AI-Generated Maps and Geospatial Data: What You Own and What You Licensed
Product teams now generate basemaps from prompts, extract building footprints from imagery with a segmentation model, and conflate half a dozen sources into one location dataset. Then someone asks who owns the result. The answer is rarely "copyright" in either direction — the protection is thinner than expected, and the restrictions come from contracts nobody on the team has read.
Start from the right premise: map copyright is thin
The instinct that an expensively assembled map must be heavily protected is wrong in the United States. The sweat-of-the-brow theory was rejected decades ago in the context of directories, and the reasoning applies directly to geospatial data: effort is not authorship. Where a road runs is a fact, and facts belong to nobody.
What survives is expression. A cartographic style — the palette, the symbol vocabulary, label placement and typography, the generalisation rules that decide which features disappear at zoom nine — is a set of creative choices and can be protected. So can an original selection and arrangement, if the selection reflects judgment rather than completeness. A dataset that aims to include every gas station has no protectable selection, because the selection rule is "all of them."
Four artefacts, four different answers
A rendered map image or tile
The most protectable artefact, because styling is expression. If a human authored the style specification, that authorship carries into the render. If the whole thing came out of a generative model on a prompt, US law gives you no copyright in it — and equally gives your competitor none in theirs.
Extracted vector geometry
Footprints, road centrelines, parcel boundaries pulled from imagery by a model. These are facts about the world. Copyright protection is weak to nonexistent, and your exposure comes from the imagery licence that governed the input, not from the output.
A conflated location dataset
Your records joined to third-party geometry and identifiers. This is where licences collide: each source's terms attach to the parts of the result derived from it, and the strictest term usually governs what you can ship. Provenance at the record level is the only way to answer questions about it later.
The model itself
Weights trained on licensed imagery. Many imagery agreements now speak directly to machine learning use, and some purport to restrict the resulting model, not just its outputs. Read the ML clause before you treat training as an internal analytics use.
The imagery licence is the real constraint
Commercial satellite and aerial imagery is almost never sold outright. It arrives under a licence with a defined territory, duration, user count, and — critically — a derivative works clause. A team that reasons "the footprints are just facts, so we can publish them" is answering a copyright question correctly and a contract question incorrectly. Typical restrictions include prohibitions on redistributing derived vectors, caps on how many named users may access the imagery, requirements that derived products carry the provider's attribution, and increasingly explicit consent requirements for training models. Breach of those terms does not need a copyright theory to hurt; it is a contract claim, and it usually comes with an audit right.
Government imagery is the notable exception worth exploiting. US federal works are generally not subject to domestic copyright, so public programme imagery and national datasets can often be used freely — which is why the compliance-cheapest architecture is frequently to build the public layer on government sources and keep commercial imagery confined to internal analysis.
ODbL: the share-alike trigger teams walk into
OpenStreetMap data is free, and the Open Database Licence attached to it is where most well-intentioned engineering teams create a real obligation without noticing. ODbL separates two things:
- A Produced Work — a rendered image, a printed map, a tile served to a user. Publish it with attribution and you are done. Your surrounding application stays proprietary.
- A Derivative Database — a database containing or built from the licensed data. Use one publicly and you must offer that derivative database under ODbL.
The failure mode is gradual. An engineer enriches an internal table with OSM geometry to fix bad addresses. A year later that table is the backing store for a customer-facing API. Nobody made a licensing decision at any point, and the product now publicly uses a derivative database. Track which columns came from which source at ingest, because reconstructing provenance after the fact across a conflated dataset is close to impossible.
Europe adds a right the US does not have
The EU sui generis database right protects substantial investment in obtaining, verifying, or presenting the contents of a database, independent of any copyright in the contents. That is precisely the sweat-of-the-brow protection US law refused. Systematic extraction of a substantial part of a European map database can infringe it even though every individual data point is a fact. If your scraping or model-training pipeline touches European geospatial providers, the US fact-based analysis does not travel — and repeated extraction of insubstantial parts can itself be treated as substantial.
A rights checklist before you ship
- Record provenance per layer and per field at ingest, not at launch. Source, licence, licence version, and date.
- Read the derivative works and ML clauses in every imagery agreement, and treat model training as a use requiring express permission unless the contract says otherwise.
- Decide Produced Work versus Derivative Database explicitly for anything touching ODbL data, and write the conclusion down where the next team will find it.
- Keep human authorship of the style if you want anything protectable. Document the design decisions and who made them; that record is what supports a claim over the rendering.
- Screen outputs for known fictitious entries if you train on scraped map data. A model that reproduces a provider's trap features has handed the provider its evidence.
- Segment by geography for European sources, where database right applies regardless of the copyright analysis.
- Publish attribution where required in the interface, not only in a legal page. Most open licences require attribution reasonably visible to the user of the map.
Frequently Asked Questions
Can you copyright a map?
You can copyright its expression, not its geography. Road positions, elevations, and boundaries are facts and are unprotectable regardless of collection cost. Cartographic styling, symbol design, label placement, colour choices, and an original selection and arrangement can be protected. That is why map copyright is called thin — competitors may depict the same terrain, just not copy your rendering of it.
Do I own a map an AI model generated?
In the US, purely machine-generated output has no human author and so no copyright. Human contributions layered on top — a hand-authored style specification, curated labels, an original legend — are protectable as to those contributions. The practical consequence is that a competitor copying your fully AI-rendered basemap may not be infringing anything at all.
Does extracting footprints from satellite imagery infringe the imagery?
The copyright answer is weak because the extracted coordinates are facts. The licence answer is usually where the problem is: commercial imagery is licensed, not sold, and those licences commonly restrict derivative products, redistribution, and machine learning use. Teams are caught by the contract rather than the statute.
What triggers ODbL share-alike?
Publicly using a Derivative Database rather than merely publishing a Produced Work. A rendered map or tile needs attribution only. Conflating OSM geometry into your own dataset creates a derivative database, and shipping a product on top of it obliges you to offer that database under ODbL. The common path into this is enriching an internal table and later exposing it through an API.
Are trap streets still relevant?
As evidence rather than as protection. A fictitious entry does not make facts copyrightable, but finding your invented entries in a competitor's dataset is near-conclusive proof of copying rather than independent collection, which supports licence-breach and misappropriation claims. Providers still seed them and models trained on scraped data reproduce them.
Is European map data treated the same way?
No. The EU sui generis database right protects substantial investment in a database independent of copyright in its contents, so systematic extraction of a substantial part can infringe even where every data point is a fact. Repeated extraction of insubstantial parts can also be treated as substantial. A US fact-based analysis does not carry over.