AI Is Disrupting Every Category of IP Law
Artificial intelligence creates genuine legal uncertainty across every IP category. Who owns content AI generates? Can a company patent an AI's invention? Does scraping data for training violate copyright? These questions do not yet have settled answers in India — or most global jurisdictions. For founders building AI-driven startups, understanding the current legal landscape and building IP strategies robust to uncertainty is essential.
AI-Generated Content — The Copyright Question
Generative AI tools produce text, images, music, code, and video at scale. When AI generates content, who owns it — the AI developer, the user who provided the prompt, or no one?
Under the Copyright Act 1957, copyright vests in the creator — defined as the person who uses skill, judgment, and creative effort to produce something original. Where a human provides substantial creative direction to an AI — detailed prompts specifying style, content, structure, and creative choices — the human's contribution may support a copyright claim. Where the human input is minimal (a simple prompt like "write a product description"), the degree of human creative authorship is questionable.
The US Copyright Office's evolving guidance — not binding in India but highly persuasive — takes the position that purely AI-generated content without human authorship is not copyrightable, while AI-assisted content with substantial human creative input is protectable. Indian startups building AI content products should: document human creative direction behind significant AI-generated outputs; register copyright at copyright.gov.in with accurate authorship disclosure; and monitor evolving guidance from India's Copyright Office as this area develops rapidly.
AI Inventions and Patent Law
AI systems increasingly generate technical innovations — new drug molecules, novel engineering designs, improved manufacturing processes. Can these be patented, and who is the inventor?
The consistent global position is that AI cannot be named as a patent inventor — inventors must be human beings. In the DABUS cases decided across the US, UK, EU, and Australia, courts and patent offices uniformly rejected applications naming AI as the inventor. The human who directed the AI, set the objective, and is responsible for the innovation is the appropriate named inventor, regardless of how substantially AI contributed to the discovery process.
For patentability, AI-generated innovations must meet the same criteria as any other invention — novelty, inventive step, and industrial applicability. The method of arriving at the invention (using AI) does not affect this analysis. However, if the invention is purely the result of an AI's mathematical computation, the Section 3(k) bar on mathematical methods may apply.
Training Data Rights — A Critical Risk Area
Machine learning models learn from data — and the rights associated with that training data are a significant and growing IP risk for AI startups. Most training data was created by humans who hold copyright in it. Using that data without appropriate rights may constitute copyright infringement.
Managing training data rights: use publicly licensed datasets obtained with rights permitting AI training; obtain explicit data licences from providers that expressly cover training use; use synthetic data where model performance allows; for user-generated content on your own platform, ensure Terms of Service include a training licence clause (with appropriate consumer protection compliance); and for scraped web data, assess whether source website terms of service prohibit automated collection and AI training use.
Beyond copyright, AI startups training on user data must comply with the Digital Personal Data Protection Act 2023 — personal data in training sets requires appropriate consent for training use.
Deepfakes and Personality Rights
AI tools generating realistic video, audio, and images of real individuals intersect directly with India's developing personality rights doctrine. The Delhi High Court's April 2026 ruling in the Allu Arjun case, treating an actor's distinctive likeness as a form of copyright, has significant implications for generative media startups.
Creating or distributing deepfake content using a recognisable individual's likeness without consent potentially exposes the startup to multiple claims: personality rights infringement, copyright infringement under the emerging AI copyright doctrine, defamation, and IT Act 2000 offences. AI startups building generative media products must implement robust consent verification systems, content moderation against non-consensual deepfake creation, and clear terms prohibiting generation of realistic media of identifiable individuals without their consent.
Algorithm Protection Strategy
Proprietary algorithms are among the most commercially valuable and most IP-vulnerable assets an AI startup possesses. Multilayered protection is essential because no single IP right covers all aspects.
Keep model architecture, training hyperparameters, and proprietary training methodologies as trade secrets — restrict access through NDAs, role-based controls, and secure API delivery so the model is never directly exposed. Register copyright in the source code implementing the algorithm. File patents for specific novel technical methods if they meet patentability criteria and produce technical effects beyond pure mathematical computation. Brand the AI product with registered trademarks for the commercial identity. Most AI startups use trade secrets as the primary protection for model internals, supplemented by patents for specific technical methods where applicable.
AI Licensing Models — Key Contractual Issues
AI startups monetising through licensing face specific structuring challenges. Model-as-a-service (API access) licences must address: permitted use scope; data handling — what happens to inference data and whether the startup can use it for further model training; output ownership — who owns content generated by the model in response to customer queries; and liability limitations for model errors, hallucinations, and bias outputs.
Output ownership is particularly commercially important. Many enterprise customers expect to own the outputs their queries generate, while AI startups may want to retain rights to use anonymised outputs for model improvement. Negotiate this clause specifically for each significant enterprise customer rather than using a standard one-size-fits-all approach.
For the broader emerging technology IP landscape, explore the related guides at the Startup IP Hub.