Who Owns AI Outputs? Allocating IP, Liability and Indemnities in AI Contracts, India

Introduction, why India’s evolving IT rules change AI contracting

AI contracts India teams are now negotiating in a materially different regulatory environment. India’s information technology framework is evolving to address synthetic media and generative AI, and proposed and amended rules under the Information Technology Act, 2000 point toward operational obligations such as labelling of synthetic content, traceability, designated grievance/compliance contacts and cooperation with takedown notices. As these obligations harden, what used to be soft governance commitments increasingly become contractual requirements. For buyers and vendors, that means IP ownership, warranties, liability caps and indemnities can no longer be treated as boilerplate; they must be drafted to allocate concrete regulatory risk.

Search engines increasingly surface AI-generated summaries for queries like this one, but summaries cannot draft the clause you need to sign next week. This guide fills the gap with clause-level drafting guidance, negotiation levers and a compliance mapping tuned to Indian law.

Methodology note: portions of the sample clauses in this guide were prepared with drafting-assistance tools and then reviewed against the primary regulatory sources cited throughout. Because the regulatory position on synthetic media and AI in India is developing, every clause should be verified against the rules in force at the time of contracting. This article is for general information and does not constitute legal advice.

Ownership vs licence, who owns AI‑generated outputs in India?

The single most contested question in AI contracts India negotiations is ownership. When a user submits a prompt and a model returns text, code, images or audio, who holds the rights in that output, the buyer who prompted it, the vendor whose model produced it, or neither? Indian statutory law does not resolve this cleanly for machine-generated material, which is precisely why the contract, not the statute, will usually decide the outcome.

Statutory and judicial background

Indian copyright law, governed by the Copyright Act, 1957, is built around human authorship, and there is no bespoke statutory regime that vests ownership of purely machine-generated outputs. This creates uncertainty at the margins: outputs with substantial human direction may attract protection in the hands of the human contributor, while wholly autonomous outputs sit in a grey zone. Because the statute does not squarely address this point, parties should not assume any default allocation, they must contract expressly.

Two other bodies of law shape the picture. First, the Digital Personal Data Protection Act, 2023 governs how personal data used in prompts, fine-tuning or outputs may lawfully be processed, which constrains what a vendor may do with customer-derived material. (Note that the Act’s operative provisions come into force as notified by the Central Government, and the detailed rules under it should be checked for current status.) Second, the Supreme Court’s privacy jurisprudence, recognising informational privacy as a fundamental right in the Justice K.S. Puttaswamy v. Union of India line of cases, informs the standard of care around personal data embedded in training sets and outputs. Ownership drafting therefore cannot be divorced from data governance.

Ownership models

In practice, AI contract clauses allocate output rights through one of several models. The right choice depends on how the buyer intends to commercialise the outputs and how much residual value the vendor wants to retain.

• Assignment. The vendor assigns all right, title and interest in the outputs to the buyer. This gives the buyer maximum control but often attracts a price premium and vendor resistance.
• Exclusive licence. The buyer receives exclusive rights, usually bounded by territory, field of use or time, while the vendor retains underlying title. This is a common middle ground.
• Non‑exclusive licence. The buyer may use outputs but the vendor may licence similar or identical outputs to others. Cheapest for the buyer, riskiest for exclusivity-dependent use cases.
• Joint ownership. Both parties hold rights, typically with reciprocal accounting or reporting. Simple in principle but litigation-prone without a clear governance mechanism.
• Output-specific rights. Ownership varies by output category, for example, the buyer owns bespoke deliverables while the vendor retains rights in generic or template material.

Practical drafting options and negotiation checklist

When negotiating ownership of AI outputs India counsel should work through a consistent checklist rather than accepting the vendor’s first draft:

• Downstream rights. Confirm whether the buyer may modify, adapt, sublicense and commercialise outputs without further consent.
• Re-use by vendor. Restrict or permit the vendor’s re-use of outputs and of the prompts that generated them.
• Sublicensing. Specify whether sublicence rights are permitted, and on what terms sublicensees are bound to compliance obligations.
• Moral rights and attribution. Address any attribution or labelling requirements, which may interact directly with synthetic-media duties under India’s IT framework.
• Consideration levers. Where the vendor resists assignment, negotiate royalties, usage reporting, source-code or model escrow, or an indemnity uplift as a trade-off.

Data, training sets and provenance, upstream obligations and IP risk

Ownership of the output is only half the risk. The far larger exposure often sits upstream, in the data on which the model was trained and fine-tuned. If a model was trained on infringing content or on personal data processed without a lawful basis, the buyer can inherit downstream liability. Well-drafted AI contracts India should therefore extend beyond outputs to interrogate provenance.

Warranties about training data

Buyers should require the vendor to warrant that training data was lawfully acquired, that the vendor holds or has cleared the necessary third-party rights, and that any personal data used was processed consistently with the Digital Personal Data Protection Act, 2023. A robust warranty addresses three risk vectors: third-party contractual rights, copyright in scraped or licensed corpora, and personal data captured in training sets. Where the vendor cannot warrant provenance absolutely, negotiate a qualified warranty backed by an indemnity for third-party infringement claims.

Representations on model provenance

Provenance representations should cover the model itself: its origin, whether it is a foundation model licensed from a third party, how it was fine-tuned, and whether model weights incorporate any restricted or open-source components with reciprocal licensing obligations. This matters because a permissive-looking output can carry hidden licence conditions inherited from the base model. Vendors supplying fine-tuned or wrapped third-party models should disclose the provenance chain so the buyer can assess flow-through obligations.

Audit and certification rights

Because warranties are only as good as the ability to test them, buyers should secure audit and certification rights. These can range from a right to receive third-party audit reports and safety-testing summaries, to a contractual right to inspect data-governance documentation, to periodic vendor certifications of continued compliance. For higher-risk deployments, negotiate the right to commission an independent audit at the vendor’s cost where a material breach is reasonably suspected.

Liability allocation for AI outputs, practical approaches

Generative systems produce errors, hallucinations, defamatory statements, biased or infringing content, and liability for AI outputs must be allocated before, not after, an incident. The drafting challenge is to size caps and carve-outs so that risk falls on the party best able to control it, without making the deal commercially unviable for either side.

Risk matrix

Start by categorising the harm the deployment could cause. A structured risk matrix lets counsel map each harm type to a liability treatment:

• Reputational harm. Offensive, biased or defamatory outputs damaging brand or third parties.
• Financial harm. Erroneous outputs relied upon in decisions, causing direct loss.
• Regulatory harm. Non-compliance with applicable IT rules or intermediary obligations triggering enforcement action.
• Privacy harm. Unlawful processing or disclosure of personal data under the DPDP Act.
• IP infringement. Outputs reproducing third-party protected material.

Caps, exceptions and carve-outs

Liability caps are the central negotiation. Vendors typically seek a cap set at fees paid over a defined period; buyers seek higher caps for high-risk categories. The critical work is in the carve-outs, the categories that sit outside the cap or attract a super-cap. Common carve-outs in AI contracts India include gross negligence, wilful misconduct, breach of confidentiality, IP infringement indemnities and, increasingly, regulatory penalties arising from the vendor’s non-compliance with applicable IT rules. Buyers should resist a vendor attempt to exclude all consequential loss where regulatory penalties are foreseeable and directly attributable to the vendor’s default.

Incident response, recall and remediation

Liability drafting should be paired with operational obligations that reduce loss. Require the vendor to notify the buyer of material incidents within a defined window, to cooperate on containment, and, where feasible, to disable, roll back or recall problematic outputs or model versions. For synthetic-media incidents in particular, coordinate the contractual incident-response clause with any applicable takedown and grievance/point-of-contact duties under India’s IT framework so obligations do not conflict during a live event.

Warranties, indemnities and remediation clauses, buyer and vendor playbooks

This is where most AI contracts India negotiations are won or lost. Warranties allocate the promise, indemnities allocate the money, and remediation allocates the fix. Each should be drafted deliberately, with buyer-side and vendor-side positions understood in advance.

Core warranties

At minimum, a buyer should seek warranties that the service complies with applicable law including the Information Technology Act, 2000 and rules made under it and the Digital Personal Data Protection Act, 2023; that outputs will not knowingly infringe third-party IP; that training data was lawfully sourced; and that deliverables are free of malware or malicious code. Vendors will seek to qualify these with knowledge and materiality thresholds. A balanced outcome usually qualifies the IP warranty by knowledge while keeping the legal-compliance and malware warranties absolute.

Indemnity drafting

A generative AI indemnity clause turns on four variables: scope, trigger, defence control and limitation. Scope defines the losses covered, typically third-party claims for IP infringement, privacy breach and regulatory penalties. Trigger defines what activates the indemnity, a claim, a demand, or actual liability. Defence control determines who conducts the defence and who may settle. Limitation ties the indemnity to, or excludes it from, the general liability cap. The negotiation ranges from buyer-favour to vendor-favour:

• Buyer-favour. “The Vendor shall defend, indemnify and hold harmless the Buyer against all losses, damages, penalties and costs arising from any third-party claim that the outputs, model or training data infringe any intellectual property right, breach the Digital Personal Data Protection Act, 2023, or violate applicable rules under the Information Technology Act, 2000, without limitation by the liability cap. Draft, for negotiation; verify against law.“
• Vendor-balanced. “The Vendor shall indemnify the Buyer against third-party IP infringement and privacy claims arising from the Vendor’s model or training data, subject to the Buyer’s prompt notice, the Vendor’s sole conduct of the defence, and an aggregate cap of [●], provided the claim does not arise from the Buyer’s modification or unlawful use of the outputs. Draft, for negotiation; verify against law.“
• Vendor-favour. “The Vendor’s indemnity is limited to direct third-party IP claims proven to arise solely from unmodified outputs, capped at fees paid in the preceding twelve months, and excludes any claim relating to the Buyer’s prompts, data inputs or downstream use. Draft, for negotiation; verify against law.“

Negotiation note: the most contested item is whether regulatory penalties under India’s IT rules sit inside or outside the indemnity. Buyers should push for coverage where the penalty stems from the vendor’s failure to meet its own statutory duties, such as any applicable labelling or traceability obligations.

Remediation

Money is not always the buyer’s priority; continuity often is. A strong remediation regime should combine service credits for missed performance, rollback rights to a prior stable model version, an obligation to retrain or re-tune where systemic errors are identified, mitigation obligations, and coordinated public communications where an incident becomes public. For synthetic-media harms, remediation and any applicable takedown duties should be sequenced so the vendor acts within the required window.

Contractual compliance with India’s IT framework & DPDP, mapping duties to clauses

The distinguishing feature of AI contracts India in the current environment is that regulatory duties are increasingly directly contractible. Rather than a generic “comply with law” clause, counsel should translate each specific duty into an operative obligation that can be monitored and enforced.

Key IT-framework duties that are contractually enforceable

India’s evolving rules for intermediaries and synthetic/generative content point toward obligations that map naturally onto contract clauses. Because the precise requirements depend on the rules in force, verify the current text before relying on any specific duty:

• Labelling. Where synthetic or AI-generated content must be identifiable, build a warranty and an operational obligation requiring the vendor to apply and maintain labels.
• Traceability. Require logging and record-retention sufficient to trace generated content, and grant the buyer access to those logs.
• Grievance / point-of-contact. Contract for a designated contact and defined response times for takedown and notice handling, consistent with intermediary grievance-officer requirements.
• Content takedown and notice cooperation. Oblige the vendor to cooperate with lawful notices and effect removal or disabling within the applicable window.
• Record retention. Specify retention periods and formats consistent with applicable rules and with evidentiary needs.

Drafting checklist

Translate each duty into a clause and a control: a labelling clause tied to acceptance testing; a logging and audit clause with defined retention; a notice clause with response SLAs; a regulator-cooperation clause requiring the vendor to assist in responding to lawful requests; and a point-of-contact clause naming a role, not just an individual, so continuity survives staff changes. Each clause on synthetic-media compliance should reference the underlying obligation so the mapping is auditable.

Cross-reference to DPDP obligations

Where personal data is processed, the contract must satisfy the Digital Personal Data Protection Act, 2023 (once its provisions and rules are in force). Include data-processing clauses that specify lawful bases, purpose limitation, restrictions on using personal data for further model training without an independent lawful basis, and a mechanism for handling data-principal requests. Data protection AI contracts should also allocate responsibility for breach notification and cooperation on data-principal rights, ensuring the vendor’s model-training practices do not create a compliance gap for the buyer.

Procurement and vendor selection, operational & negotiation checklist

Good AI procurement India practice front-loads risk assessment into vendor selection, so the contract negotiation begins from an informed position rather than a marketing deck.

RFP red flags and must-haves

Treat as red flags any vendor unwilling to disclose model provenance, any refusal to warrant training-data lawfulness, and any blanket exclusion of liability for regulatory non-compliance. Must-haves for the RFP include a commitment to applicable labelling and traceability obligations, a documented incident-response capability, and willingness to accept audit rights. Building these into procurement criteria strengthens the buyer’s negotiating leverage before drafting begins.

Due diligence

Due diligence should request model provenance documentation, red-team and safety-testing results, and any third-party audit reports. For deployments touching personal data, request the vendor’s data-governance documentation and evidence of DPDP-aligned processing. The depth of diligence should scale with the risk category identified in the liability matrix.

SLA and performance metrics

Define measurable performance obligations: accuracy or error-rate thresholds appropriate to the use case, explainability expectations where decisions affect individuals, and uptime commitments for the service. Tie these to service credits and to the remediation regime, so a persistent failure to meet accuracy or availability targets triggers concrete remedies rather than abstract dispute.

Practical clause bank, AI contracts India model clauses to adapt

The clauses below are drafting starting points for AI contracts India negotiations. Each is a template to be adapted to the transaction and verified against current law. Every clause is labelled Draft, for negotiation; verify against law.

• IP assignment / licence. “Subject to payment, [the Vendor assigns to the Buyer all right, title and interest in the Outputs] / [the Vendor grants the Buyer an exclusive, worldwide licence to use, modify and sublicense the Outputs within the Field of Use].” Draft, for negotiation; verify against law.
• Training-data warranty. “The Vendor warrants that all training data was lawfully acquired, that it holds the rights necessary to use such data, and that any personal data was processed in accordance with the Digital Personal Data Protection Act, 2023.” Draft, for negotiation; verify against law.
• Indemnity. “The Vendor shall defend and indemnify the Buyer against third-party claims arising from IP infringement, privacy breach or non-compliance with applicable rules under the Information Technology Act, 2000 caused by the Vendor’s model or training data.” Draft, for negotiation; verify against law.
• Liability cap. “Each party’s aggregate liability is capped at [●], save that the cap shall not apply to indemnified IP claims, breach of confidentiality, gross negligence, wilful misconduct, or regulatory penalties arising from the Vendor’s default.” Draft, for negotiation; verify against law.
• Labelling and attribution. “The Vendor shall ensure all synthetic or AI-generated Outputs are labelled as required by applicable law and shall maintain such labelling throughout the term.” Draft, for negotiation; verify against law.
• Audit and right to inspect. “The Buyer may, on reasonable notice, review the Vendor’s data-governance records and audit reports, and may commission an independent audit at the Vendor’s cost where a material breach is reasonably suspected.” Draft, for negotiation; verify against law.
• Regulatory cooperation. “The Vendor shall maintain a designated point-of-contact and shall cooperate with the Buyer in responding to takedown notices, data-principal requests and lawful regulatory requests within the applicable timeframes.” Draft, for negotiation; verify against law.

Adapt each clause to your transaction and to the rules in force at the date of contracting.

Conclusion, checklist and next steps for counsel

Drafting AI contracts India in a fast-evolving regulatory environment means treating regulatory duties as contract terms, not aspirations. The essentials are clear: fix output ownership expressly because the statute will not resolve it; warrant training-data and model provenance and back the warranty with an indemnity; size liability caps with carve-outs that keep regulatory penalties and IP claims where they belong; map each applicable IT-framework and DPDP duty to an operative clause with a monitorable control; and pair liability with a practical remediation and incident-response regime. Counsel who work through this checklist, and who negotiate from a risk matrix rather than a template, will produce AI contracts India that survive both commercial pressure and regulatory scrutiny.

Adapt the model clauses above to your transaction, verify each against current law, and treat them as the start of a negotiation, not the end of one.

This article is for general information and does not constitute legal advice. Sample clauses are drafts to be adapted and verified with qualified counsel.

Sources

1. Ministry of Electronics & Information Technology (MeitY)
2. Gazette of India (official publications)
3. India Code / Legislative Department (Acts of Parliament)
4. Supreme Court of India (Judgments portal)
5. Bar Council of India
6. OECD AI Policy Observatory
7. NITI Aayog