1. What it means

Data Sovereignty = principle that data is subject to the laws and governance of the nation where it was collected or where it physically resides.

For AI training, this creates 3 obligations:

a) Residency: Data must stay inside the border for processing. India, China, Russia mandate this for certain categories.
b) Jurisdictional control: Who can legally compel access to it. The US CLOUD Act says a US-domiciled vendor must hand over data even if server is in Frankfurt. The EU says that is incompatible with GDPR.
c) Provenance & Purpose: You must prove lawful basis, consent, and purpose limitation for training. GDPR, DPDP Act, and upcoming EU AI Act all require it.

In AI, sovereignty is enforced at the context/retrieval layer, not just model layer. When your RAG agent pulls data via MCP, that retrieval may be subject to GDPR, Schrems II or local residency laws, regardless of where the model runs.

2. Why AI Training is a Sovereignty Flashpoint

1. Massive cross-border scraping: Models train on billions of web images, text from EU, India, US without explicit consent.
2. Biometric and sensitive data: Facial recognition models create biometric templates - special category under GDPR Article 9.
3. Transfer for compute: Training happens in US clouds but data subjects are in EU/India.
4. Right to Know / Explainability: New laws require disclosure of training sources. California and EU AI Act Article 50 transparency obligations live from Aug 2026.

3. Global Legal Frameworks

EU: GDPR + AI Act + Data Act + Schrems II
GDPR applies extraterritorially. Any AI system that operates through large-scale indiscriminate collection may be subject to GDPR even when provider does not intend to target EU. Personal data processed by AI must remain under EU jurisdiction, contractual assurances with US-headquartered providers are insufficient.

India: DPDP Act, 2023 + DPDP Rules, 2025

• Section 3(c)(ii): Act does NOT apply to personal data made or caused to be made publicly available by the data principal. This was initially read to allow scraping for AI.
• Rules 2025 tighten it: AI companies must obtain free, specific, informed consent for each specified purpose, with granular notices. Consent must be separate for training vs other uses.
• Section 17(5) exemption debate: IAMAI asked Centre to exempt processing solely for training AI models on publicly available data. 75% of Indian firms remain dependent on public data for training. Also no text-and-data-mining exception in Copyright Act, 1957 - mass training without licence could violate Section 14 reproduction right.

US: Fragmented
CCPA/CPRA + state laws, FTC oversight on algorithmic discrimination, CLOUD Act vs EU. No federal sovereignty law, but FTC treats biased AI targeting as civil rights violation.

China: PIPL + DSL + CSL
Strict data localisation for important data, government-approved transfer mechanisms for AI training abroad.

4. Landmark Case Laws

A. The Clearview AI Saga - The Defining Sovereignty Case

Facts: US company scraped 30+ billion facial images from Facebook, LinkedIn, YouTube without consent to train facial recognition AI for law enforcement.

Holdings across EU:

• Netherlands AP, 2024: Fined €30.5M. Dutch DPA ruled legitimate interest cannot be legal basis for biometric processing of this magnitude. Active authorization required, opt-out models closed.
• France CNIL, Italy Garante, Greece, Austria 2022-23: €20M fines each, ordered deletion. Found Clearview is controller under Article 4(7) GDPR because it determines purposes and means for setting up database and training algorithm.
• UK ICO: Initially £7.5M fine, later won on appeal on narrow foreign law enforcement exemption. UK proceedings noted Clearview collected photos for training.

Principle: Publicly available on internet ≠ lawfully available for AI training. Scraping = processing under GDPR.

B. Schrems I & II (CJEU C-362/14 and C-311/18)

Schrems I (2015): Invalidated Safe Harbor EU-US.
Schrems II (2020): Invalidated Privacy Shield. Held US surveillance laws not adequate. Created Transfer Impact Assessments (TIA) + Standard Contractual Clauses (SCCs) requirement.
Impact on AI: Any AI training that moves EU personal data to US cloud for compute is now a restricted transfer. Many orgs overcorrect by refusing all transfers when compliant architectures exist. Led to DPC v Meta 2023 - €1.2 billion fine for EU-US data transfers used for ad targeting and AI model training.

C. ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, CS(COMM) 1028/2024, Delhi High Court [24 July 2026]

First Indian AI training case.
ANI alleged OpenAI's scraping, storage and reproduction of its literary works amounted to copyright infringement - training claim and output claim (verbatim copies).

Interim order by Justice Amit Bansal: Refused injunction. Held storage of ANI's work for training purposes is prima facie protected by Section 52(1)(a) of Copyright Act - fair dealing for private or personal use including research. Court candid that when Section 52 last amended in 2012, legislature could not have imagined AI.

Status: Appeal notice issued 15 September 2026, division bench refused ex-parte stay. Fact-specific, cannot be general assurance that AI training is lawful.

D. hiQ Labs v. LinkedIn, 9th Cir. (2019, 2022)

US court held scraping publicly available LinkedIn data likely does NOT violate CFAA. Crucial for AI startups arguing public data is fair game. Distinction: US focuses on access, EU/India focus on privacy/copyright purpose.

E. The New York Times v. OpenAI/ Microsoft (S.D.N.Y. Dec 2023)

NYT alleges millions of articles used for training without licence, outputs regurgitate verbatim. Will test US fair use four factors vs data sovereignty claims of publishers. Parallel to ANI case but under US law.

5. How to Build Compliant Sovereign AI Training

1. Data Mapping: Document source country, legal basis, consent artefact for every dataset shard.
2. Lawful Basis by Region: EU - consent or legitimate interest with DPIA, cannot rely on public availability alone. India - granular purpose-specific consent under DPDP Rules + anonymisation if using personal data.
3. Architectural Sovereignty: Keep data plane swappable - train in-region, use sovereign cloud, Zero Data Retention contracts.
4. Technical Controls: Encryption at rest/in-transit, federated learning, differential privacy, synthetic data substitution.
5. Contractual Sovereignty: MSAs must include AI IP indemnification, Zero Data Retention, and explicit prohibition on using customer data for model training.

Bottom line for India: You are in a grey zone. The ANI judgment gives temporary breathing room under copyright fair dealing, but DPDP Act's consent-centric framework and lack of TDM exception means using personal data of Indians - even public posts - for AI training without verifiable consent is high risk until Section 17(5) exemption is notified.