Rights and contracts
Media asset permission conflict dataset: when rights fields disagree
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A media asset permission conflict dataset records one requested use, every document that grants or limits it, and one of five decisions: approve, approve with conditions, substitute, hold for rights review or decline. Conflicts come in four kinds, and only priority conflicts have a winner. In National Gallery of Art image data, 6,213 of 60,178 rights-restricted images carry no resolution cap.
License terms
Media asset permission
Key takeaways
- Sort every conflict first: priority conflicts have a contractual or legal winner, stacking conflicts need every permission at once, signal conflicts are evidence problems, and merit leaks are not permission at all.
- In the National Gallery of Art's open data (27 October 2026), 6,213 of 60,178 rights-restricted image records carry no resolution cap, although each is larger than the 900-pixel cap on 53,956 others.
- Curatorial selection does not settle usage authority: 483 of the 2,471 objects on view with published images have only restricted images.
- Creation date is a poor proxy: 1,242 objects made before 1900 have restricted images, and 22,868 made in 1900 or later are open access.
- Hold for rights review is a correct answer whenever the governing document is missing or two signed grants overlap; grade it that way.
What does a media asset permission conflict dataset need to capture?
It needs one requested use of one asset, every record that grants, limits or describes permission for that use, and the decision a rights reviewer made, chosen from a short label set: approve, approve with conditions, substitute, hold for rights review or decline. The conflict is the reason the case exists. Two records disagree, several permissions must all hold at once, or a status field says something the governing document does not.
Those cases are where AI help is most tempting and most risky. A creative operations team pushing thousands of requests a year through a digital asset management system wants a model to clear the easy ones and flag the rest. A model trained only on past approvals learns that popular assets usually get used. A model trained on conflict cases learns when the right answer is to stop, attach a condition or offer a different asset.
The sections below define the decision, sort conflicts into four kinds, count how often two permission fields disagree in a large public image archive, and set out a label schema, an illustrative casebook, routing rules and the limits of what any licensor's records can show. Clearing a single asset for a single use and modeling a license that expires are covered on their own pages; this one is about the cases where the permissions themselves pull in different directions.
Four conflicts that look alike but resolve differently
Classify the conflict before labeling the decision, because only one of the four kinds has a winner. A priority conflict is two grants of the same right to different parties, and contracts, sometimes backed by statute, decide which prevails. A stacking conflict is a set of different permissions that must all hold at once, so the narrowest one controls the use. A signal conflict is a disagreement between records about what was granted, which makes it a question of evidence. A merit leak is creative preference presented as if it were permission.
Copyright law shows how a priority conflict can be settled on paper. The Copyright Act defines a transfer of copyright ownership to include an exclusive license but not a nonexclusive one. Section 205(e) then says when a nonexclusive license prevails over a conflicting transfer: the license must be evidenced by a written instrument signed by the owner of the rights or an authorized agent, and it must have been taken before the transfer was executed, or in good faith before the transfer was recorded and without notice of it. A dataset can store those dates and signatures; a model cannot infer them.
Stacking conflicts have no winner at all. Section 2(b)(1) of the Creative Commons BY 4.0 legal code says publicity, privacy and similar personality rights are not licensed under it, and New York Civil Rights Law section 50 makes it a misdemeanor to use a living person's name, portrait, picture, likeness or voice for advertising or trade without first obtaining that person's written consent. A photo can carry a valid copyright license and still be unusable in an advertisement because a second permission is missing. Each kind, what settles it and the label to use until it is settled:
| Conflict | What disagrees | What settles it | Default label until resolved |
|---|---|---|---|
| Priority | Two signed grants cover the same right, period and territory for different parties | Contract terms and signature dates; for an exclusive grant against a nonexclusive one, the conditions in 17 U.S.C. 205(e) | Hold for rights review |
| Stacking | A copyright license allows the use, but a release, consent or third-party clearance is missing or narrower | Every required permission on file for this use, with the narrowest one setting the scope | Substitute or decline; hold only while the missing permission is being sought |
| Signal | Two records about the same grant disagree, such as a status field and the document it summarizes | The governing document, located and read | Hold for rights review |
| Merit leak | A test score, award or curatorial pick is offered as the reason to use an asset | Nothing about merit; the case is decided on the permissions alone | Whatever the permissions support, with the score ignored |
How often do permission fields disagree in a public image archive?
Often enough to matter, even in a carefully run archive. The National Gallery of Art publishes the image records behind its website as open data, and each record carries two permission fields. In openaccess, 1 marks an open access image with no usage restrictions under the museum's policy and 0 marks a rights-restricted image available only under fair use. The data dictionary describes maxpixels as a limitation used to enforce fair use: above that size, the image server will serve the image only from a copy resized to fit within it.
Read together, the fields should agree, with every restricted image carrying a cap. In the snapshot at commit d2a43c3 (27 October 2026), the 129,611 image records split into 69,433 open access images, none of them capped, and 60,178 restricted images. Of the restricted images, 53,965 carry a cap, 53,956 of them at 900 pixels and 9 at 4,000, while 6,213, or 10.3 percent, carry none. None of the 6,213 is small enough to make a cap unnecessary: the shortest longer side among them is 1,200 pixels and the median is 5,166.
The records cannot say which field is right. A blank cap could mean the rights holder allowed larger display, that the field was never filled in, or that size is limited somewhere else. The source images behind the uncapped records were created in every year from 2004 to 2025, plus one image each dated 1980 and 2026, with the largest groups in 2012 and 2014, which looks more like accumulated practice than one bad import, and photographs make up 3,234 of the 6,007 objects whose restricted images are all uncapped. That is the kind of case a permission dataset should teach a model to send to a person instead of resolving it.
Is a work chosen for the galleries cleared for reuse?
No, and the archive shows how far apart the two questions sit. Selection for display is the closest public stand-in for creative merit in a museum, since curators chose these works for the walls. Of the 2,471 objects with published images whose recorded location is a public room on the museum's campus, 483, or 19.5 percent, have only restricted images. Among the 117,553 objects with images that are not on view, 47.4 percent are restricted. Display says something about the collection and nothing about the license.
Age is a weaker shortcut than it looks. Of the 42,201 objects with images whose recorded end year is before 1900, 1,242 have restricted images; of the 77,745 dated 1900 or later, 22,868 are open access. A rule that cleared everything old and blocked everything new would mislabel 24,110 objects in this one archive. Permission follows documents, not dates.
For a dataset the lesson is structural. Keep creative signals, such as a reviewer's score, an A/B test result or a pick for a campaign, in fields that the authority label never reads, and build evaluation pairs in which the creative favorite is not cleared and the cleared asset is creatively weaker. A model that scores well only when merit and authority happen to agree has learned the wrong rule.
The label schema: five decisions and what each must cite
Two labels travel with each case. The decision label is what a rights specialist concluded from the documents, recorded at the time or re-derived in review. The outcome label is what happened next, read from later records: a takedown notice, a claim or invoice from a rights holder, a renewal, or nothing within a window fixed in advance, such as a year after first use. Silence is weak evidence, because many unlicensed uses are never noticed, so cases whose window is still open stay censored rather than counting as clean.
Each case gets exactly one of five decision labels, and a label counts only when the answer cites the records that justify it. Conditions and substitutes are full answers in this schema, because most conflict cases end there rather than in a clean yes or no:
| Label | Correct when | The answer must cite | Mislabel to catch |
|---|---|---|---|
| Approve | Every required permission covers the requested use as asked | Each governing document and the clause or field that covers the use | Approving from a status field without reading the document behind it |
| Approve with conditions | The use fits only if it is narrowed or a duty is met, such as a credit line, an end date or a channel limit | The permission, the condition and who must carry it out | Dropping the condition, which turns a partial yes into an unlicensed use |
| Substitute | The requested asset fails and a cleared alternative meets the brief | Why the original fails and the alternative's own permissions | Picking the alternative on creative fit without checking its rights |
| Hold for rights review | The governing document is missing or ambiguous, or two signed grants overlap | The specific conflict and the document that would settle it | Guessing a winner, or scoring a hold as a failure to answer |
| Decline | No permitted path exists for this use and no cleared alternative is available | The missing or blocking permission | Declining a use that a condition or a substitute would have allowed |
Casebook: eight illustrative conflicts with gold labels
These cases were written for this page and are illustrative; no company, asset or person in them is real. Each gold label comes from the five-label schema, and the last column names the trap a model is most likely to fall into:
| Case (illustrative) | Facts at decision time | Gold label | Trap (hard negative) |
|---|---|---|---|
| 1. Packaging shot | Taken by an in-house photographer on assignment; no people, no third-party marks; packaging use requested | Approve | Over-caution: holding a clean case teaches a model that every request is a conflict |
| 2. Shorter cut of a licensed spot | Music licensed for a 30-second spot in one campaign, medium and term; a 15-second cut is requested for the same campaign, medium and term | Approve with conditions | Leaving out the condition that the cut stays inside the original campaign, medium and term |
| 3. Credited web use | License allows editorial web use with a credit line in a set form; a web article is requested | Approve with conditions | Approving without the credit line the license requires |
| 4. Campaign hero image | Top-scoring frame in creative testing; the model release excludes alcohol advertising and the client is a brewer; another frame from the shoot has an unrestricted release | Substitute | Merit leak: the test score is offered as the reason to approve |
| 5. Sponsored social post | Stock image licensed for editorial use only; a commercially licensed image from the same series is already in the library | Substitute | Reading editorial use as covering any online post |
| 6. Overlapping grants | Agency signed an exclusive web license for a territory with one client and a nonexclusive license for the same territory with another; dates and recording unclear | Hold for rights review | Picking the later or larger client instead of applying the documents |
| 7. Release on file, says who? | Asset system field reads release on file; embedded file metadata reads no release; no release document is linked | Hold for rights review | Trusting whichever field the model read first |
| 8. Open license, recognizable face | Photo of a recognizable person under Creative Commons BY 4.0; requested for a paid advertisement in New York; no release and no alternative | Decline | Treating an open copyright license as consent to commercial use of a likeness |
Routing rules that send a case to rights review
A model, or a new coordinator, should send a request to a rights reviewer whenever any rule below fires. The same rules serve as the checklist a domain reviewer runs before cases enter a training or evaluation set:
At a glance
Routing rules that send a case to rights review
- Two signed grants cover the same right, period and territory for different parties.
- A status, tag or flag disagrees with the document it summarizes, or that document cannot be found.
- A person is recognizable, the use is advertising or trade, and no written release covering that use is on file.
- The license lists channels or media and the requested one is not among them.
- The asset contains third-party material, such as artwork, music, a logo or a building, with no recorded clearance.
- Ownership changed hands after the grant, through an assignment, an acquisition or a disputed work-for-hire claim.
- A takedown, objection or revocation request is logged against the asset or a person in it.
- The stated reason to use the asset is creative performance rather than a cited permission.
- Before any case enters the dataset, a rights specialist confirms its gold label and records which document settled it.
What a licensor's asset records can and cannot show
A media company's records can show the request, the documents on file, the decision, who made it and what followed, such as a takedown, a renewal or a complaint. That is enough to teach the routing behavior above and to test whether a model cites the right document. They cannot show permissions given orally or implied by conduct, how a court would weigh two grants, or whether a decision that drew no complaint was legally sound.
The strongest objection is that a sector label proves little. A dataset sold as media permission cases may come from a company whose client contracts forbid sharing them, whose releases contain personal data, or whose legal context differs from the buyer's. All three points are true, and each changes the scope rather than the idea: client agreements decide whether decision records may leave the company, releases and request logs need de-identification, and a gold label made under one state's law should carry that jurisdiction as a field.
Licensing the decision records is also separate from licensing the creative works. A dataset can describe each asset by identifier, type and permission history without delivering the image, film or track, and whether the works themselves could ever be licensed for training is a question for their rights holders, not for the team that kept the records. None of this is legal advice; counsel should review any specific clearance decision.
Requesting permission-conflict records for training or evaluation
Ask for cases, not catalogs. A useful request names the decision (one requested use of one asset), the five labels, the documents each case must link to (license, release, client agreement, status history and reviewer notes), a minimum share of hold and decline cases, and an acceptance test in which a reviewer re-derives the gold label for a sample of cases from the linked documents alone.
SourceX licenses data directly from the companies that hold it and supplies it to AI developers in its own name, sourcing each dataset on request rather than from stock. Through its data request form, a buyer can describe the decision, the label schema, the outcome window, the de-identification needed for people named in releases and request logs, and the permitted use, and SourceX follows up about potential sources. A request written this way states a need; it does not imply that such records are available today.
Methods note: the National Gallery snapshot and counting rules
Data: published_images.csv, objects.csv and locations.csv from github.com/NationalGalleryOfArt/opendata at commit d2a43c3a834a2c52e05270f3ce04418505d6c3e0 (27 October 2026), cloned on 28 October 2026 and read by nga_permission_signals.py, a standard-library Python script with no randomness that runs no repository code. None of the three files begins with a byte-order mark, and the script would strip one.
Rules as applied: every row of published_images.csv is an image record. Open access means openaccess equals 1 and restricted means 0; no other value occurs. A cap is a maxpixels value that is not empty after trimming spaces, and the longer side is the larger of width and height (a blank dimension would count as 0; none occurs among the uncapped images). Images join to objects on depictstmsobjectid, which every image row carries and which always matches objects.csv; an object is open when all its images are open and restricted otherwise, and no object mixes the two. On view means the object's locationid matches a row of locations.csv whose site is East Building, West Building or Sculpture Garden (WSG). The date band uses the integer endyear, before 1900 or 1900 and later; 78 objects whose end year is blank or not an integer are left out of both bands. Image creation years are the first four characters of the created field, and object type is the classification field as recorded. An object counts as uncapped when every one of its restricted images lacks a cap; 4 restricted objects mix capped and uncapped images and fall in neither group. The median longer side is the middle value of the 6,213 sorted values.
Limits: the two fields record how the museum serves images, not the rights documents behind them, so the count measures disagreement between fields, not error. The museum refreshes the data daily, so later snapshots will differ. Display location stands in for curatorial selection. The conflict typology, label schema, casebook and routing rules are editorial frameworks drawn up on 28 October 2026, and the cases are illustrative. Saved outputs hold counts only, with no titles, names or attributions.
Frequently asked questions
What is a media asset permission conflict?
It is a case where the records governing a requested use of an asset disagree, overlap or fall short: two signed grants of the same right, a license without the release the use also needs, or a status field that contradicts its source document. Each kind calls for a different decision, so a dataset should label the kind of conflict as well as the outcome.
Does a Creative Commons license clear a photo for advertising?
It covers the copyright the licensor holds, on the license's terms, but Section 2(b)(1) of CC BY 4.0 says publicity, privacy and similar personality rights are not licensed. If the photo shows a recognizable person, an advertising use may also need that person's written consent, as New York Civil Rights Law section 50 requires there. Counsel should review the specific use.
Why is hold for rights review a valid label?
Some conflicts cannot be settled from the records at hand, and a model that guesses a winner can authorize an unlicensed use. A hold that names the conflict and the document that would settle it is the correct and useful answer, and graders should score it that way.
Can creative performance data predict whether an asset is cleared?
It should not be used that way. In the National Gallery of Art's open data, 483 of the 2,471 objects on view with published images have only restricted images, so even expert selection says little about usage authority. Keep creative scores in separate fields and test models on pairs where the favorite is not cleared.
Do you need the images themselves to build a permission conflict dataset?
No. The decision records, grant documents, release metadata and status histories carry the reasoning, and each asset can be represented by an identifier and a description. Whether the creative works may be licensed for AI training is a separate question for their rights holders.
Sources
- The National Gallery of Art's open data files at commit d2a43c3 (27 October 2026): 129,611 published image records and 146,458 object records, whose permission, location and date fields were counted for this article. Source
- The National Gallery of Art data dictionary defines openaccess (1 open access with no usage restrictions, 0 rights restricted under fair use), maxpixels as a limitation used to enforce fair use, the image created date and public locations. Source
- The National Gallery of Art releases its collection dataset under CC0, says images are not included, and says the data are updated frequently, usually once a day. Source
- The U.S. Copyright Office quotes the statutory definition: a transfer of copyright ownership includes an exclusive license but not a nonexclusive license. Source
- Section 205(e) of the Copyright Act sets when a signed nonexclusive license prevails over a conflicting transfer of copyright ownership. Source
- Creative Commons BY 4.0 legal code, Section 2(b)(1): moral rights and publicity, privacy and similar personality rights are not licensed under the public license. Source
- New York Civil Rights Law section 50 makes advertising or trade use of a living person's name, portrait, picture, likeness or voice without prior written consent a misdemeanor. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.