Skip to content

Rights and contracts

AI dataset performance claims rights: can the buyer cite your name?

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A buyer should use your company name in benchmark or performance claims only when the license expressly allows that specific claim. A safe supplier default is no name, logo or identifying description in any performance claim; anonymous or aggregate claims only when the buyer holds evidence; and every named claim approved in writing for exact wording, medium and period, with a right to withdraw.

License terms

AI dataset performance

This page covers AI dataset performance claims rights: can the buyer cite your name.Illustrative

Key takeaways

  • A performance claim that carries your name can read as an endorsement, which is different from announcing a deal.
  • Classify each request by how identifiable the attribution is; a precise description can identify you as surely as a name.
  • The buyer making the claim needs evidence for it; ask for protocol, data version, metric, sample size and date before approving.
  • Approve exact wording for a named medium and period, tied to a data version, with a right to withdraw.
  • Comparison claims, regulated settings and claims already published without approval go to counsel first.

Can the buyer use your company name in its benchmark claims?

Only if the license says so, and only for the specific claim you approved. The safest default for a data supplier is that the buyer may not name you, show your logo or describe you in a way that identifies you in any statement about model performance, and may make anonymous or aggregate performance claims only when it can back them up. Any named claim then becomes a separate, written approval of exact wording.

This question is narrower than whether a buyer may announce a data deal. An announcement says the two companies worked together. A benchmark or performance claim says something measurable happened because of your records: a model scored higher, made fewer errors or was validated on your data. That ties your name to a result you did not produce, may not be able to check and cannot easily retract once it is in a press release, a model card or a sales deck.

The rest of this page sets out why that distinction matters, a matrix for deciding each kind of request, and the assumptions that should change your answer. Treat it as an issue map for your counsel, not as a substitute for their review of your agreement.

Why a performance claim carrying your name reads as an endorsement

U.S. advertising rules can treat a company's name attached to a product claim as an endorsement, which is why performance claims deserve more care than a logo on a customer page. Under the FTC's Endorsement Guides, an endorsement is an advertising or promotional message that consumers are likely to believe reflects the views of a party other than the advertiser, and the guides say the name or seal of an organization can be an endorsement.

The same guides say an advertiser may keep using an endorsement only while it has good reason to believe the endorser still holds the views presented, and that an advertiser can be liable for a deceptive endorsement even when the endorser is not. For a supplier, the practical reading is simple: if a buyer's claim implies you vouch for its model, you have an interest in the wording and in being able to withdraw it later.

Two federal statutes sit behind this. Section 5 of the FTC Act declares unfair or deceptive acts or practices in or affecting commerce unlawful. Section 43(a) of the Lanham Act, 15 U.S.C. 1125(a), gives a civil action against uses in commerce likely to cause confusion about a person's affiliation, connection or association with another, or about another person's sponsorship or approval of goods, services or commercial activities. Whether either applies to a particular claim is a question for counsel; the point for the license is that both parties have reasons to want the wording settled in advance.

How identifiable is the attribution being proposed?

Start by classifying how identifiable the attribution is, because the same performance number carries very different risk at each level. Removing your company name from the records themselves does not settle this; a description can identify you as surely as a name.

  • Named with logo or trademark: your name and brand appear next to the claim, for example in a launch post or a slide.
  • Named in text: the claim names your company, for example that a model was evaluated on your service records.
  • Identifiable description: no name, but a description that only fits you or a handful of companies, such as the largest operator of a given service in one state.
  • Category description: a description shared by many companies, such as a U.S. field-service business with more than five years of work orders.
  • Unattributed aggregate: the claim describes results on licensed operational data without describing the source.
  • Machine-readable attribution: dataset identifiers or evaluation entries in a model card or leaderboard. Hugging Face's documentation, for example, says model cards should describe which datasets were used to train a model and its evaluation results, that listed datasets are displayed on the model page and can be used to filter models, and that evaluation results can be submitted through pull requests and marked as community-provided.

The publicity and attribution approval matrix

Use the matrix to set the license default for each claim type and the conditions under which you would approve an exception. The defaults below are a conservative starting point for a supplier, not a market standard; adjust them with counsel to fit your agreement and your appetite for visibility.

The publicity and attribution approval matrix
Claim typeDefaultConditions for approvalApprover and evidence
Named result in marketing, press or sales materialNot permittedExact wording approved in writing, per use, with a right to withdrawExecutive owner and counsel; test protocol and data version
Named or identifiable comparison against a competitor's modelNot permittedRarely worth approving; requires the full comparison methodCounsel; method, both test conditions and dates
Statement that the supplier validated or verified resultsNot permittedOnly if your own team actually ran and checked the evaluationExecutive owner; your own evaluation record
Category description with a performance resultPermitted with noticeDescription approved once; claim states sample size, date and metricData owner; claim text and method summary
Unattributed aggregate result on licensed dataPermittedBuyer holds substantiation before publishing; no record content shownNone in advance; buyer keeps evidence on request
Peer-reviewed paper or technical reportCategory description onlyPre-publication review window for identification riskData owner and counsel; manuscript draft
Model card, dataset listing or leaderboard entryNo supplier identifierAny identifier or link approved in writing firstData owner; proposed metadata fields
Investor, board or due diligence material under confidentialityPermitted under NDANamed only where the license allows confidential disclosureCounsel; confidentiality terms

Which requests fall outside the matrix and need their own clause?

Some uses fall between the rows of any matrix, and they should be decided expressly in the license instead of left to the buyer's reading of a general publicity clause.

  • Claims made after the license ends. Decide whether a named or category claim may stay published, must be removed, or may remain only in archived material.
  • Claims by affiliates, resellers or customers of the buyer. A downstream customer repeating a claim in its own marketing is outside your approval unless the license reaches it.
  • Claims about a model trained on several suppliers' data. A named claim can suggest your records drove the result even when they were a small share of the training mix.
  • Leaderboards and public benchmarks built partly from your records. Listing them can identify you through the benchmark's name or description.
  • Sales conversations with your own customers or competitors. A buyer telling a prospect that its model learned from your records is a claim even if nothing is published.
  • Synthetic or derived examples shown alongside a claim. They can carry recognizable details of your operations and should be reviewed as examples, not just as claims.

Who has to prove the claim, and what should you see before approving it?

The buyer making the claim needs the evidence for it, and the license should say so. The FTC's advertising substantiation policy statement reaffirms that advertisers need a reasonable basis for objective claims before the claims are disseminated, and lists factors including the type of claim, the consequences of a false claim and the amount of substantiation experts in the field consider reasonable. In April 2025 the FTC announced a proposed order against Workado, an AI detection company, that would bar efficacy claims unless the company has competent and reliable evidence supporting them when they are made, and it approved the final order in August 2025.

For a supplier asked to approve a named or category claim, that translates into a short evidence request: the evaluation protocol, the data version and date range used, the metric and how it was computed, sample size, the comparison baseline, the date of the run and who ran it. You are not auditing the buyer's model. You are checking that the claim describes what was actually measured on your data and does not imply you checked it.

Write the approval so it covers exact wording, a named medium and a period. Approving a sentence for a launch blog post is different from approving it for every future deck. A right to withdraw approval when the data version, the model or the evaluation changes keeps the claim honest over time and mirrors the endorsement principle that a claim should be used only while it remains true.

Which assumption would flip your answer, and how should the claim show doubt?

Three assumptions move the decision more than any other, and the claim's wording should show the uncertainty behind each one.

To show uncertainty, the approved claim should state the sample size, the date, the metric and, where the buyer has it, a confidence interval or range across runs rather than a single best score. A sentence that reads as a permanent fact about your data should be rewritten as a dated result of one evaluation.

  • The test data was really yours and really held out. If the evaluation set overlapped with training data, the named result may be inflated. Ask how the split was made and whether the claim was rerun on a clean held-out set.
  • Your data version is the one described. Claims drift as buyers retrain. Tie each approval to a dataset version and evaluation date, and treat a new run as a new request.
  • The description cannot identify you. Test the category wording against your market: if a knowledgeable competitor could name you from it, treat it as a named claim.

Worked scenario (illustrative): three requests from one buyer

This scenario is illustrative and invented. A regional freight brokerage licenses three years of de-identified dispatch notes to a developer building a carrier-communication model. The license uses the matrix defaults above. Within six months the buyer sends three requests.

  • A launch blog post saying the model was trained and validated on the brokerage's dispatch history, with its logo. The supplier declines the logo and the word validated, because its team did not check the evaluation, and offers a category description instead.
  • A technical report stating a 12 percent reduction in routing errors on held-out dispatch notes from a mid-sized U.S. freight brokerage. The supplier asks for the protocol and data version, confirms the held-out split, and approves the sentence for the report only, with the sample size and evaluation date added.
  • A leaderboard entry whose benchmark name includes a shortened form of the brokerage's name. The supplier treats it as a named claim, refuses the name and approves a neutral benchmark title.
  • The exception: a year later the brokerage decides it wants to be named, as a signal to its own customers. It approves one named sentence in a joint case study, limited to that document, dated, with a right to withdraw if the buyer retrains on a different data version.

Signals that a claim request needs counsel before anyone answers

Most requests can be handled with the matrix and an evidence request. A few should go to counsel before anyone replies, because the wrong answer is hard to reverse or because law outside the contract may apply.

  • A comparison claim against a named competitor, or a claim that your data makes a model safer or more accurate in a regulated setting such as lending, insurance or employment.
  • Any claim that would appear in a securities filing, investor offering material or a regulator submission.
  • A request to describe your customers, staff or partners as the source of examples, which raises privacy and confidentiality questions beyond attribution.
  • A claim already published without approval. Preserve a copy, check the license's remedy and notice clauses, and let counsel make the first contact.
  • A buyer asking to name you in jurisdictions where your agreement's governing law or your own customer contracts impose separate consent requirements.

How are attribution terms set in a license arranged through SourceX?

A license arranged through SourceX is signed by the company, scope and terms included, before any data is delivered, after which the de-identified dataset can be used, sublicensed and distributed only on the agreement's terms. Publicity and attribution limits are part of what the company reviews before signing, not a courtesy requested afterward.

SourceX handles rights review and contracting alongside data preparation, so the attribution defaults a company prefers can be set alongside the scope of use. Eligibility is answered up front by a short company assessment that asks for no data.

Frequently asked questions

Can a buyer say its model was trained on our data without naming us?

Usually yes, if the license allows unattributed claims and the description cannot identify you. Check the wording against your market: a description that fits only a few companies works like a name and should be approved as one.

Is a publicity clause enough to cover benchmark and performance claims?

Often not. Publicity clauses tend to address announcements and logo use, while performance claims tie your identity to a measured result. Add a separate attribution clause that covers claims, evidence, media, duration and withdrawal.

Who is responsible if a named performance claim turns out to be wrong?

The party making the claim needs a reasonable basis for it, and the FTC's guides say an advertiser can be liable for a deceptive endorsement even when the endorser is not. Your exposure depends on the facts and your contract, so counsel should review any claim already published.

Should we ask for the buyer's evaluation results before approving a claim?

Ask for enough to confirm the claim describes what was measured on your data: protocol, data version, metric, sample size, baseline and date. You are not validating the model, and the approval should say so.

Can we withdraw approval for a claim we already allowed?

Only if the license gives you that right, which is why approvals should be tied to exact wording, a medium, a period and a data version. Without a withdrawal right, a removal request depends on the buyer's cooperation.

Sources

  • The FTC Endorsement Guides define an endorsement as a promotional message consumers are likely to believe reflects the views of a party other than the advertiser, and say an organization's name or seal can be an endorsement. Source
  • The Endorsement Guides say an advertiser may keep using an endorsement only while it has good reason to believe the endorser still holds the views presented, and that an advertiser may be liable for a deceptive endorsement even when the endorser is not. Source
  • Section 5 of the FTC Act declares unfair or deceptive acts or practices in or affecting commerce unlawful. Source
  • 15 U.S.C. 1125(a) provides a civil action against uses in commerce likely to cause confusion about affiliation, connection or association, or about sponsorship or approval of goods, services or commercial activities. Source
  • The FTC's advertising substantiation policy statement requires a reasonable basis for objective advertising claims before dissemination and lists factors such as claim type, consequences of a false claim and what experts consider reasonable. Source
  • In April 2025 the FTC announced a proposed order requiring AI detection company Workado to have competent and reliable evidence supporting efficacy claims when they are made. Source
  • Hugging Face's model card documentation says a card should describe which datasets were used to train a model and its evaluation results, and that datasets listed in metadata are displayed on the model page and used for filtering. Source
  • Hugging Face's evaluation results documentation says results stored in model repos appear on benchmark leaderboards and can be submitted through pull requests and marked as community-provided. Source
  • In August 2025 the FTC approved the final consent order against Workado after a public comment period. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies