Liability caps and super-caps in data licensing agreements

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why a fees-paid cap often fails a training data deal

A general cap tied to fees paid is the market starting point, but it is a poor proxy for risk when the license fee is small and the model built on the data is valuable. Published standard data agreements, including master data license frameworks that place individual purchases on order forms, are useful benchmarks for how a liability section is drafted [1][2]; check whether the cap is keyed to amounts paid under the agreement or under a specific order. On a modest first purchase, 12 months of fees can be a very small number next to the cost of a contaminated model.

The mismatch is structural, not a drafting accident. A licensor prices the data on its value to the licensor, while the buyer's exposure scales with compute, product launches and downstream customers. Treat the following as a working hypothesis to test in each negotiation rather than a market statistic: the smaller the deal, the less the general cap matters relative to the carve-outs.

Common cap measures counsel will see:

• Fees paid in the prior 12 months: the most common default; near zero early in a relationship or on a one-time purchase.
• Fees paid or payable under the order form: better for one-time datasets, since it counts the full committed price.
• Greater of a fixed dollar floor or a fees multiple: protects small deals; the floor is the real negotiation.
• Aggregate versus per-claim: an aggregate cap shared across all claims can be exhausted by one incident.
• Per-order versus agreement-wide: under a master agreement, check whether each order form carries its own cap or draws on one pool.

Which carve-outs belong in a super-cap and which stay uncapped

A super-cap is a separate, higher ceiling for named risks that the general cap would undervalue, and in data licenses it usually covers confidentiality, data protection and IP indemnity. Fully uncapped items are normally limited to fraud, gross negligence or wilful misconduct, and sometimes the licensee's use outside the licensed scope. Drafting the triggers precisely matters more than the dollar figure.

Data protection breach. Licensed operational data can carry personal information that was supposed to be removed. Under the GDPR, anonymous information sits outside the regulation only when identification is not reasonably likely given the means available [3], and California's definition of deidentified information depends on reasonable measures, a public commitment not to reidentify, and contractual obligations on recipients [4]. Because PII detection tools warn they cannot guarantee finding every identifier [7], residual personal data is a foreseeable failure mode, and a super-cap for it is reasonable rather than exotic.

Re-identification by the licensee. This is the mirror image, and licensors will ask for it to be uncapped or super-capped against the buyer. HIPAA limited data sets, for example, may be disclosed only under a data use agreement in which the recipient agrees not to identify the individuals or contact them [5]. Expect the licensor to treat a re-identification attempt as outside any cap; see re-identification risk assessment methods for what you can verify before signing.

Biometric and statutory damages exposure. Statutes with private rights of action change the arithmetic. Illinois BIPA sets liquidated damages per violation, though the 2024 amendment treats repeated collection of the same identifier from the same person as a single violation [6]. If voice recordings or facial images are in scope, a fees-based cap will not track the exposure.

IP indemnity. Third-party claims that the licensed data infringes copyright or misappropriates trade secrets are usually handled by an indemnity with its own cap. Scope, defense control and remedies are covered in IP indemnities for licensed training data and the indemnification glossary entry; this page covers only how that indemnity sits against the overall limitation.

Confidentiality. Uncapped confidentiality is common in NDAs but dangerous in data licenses, because the dataset itself is the confidential information. A licensor may argue any out-of-scope training is a confidentiality breach and therefore uncapped. Buyers should define confidentiality breach as unauthorized disclosure to third parties and route use-scope disputes to the license remedies instead.

How consequential damages exclusions hit retraining costs

A standard exclusion of indirect, consequential, special and lost-profit damages can quietly bar a buyer's largest real loss: the cost of purging tainted data and retraining a model. Whether retraining is "direct" or "consequential" is fact-specific and jurisdiction-dependent, so the safe course is to name it. The same applies to regulator fines, notification costs and customer credits.

Drafting moves that remove the ambiguity:

1. Define recoverable direct damages to include reasonable costs of identifying, removing and replacing nonconforming data, and of retraining or fine-tuning affected models up to a stated amount.
2. Exclude those costs from the consequential bar expressly, rather than relying on case law about foreseeability.
3. Tie the remedy to the warranty so a breach of a provenance or de-identification warranty triggers the replacement-and-retrain remedy; see data warranties buyers should require.
4. Coordinate with deletion terms, since a remedy that requires deleting data from a trained model collides with what the license says about model retention after termination and deletion and return.

Under the GDPR, data subjects have their own compensation rights and supervisory authorities can impose fines [3]. Contract allocation does not change who regulators pursue, but it determines who bears the cost between the parties.

Should licensor and licensee caps be symmetrical

Symmetry is a reasonable default for the general cap but rarely right for the carve-outs, because each side controls different risks. The licensor controls what is in the dataset and whether it had the rights and consents to license it. The buyer controls access, security, scope of use and any attempt to re-identify.

A balanced position, offered as a hypothesis for negotiation rather than a market norm:

• Mutual general cap and mutual consequential damages exclusion.
• Licensor super-cap for data protection breach caused by data as delivered, and for IP indemnity.
• Licensee super-cap for security incidents in its environment and out-of-scope disclosure.
• Mutual uncapped fraud and wilful misconduct; re-identification by the licensee likely uncapped on the buyer side.
• No uncapped "breach of license scope" against the buyer; route it to a defined remedy such as true-up fees or deletion.

Clause structure worksheet for counsel

Use this table to map the limitation clause before redlining, one row per layer. It is designed to sit alongside the AI data license negotiation checklist.

Illustrative example: invented to show structure; it does not describe an available dataset.

Two checks worth running before redlining: confirm every carve-out in the indemnity section also appears in the limitation section, and confirm the "payable" fees base covers committed future orders under a master agreement and order forms.

Sector overlays that move the numbers

Regulated data types justify carve-outs that a general-purpose license would not need. Health data shared as a limited data set comes with a mandatory data use agreement whose recipient obligations should map to uncapped or super-capped buyer duties [5]; the HIPAA limited data set guide covers that structure. Biometric data brings statutory damages [6], and EU personal data brings GDPR compensation and fines [3].

For operational data such as support transcripts or finance workflows, the realistic failure mode is residual identifiers in free text. PII detection tools state they cannot guarantee finding all sensitive information [7], so negotiate the data super-cap around that residual risk rather than around a promise that it is zero.

Where SourceX fits in the liability conversation

SourceX sources operational datasets from US companies and manages the commercial process, including licensing agreements. Every dataset is rights-reviewed for ownership and consents and delivered under a license that defines the records, allowed uses, term and delivery. Personal details such as names, emails, phones and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Counsel sizing data carve-outs should still treat residual identifiers as a live risk; you can describe the data you need on the SourceX buyers page.

Terms, including pricing, are agreed per deal, and nothing is contracted until a supplier agrees. For the wider framework, start from the data licensing for AI training hub or the AI data buyer's guide.

Licensing operational data with a liability framework you can defend

SourceX looks for US businesses that hold the data you describe, assesses data and licensing permissions, and agrees pricing and allowed uses in a license before anything is delivered. Every release is approved by the supplying company, and diligence materials on source, rights, preparation and allowed use are prepared per dataset. Tell SourceX what data your team needs.