Geospatial Licensing: AI Extraction Risk, Not Just Training

A product team wants to ship a feature that flags storm-damaged roofs from aerial imagery. No training pipeline, no dataset, no model of their own—just a call to an off-the-shelf vision model over imagery they're already licensed to display, and a structured list of damaged roofs comes back.

They're still not in the clear.

Most of the conversation around geospatial imagery licensing focuses on the dramatic case—someone scrapes a platform's imagery and trains a competing model on it. That's real, and it's explicitly banned everywhere. But it's not the case most teams actually run into. The dominant pattern in 2026 isn't training foundation models on licensed imagery—it's shipping product features that call an existing AI model to extract, detect, or index information from that imagery. And the license language that matters here has almost nothing to do with "training."

What the API actually authorizes

Start with Google, because its terms spell out the extraction case explicitly rather than leaving it implied. Section 3.2.3(c), "No Creating Content From Google Maps Content," bans customers from creating content derived from Google's imagery—independent of whether any model gets trained. The named examples: tracing or digitizing roadways, building outlines, utility posts, or electrical lines from satellite basemap imagery; creating 3D building models from 45° imagery; building terrain models from elevation data; and constructing an index of tree locations from Street View imagery. Only the last item in that list—training, testing, validating, or fine-tuning machine learning and AI models—is actually about training. The other examples are extraction use cases: run the imagery through something, get back a structured answer, exactly the pattern most AI features use.

Apple's Maps Terms of Use draw the same line without separating it into a list. Section 1.3(vi) bans using the service to "copy, extract, scrape or reutilize any portion of the Service, including...creation of any databases based upon data or content provided through the Service, or training of any model." Extraction and database creation are named restrictions in their own right, sitting next to—not folded inside—the training ban. Running Apple's imagery through a vision model to flag, say, solar panels or driveway conditions, and saving the results, is "creation of a database" whether or not any training happened.

Training isn't the only trigger

This is the distinction that catches teams building AI product features, as opposed to teams building foundation models: reading "no training of any model" and concluding a pure-inference feature is fine because nothing gets trained. That's technically correct as far as it goes—calling a vision-language model to classify or describe an image isn't "training" in the machine-learning sense—but it's the wrong test. None of the restrictions above require training to apply. Extracting structured information, creating a database of results, indexing locations, or digitizing features from imagery are each independently restricted, with or without a training step anywhere in the pipeline.

That gap matters more as the tooling shifts. Training a custom model used to be the deliberate, expensive part of a GeoAI product—the part a team would naturally stop and think about licensing for. Calling a vision model to extract information from an image is now a few lines of code, wired into a feature without anyone treating it as a separate legal question. The product may never touch a training pipeline at all; the imagery still gets processed, structured, and stored as new data in exactly the way these licenses restrict.

The same restriction, one level up

Consumer mapping isn't the edge case here—it's the entry point. Go up a level to the satellite and aerial imagery that actually feeds most GeoAI products, and the same split between viewing and processing shows up again, sometimes even more explicitly. Vantor (the imagery business formerly branded Maxar Intelligence, renamed in its October 2025 rebrand) licenses its high-resolution satellite and aerial imagery under a SecureWatch agreement that separately bans using the imagery "to improve the accuracy of any other satellite imagery via algorithmic processing or any other method," and using "Information Product or Derivatives thereof to train machine learning algorithms for feature extraction or for any purpose." Feature extraction—detecting buildings, roads, vehicles, storm damage—is named directly, not implied. Vantor's own definition of "Information Products" includes "extracted data layers, shapefiles, vectors, summary, analysis or other report"—the license anticipates and restricts exactly the inference-output products teams are building today, whether or not the word "training" ever applies.

Planet Labs draws a narrower line: its Partner Program terms bar using "any aspect of the Products or Derivative Products as training or validation of any machine-learning model designed to replicate the Products or Derivative Products." That's a training-specific restriction, scoped to models built to reproduce Planet's own imagery—a genuinely lighter-touch case than Google, Apple, or Vantor. Which kind of restriction a given provider actually uses is worth confirming before you build, not after.

This is not a mapping-industry quirk

The same split shows up everywhere a platform sits on valuable imagery or content. Getty Images prohibits unauthorized scraping and training on its library entirely, and built its own generative model trained only on licensed content, marketed as commercially safe to use. Shutterstock took the opposite route and signed a six-year data-licensing deal with OpenAI, with Google, Meta, Amazon, and Apple following suit. Same content category, two different answers to who's allowed to train on it—and the answer only lives in the contract, not the product page.

Reddit v. Anthropic is the sharper version—a training case, since that's where the lawsuits have landed so far, but the same logic applies. Reddit alleges Anthropic scraped its platform over 100,000 times since mid-2024 to train Claude without a licensing agreement, after OpenAI and Google had each separately signed paid licensing deals with Reddit for the same access. The case is pending in the Northern District of California, hearing scheduled March 24, 2026. The lesson isn't "don't use Reddit data." It's that a licensed path existed, other companies took it, and the one that didn't is now in federal court. Geospatial vendors are sorting into the same two camps—license-for-use or prohibit-entirely—for training and extraction alike.

Why the restriction exists

Street-level, aerial, and satellite imagery are all expensive to create and keep current, which is exactly why the platforms that own them defend that layer so hard, against both training and extraction. Google Street View's cumulative coverage runs to roughly 12 million miles across 110 countries, captured and re-shot by a fleet of contracted drivers running roof-mounted camera rigs. Nearmap's aerial program covers more than 1,740 urban areas and over 80% of the U.S. population, flying over 308,000 unique square miles a year, with recapture up to three times annually. One level up, Vantor's WorldView Legion constellation—six satellites, fully on orbit as of this year—is built to collect more than 6 million square kilometers of Earth imagery a day.

None of these companies publish what that costs to sustain, but the dollar figure isn't the point: running vehicle fleets, aircraft, or a satellite constellation on a repeating capture schedule across that much geography isn't something a startup can casually replicate. The imagery is the hard part of the geospatial stack, not the API layer or the AI model on top of it. That's why access stays open and reuse stays closed—the display layer is replaceable; the capture program isn't.

"We're not training anything" is not a defense

The reason teams walk into this restriction unknowingly is that nothing about the technical or commercial experience signals it, whether it's a self-serve API or a six-figure satellite imagery contract. The documentation is clean, quotas are published (MapKit JS and the Apple Maps Server API share a free daily allowance of 250,000 map views and 25,000 service calls, with a "contact us" path for more), and the sales process looks like any other sanctioned commercial pipeline. The restriction lives in a separate legal document most engineering teams never open, or that procurement signs without a technical review of what it actually bans.

That gap matters because "we accessed it through the official API," "we're a paying customer," and "we're not training anything, just calling a model" all answer the wrong question. Every one of these companies' terms restricts what you do with the output—extract it, index it, structure it, store it—not access method, payment status, or whether a training run happened anywhere in the pipeline.

What's actually at stake

Building an AI feature on restricted imagery without checking the license first creates three separate categories of risk, and they compound.

The first is legal exposure: breach of contract, IP claims, and indemnification obligations already written into the terms you skipped. The second is product risk—a feature built on unlicensed extraction may need to be pulled, rebuilt on different imagery, or renegotiated after launch, a far more expensive fix than a licensing review would have been. The third hits later and harder: data provenance is now a standard diligence item for investors and acquirers evaluating AI companies, and "where does the imagery behind this feature come from" isn't a question most teams want to be answering for the first time during a term sheet negotiation.

Clearview AI is the clearest evidence this isn't theoretical. The company built a facial-recognition model on roughly 60 billion photos scraped from Facebook, YouTube, Venmo, LinkedIn, and Twitter without consent, and has since been fined $33.7 million by the Dutch Data Protection Authority, $9.5 million by the UK's Information Commissioner's Office, and agreed to a $51.75 million U.S. class-action settlement. The legal theory was consent and biometric-privacy law, not a platform ToS breach specifically—but the pattern is identical: build first, discover the sourcing problem later, pay for it at a scale that dwarfs whatever a licensing review would have cost.

Check the license before you ship the feature

None of this means GeoAI teams are stuck. Plenty of imagery and geospatial data providers do offer licensing compatible with AI features and training alike—the Shutterstock/OpenAI deal and the paid Reddit licensing arrangements both show that market exists and works. The mistake isn't building an AI-powered product on street-level, aerial, or satellite imagery. It's sourcing that imagery from whichever API or vendor is easiest to integrate instead of the one actually licensed for what the feature does with it.