Insolvency of AI Startups: Who Owns the Data, Algorithms and Trained Models?

Author~ Riya Gugliya

Abstract

India is now the third largest startup ecosystem in the world, trailing only the United States and China, with over 3,200 active AI startups as of 2026. Yet this growth sits on a legal fault line that few have noticed. When an AI company makes money, it is celebrated. When it fails, an uncomfortable question emerges one that neither its investors nor its lawyers are well-equipped to answer: who owns the data, algorithms, and trained models left behind?

These are not ordinary assets. A dataset is not a machine you can lock and auction. A trained AI model is not a building you can value from a balance sheet. They are intangible, deeply entangled with the privacy of millions of users, and governed by a patchwork of laws that were never designed to work together in an insolvency scenario. The Insolvency and Bankruptcy Code, 2016 (IBC) India's primary framework for handling corporate failure was built for factories and balance sheets. It has no answer for assets that learn, evolve, and contain the personal information of people who never signed up to become part of a creditor recovery process. This blog examines that gap, tests the law against the reality of AI businesses, and makes the case for reform before the courts are forced to invent answers on the fly.

Keywords: Artificial Intelligence, Insolvency, IBC 2016, Datasets, Algorithms, Trained Models, DPDP Act 2023, Intellectual Property, Data Privacy, Resolution Professional

Introduction - The Problem Nobody Is Talking About

There is a version of the AI startup story that everyone tells the funding round, the breakthrough product, the unicorn valuation. There is another version, quieter and less glamorous, that is increasingly common: the failed pivot, the dried-up runway, the winding-up petition. As India's AI sector matures, startup failures are not a hypothetical. They are a statistical inevitability.What makes AI insolvency genuinely different is the nature of what gets left behind. When a traditional company enters the Corporate Insolvency Resolution Process (CIRP) under the IBC, the Resolution Professional takes custody of identifiable assets and runs a structured process. When an AI startup enters CIRP, the Resolution Professional is likely holding assets they cannot define, cannot value accurately, and cannot transfer without potentially violating the privacy rights of people who had nothing to do with the company's financial failure. This is the problem. And Indian law has not yet caught up with it.

2. What Are AI Assets, really?

Datasets are the raw material user-generated data collected by AI companies for training their systems. This data is not owned outright by the company the way it owns its office furniture. Much of it belongs, in a meaningful sense, to the individuals who generated it. A fitness app does not own your step count; it holds it conditionally, subject to the purpose for which you gave consent.

Algorithms are the proprietary coded logic more clearly owned by the company, often protected as software under the Copyright Act, 1957, and sometimes as trade secrets. But their value deteriorates rapidly when the technical team that built them leaves.Trained AI models are the most legally complex of all. A model is the end product of running algorithms over datasets it has learned from data, and in doing so, it may have absorbed patterns derived from personal information in ways that cannot be easily reversed. Selling a trained model is not like selling a machine. It may be more like selling a record of everything your users ever did, distilled into mathematical weights.

What the IBC Says and What It Misses

The IBC is not entirely silent on intangibles. Section 18(f)(iv) specifically lists "intangible assets including intellectual property" among what the Interim Resolution Professional takes custody of. The Supreme Court, in Victory Iron Works Ltd. v. Jitendra Lohia (2023), confirmed that the term "asset" includes "property of any kind." Datasets and algorithms can, in principle, fall within this.

The deeper problem is the complete absence of any AI-specific framework. There are no guidelines for what happens when an "asset" is a database of personal data subject to the DPDP Act. There are no valuation standards for trained models. There is no requirement that a Resolution Professional have any technical competence in what they are handling. A generalist insolvency professional confronted with a fine-tuned language model and three petabytes of health data has no statutory guidance whatsoever. The IBC defaults to treating everything as transferable which may satisfy creditors while violating the rights of every user whose data made those assets valuable.

The Dataset Problem: Whose Data Is It Anyway?

The most commercially attractive AI asset in an insolvency estate is almost always the dataset. Under Section 6 of the DPDP Act, a data fiduciary may only process personal data for the specific purpose for which the data principal gave consent consent that must be "free, specific, informed, unconditional and unambiguous." A user who gave a health startup permission to use their data for diagnostics did not consent to having that data auctioned to a competitor in a court-supervised bankruptcy process.This conflict has real precedent abroad. The U.S. Federal Trade Commission intervened in the Toysmart bankruptcy (2000) and the RadioShack bankruptcy (2015) to prevent customer data from being sold in violation of original privacy commitments. In the RadioShack case, a coalition of 36 state attorneys general negotiated conditions that led to the destruction of most customer data rather than its unrestricted sale. India has no equivalent institutional practice the Data Protection Board of India is not a party to CIRP proceedings, and the IBC requires no privacy compliance review before insolvency sales of personal data.

Algorithms, Trained Models, and the Valuation Gap

Algorithms and source code sit on firmer legal ground protected as literary works under the Copyright Act, 1957, and potentially as trade secrets under common law. These can be sold in insolvency without the consent complications that apply to personal data.Trained models raise harder questions. Research has shown that some models can be queried to extract information about their training data, making a model sale potentially equivalent to a data transfer in disguise. When a model trained on personal health or financial data is sold to a new entity, does the original consent follow it? Can users exercise their right of erasure under Section 13 of the DPDP Act against the acquiring company? The law currently provides no answers.Compounding this is a valuation gap. The IBBI (Registered Valuers and Valuation) Rules, 2017 recognise three asset classes for registered valuers: securities and financial assets, plant and machinery, and land and buildings. AI-specific intangibles fit into none of them. The result is that AI assets get undervalued, creditors recover less than they should, and acquirers often competitors gain enormous advantage at a price the law never designed to set.

The Reform Agenda

The gaps are structural and will produce bad outcomes if left unaddressed. Three reforms are immediately necessary.

First, the IBBI should issue guidelines recognising datasets, algorithms, and trained models as distinct asset categories with their own protocols for preservation, valuation, and transfer. Resolution Professionals handling AI insolvencies should be required to engage technical experts. The registered valuer framework must be updated to include a category for intangible AI assets.

Second, the DPDP Act and IBC must be made to interact. Any CIRP involving personal data should trigger mandatory notification to the Data Protection Board of India. Transfers of personal data including through model sales should require privacy compliance review before NCLT approval.

Third, India should look to international frameworks. UNCITRAL's Model Law on Cross-Border Insolvency is already part of Indian law through the Insolvency and Bankruptcy (Amendment) Act, 2021. As AI startups increasingly hold assets and users across multiple jurisdictions, the cross-border dimensions will require their own attention.

 Conclusion

The failure of an AI startup is not just a business event. It is a privacy event, an intellectual property event, and a governance event all at once. The IBC can take custody of a dataset, but it cannot determine whether selling that dataset violates the rights of the people inside it. It can transfer a trained model, but it cannot assess whether doing so amounts to a second processing of user data without consent.India is building one of the world's most ambitious AI ecosystems. It deserves an insolvency framework that matches that ambition one that protects creditors, respects users, and ensures that when AI companies fail, they do so in a way that the law has actually thought through.

As artificial intelligence becomes a cornerstone of the digital economy, the question is no longer whether AI startups will enter insolvency. It is whether insolvency law is prepared to deal with assets that think, learn, and evolve.

Note - The information contained in this blog is for general informational purposes only. We endeavour to keep all content accurate, updated, and free from any form of misinformation or objectionable material. However, we shall not be responsible for any claims arising out of copyright infringement, plagiarism, or related issues; such responsibility lies solely with the respective authors. If you find any misinformation or objectionable content on this website, please report it to us at: editors.ilw@gmail.com