![Pocket Auto Catch Plus with Keychain - Upgraded Version of Auto Catch, Auto Spin, Catching Pocket Monsters, Collecting Items, Wristband Bracelet Auto Catcher Accessory [Exclusive Version]](https://m.media-amazon.com/images/I/71BzSq2L-8L._AC_UY218_.jpg)
Pokemon Go Scans Used to Train Military Drones: What We Know
Hundreds of millions of Pokémon Go players spent years filming streets, parks, and building façades to earn in-game rewards. Those roughly 30 billion environmental scans are now owned by Niantic Spatial and reportedly helped train a camera-based navigation model that a U.S. defense contractor is preparing to deploy in military drones and robots, per DroneXL. Whether Pokemon Go scans used to train military drones represent a direct data pipeline or a more indirect contribution is exactly what neither company will confirm.
The documented chain runs in three steps: player scans trained an early version of Niantic's visual positioning model; that model moved into commercial deployment before any defense partnership was announced; Niantic Spatial then partnered with a major defense contractor to integrate the technology into GPS-denied military systems. What the reporting cannot establish is whether the specific model Vantor intends to field carries the imprint of Pokémon Go training data. Both companies have been precise about what they've denied, and careful to leave the harder question untouched.
One more caveat before going further. The investigation that surfaced this story was published last week by Dutch newspaper Trouw. Subsequent reporting by Game Developer, DroneXL, and NL Times all trace back to that single source. Three outlets amplifying one investigation is not three independent confirmations, and the certainty level should be read accordingly.
What players contributed and what it became
Starting in 2021, Pokémon Go gave players the option to film short video sweeps of real-world locations PokéStops, parks, storefronts, building façades in exchange for in-game items, per DroneXL. Scanning was voluntary, and Niantic asked separately for permission to retain the footage. The terms players accepted granted Niantic a transferable, sublicensable license to the imagery, meaning the company could resell it to third parties, per DroneXL.
Pokémon Go and Niantic's earlier AR title Ingress together produced roughly 30 billion environmental scans from hundreds of millions of players globally, per Game Developer. Niantic Spatial used that material to build a Visual Positioning System: navigation by visual recognition rather than satellite signal. The system learns structural patterns across thousands of real-world locations so that a device can infer its position from what its camera sees, per Snopes, which rated the underlying claim true in late 2024. Niantic confirmed to Trouw that Pokémon Go scans specifically trained an "early version" of that model, per NL Times.
The proof that this system was working came three months ago. MIT Technology Review reported that Niantic Spatial had partnered with Coco Robotics, whose delivery robots each fitted with four cameras were already running routes in U.S. cities and Helsinki using Niantic Spatial's visual positioning to supplement GPS. The navigation model had moved from training into real-world operation before any defense partnership was announced. The capability was proven. What remained was where it would go next.
How Pokemon Go area scans reached defense tech
In December 2025, Niantic Spatial announced a partnership with Vantor, a defense and intelligence firm formerly known as Maxar Intelligence, to integrate its visual positioning technology into systems built for GPS-denied military operations, per DroneXL.
The partnership joins two complementary systems. Niantic's ground-level visual positioning, trained on real-world environmental scans, combines with Vantor's Raptor aerial navigation software, which uses a drone's camera against proprietary 3D terrain data. The result, as the companies describe it, is that an airborne drone and a ground operator can share coordinates in real time without a satellite link. The joint announcement named GPS "unavailability, spoofing, interference, and jamming" as the specific vulnerabilities the combined system addresses, and listed autonomous drones, ground vehicles, and augmented reality glasses among its intended platforms, per DroneXL. This is positioning infrastructure for contested environments broadly, not a narrow single-platform application.
Vantor is not a startup feeling out the defense market. Rebranded from Maxar Intelligence in October 2025, it holds a $70 million follow-on award from the National Geospatial-Intelligence Agency under the agency's Global Enhanced GEOINT Delivery program, which serves more than 400,000 U.S. government users, per DroneXL. Field testing of the integrated system was scheduled for early 2026; the available reporting does not establish what happened after that window passed.
Both companies' public statements are accurate as far as they go, and both stop short of the question that matters. Vantor told Trouw it would not use Pokémon Go data directly, then declined to say whether the model it plans to deploy was trained on those scans at an earlier stage, per Game Developer. Niantic confirmed that player personal data has not been sold to anyone and that scanning was always optional, but when asked specifically about the Vantor defense partnership, said it had "no new information to share at this time," per NL Times. Both statements are technically compatible with the game data having shaped the model Vantor intends to deploy.
There's a specific reason this gap may never close, and it's not simply corporate evasion. Jeroen van den Hoven, professor of ethics and technology at TU Delft, told Trouw that AI models trained across multiple datasets no longer contain traceable original data they contain learned patterns encoded from it, per NL Times. Deleting the raw scan files doesn't remove what the model absorbed from them. A company can truthfully say it is not using Pokémon Go footage while deploying a system whose capabilities were partly shaped by it. Van den Hoven put it plainly: "Without the huge number of scans from all those gamers, the development of this system would never have progressed so quickly."
The consent gap: agreed to the terms, not the destination
The legal architecture held. Players opted into a voluntary feature under terms that permitted exactly this kind of downstream use. But legal compliance and meaningful informed consent are not the same thing, and the distance between them is what this case keeps exposing.
Floris de Hingh, a Dutch player who started when the game launched in 2016, told Trouw he recorded footage without imagining it might have military implications. "First you think you are playing a game, and then suddenly your data can be used in a war," he said, per NL Times. His reaction isn't a claim that he was deceived. It's a description of a consent process that was formally complete and practically uninformative about where the data could end up.
Van den Hoven frames the problem as bigger than one game's terms of service. "The people who thought they were playing a game have clearly been fooled," he told Trouw, arguing that companies now routinely treat trained AI models as monetizable assets to be sold to whatever buyer appears, regardless of what users understood when they contributed data, per NL Times. "It is gradually becoming clear that companies do not necessarily use our data to genuinely improve our lives. The point is to make money. If they can sell a dataset or AI model for a good price, they will do so." His criticism isn't that disclosure was false. It's that "we may use this data to improve our services" and "a defense contractor may deploy a navigation system trained on your scans" are not equivalent acts of transparency, even when both fit under the same contractual language.
There's also a practical implication worth stating plainly. Stopping scans or deleting an account does not undo contributions already incorporated into a trained model. The window for individual choice closes at the point of training, not deployment. Once visual patterns are encoded into the model's learned behavior, the source material is no longer relevant to the model's capabilities. This isn't a quirk of Pokémon Go; it applies to any consumer application using opt-in data to train AI at scale.
The indoor dimension extends the concern beyond publicly visible spaces. A senior Niantic Spatial product executive suggested in an interview published on Vantor's website that the company would like to collect more indoor imagery, per NL Times. De Hingh had already filmed inside his apartment without thinking twice about it. Consumer AR games are structurally well suited to generate machine-readable models of private spaces at scale, and nothing in the current disclosure framework clearly governs how that data flows or what markets it can eventually reach.
What this case establishes, and why it matters beyond Pokémon Go
The documented progression is solid. Snopes confirmed in late 2024 that Niantic used player scans to build its Large Geospatial Model. MIT Technology Review confirmed three months ago that the resulting navigation system was already deployed commercially in robotics. Niantic itself acknowledged the scans trained an early version of the model. Then, late last year, Niantic Spatial partnered with a major defense contractor to integrate that technology into GPS-denied military systems.
What no reporting can currently confirm is whether the specific model Vantor intends to field carries the imprint of Pokémon Go training data. Van den Hoven's point about model provenance that tracing original data contributions once training is complete may be technically impossible regardless of intent, per NL Times means a clean resolution may not exist even with full corporate disclosure.
Pokémon Go is the clearest large-scale example yet of a structure that will keep recurring. Consumer AR applications function as distributed scan infrastructure. Players generate spatial data as a byproduct of gameplay; that data trains AI systems; those systems get adapted into markets players never considered commercial robotics, defense navigation, indoor mapping. Van den Hoven has called on the European Commission to establish clear rules governing downstream uses of consumer-contributed data, per NL Times.
The policy question this case forces into the open is whether consent to AI training and consent to downstream market transfer should require separate, specific disclosures, or whether one broad clause is sufficient to cover both. Right now, one clause covers everything. The Trouw investigation has made that arrangement very hard to keep ignoring.
Comments
Be the first, drop a comment!