-transaction dispute resolution mechanism. These mechanisms are effective when there is an identifiable human activity that can be associated with the value creation. Once that assumption fails, the real challenge emerges.
The Attribution Paradox: Algorithms as a Value Creation Process
The DEMPE model presupposes the involvement of identifiable legal persons informing the human decision-making process and executing economically significant functions. Current AI systems complicate this concept. Large language models, among other high-end predictive models are trained using large datasets to identify patterns and deliver results with minimal human supervision. The data used to facilitate this learning process is referred to as training data. In a conventional analysis, the role of the latter may be identified as routine in cross-border AI arrangements in which the former party develops and trains the model and the latter provides proprietary datasets. The distinction between regular input and value creation becomes difficult when the quality of datasets directly affects commercial performance. The allocation of residual profits therefore relies more on the technological interdependence than on legal ownership.
Domestic law displays this uncertainty. The Copyright Act, 1957 (Copyright Act) does not attract any presumption of non-human authorship. Although 2024 governmental clarifications suggested that already developed rules regarding copyright could help regulate the AI-generated works, they did not provide the answer to the more significant question regarding the attribution of ownership. The case of Asian News International v. Open AI (2024) (Open AI) highlights this ambiguity in the impending ruling in the case. The case concerns an admittedly illegal use of copyrighted content to provide training to large language models, but it raises broader concerns regarding control and ownership in AI-driven models. The Digital Personal Data Protection Act, 2023 (DPDP Act) remains silent on the issue of the proprietary rights of AI-generated outputs but only focuses on personal data processing. This has a direct effect on the legal certainty of the analysis of the transfer pricing in the context of the consistent allocation of profits as the primary right to value is not legally clear.
The Data Valuation Crisis
The valuation problem is directly influenced by this attribution dilemma because both the international framework and the Indian counterpart do not specifically acknowledge training data as an independent intangible that has to be valued separately. AI relies on training data. It dictates the accuracy of the system, its intelligence, and its profitability, which can be well beyond the value of the cost of gathering it. Such stringent regulations as the General Data Protection Regulation, 2018 (GDPR) in Europe and the Digital Personal Data Protection Act in India severely limit transnational data transfer, making high-quality datasets even more commercially significant. Where a company’s personal data constitutes a key driver of profitability, its economic value is evident, yet tax and transfer pricing regulations do little to recognise it.
The standard practices are not effective. The Comparable Uncontrolled Price method requires analogous transactions by unrelated parties, but such transactions practically never exist for specialised AI datasets. The Cost-Plus Method merely examines the costs incurred in providing the service, rather than the substantial economic value ultimately created by the resulting intangible. The Profit Split Method provides no objective means of allocating profits when AI models and training data are closely integrated.
As long as we do not formally consider training data as a distinct intangible asset, valuation debates will be inevitable.
The Comparability Collapse
Comparability presents a significant challenge in applying the arm’s length principle to AI-generated intangibles. Even if questions relating to attribution and valuation are resolved, the arm’s length principle ultimately depends on the availability of reliable comparable uncontrolled transactions. AI-generated intellectual property complicates this analysis because AI models differ in their architecture, the composition of their training datasets, and the manner in which they continue to evolve.
New-generation AI models are radically different in structure, the precise combination of training data, and the way their features continue to advance. They can be reproduced at minimal marginal cost, and hence the traditional relationship between cost and value is destroyed. True third-party licences to developed AI are very uncommon and highly customised. The old-fashioned R&D outsourcing transactions would be an inappropriate match with the high-speed, low-cost development cycles that we experience today. The majority of software licences are structured around usage and not the actual monetary value created.
In the absence of consistent, verifiable market standards, the comparability analysis that is being constructed around transfer pricing becomes an educated conjecture.
Reform Imperative: Redefining Transfer Pricing to AI
It is time to reconsider transfer pricing in the AI era. Taken together, ambiguous attribution, the failure to recognise training data as a distinct asset, pooling, and weak ex post mechanisms place considerable strain on the existing framework.
OECD Guidelines continue to function effectively for traditional, human-created intangibles. However, they are premised on identifiable human inputs, risks that can be effectively assumed and controlled, and reasonably foreseeable development trajectories. These assumptions are difficult to sustain where AI systems autonomously generate or substantially enhance commercially valuable intangibles.
The same fundamental strategy is applied in India and there is no AI-specific guidance. The EU AI Act, 2024 is concerned with transparency and safety but says nothing about the distribution of profits across borders. India may take the lead by providing specific AI-related guidance within the Advance Pricing Agreement programme, requiring companies to disclose model architecture, training data sources, and relevant functional parameters and amending Section 165 of the IT Act to recognise AI-generated IP and training datasets as distinct asset classes with dedicated valuation methodologies.
At the global level, the OECD will have to reconsider its initial assumptions regarding human agency and control of the economy when the creator is an algorithm. Economies of intangibles, as Haskel and Westlake indicate, are difficult to regulate in the first place. The challenge becomes even greater when intangibles begin to develop and enhance themselves through AI. Traditional transfer pricing rules were developed with human decision-makers in mind. The task now is to adapt them to a framework in which value creation is increasingly autonomous.