Niantic's Pokémon Go scans linked to military drone AI

Somewhere in the terms of service you accepted to catch a Charizard sits a sentence that, according to reporting rounded up by Ars Technica, now leads to a camera-based navigation model heading into military drones. The two ends of the chain are not really disputed. Starting in 2021, Niantic ran an in-game feature that rewarded players for scanning real-world buildings and streets. Niantic Spatial then partnered with a U.S. defense contractor. What is contested is the middle: whether those specific Pokémon Go scans trained that specific model.

The scale is worth pausing on. Niantic's Visual Positioning System is reportedly built on over 30 billion captured images and turns aerial and ground captures into a precise positioning sensor. In December 2025 it linked up with Vantor, the defense and intelligence firm formerly known as Maxar Intelligence, to fuse ground-level scans with Vantor's aerial software for GPS-denied environments where satellite signals are jammed or spoofed. The Decoder cites test figures of up to 70 percent error reduction and roughly 1.5 meter accuracy. In February 2026, Vantor landed a U.S. Army contract worth up to $217 million for 3D terrain work supporting military training environments.

The bigger shift is that consumer AR has quietly become defense-relevant infrastructure. Jamming and spoofing GPS is now standard practice in active conflict zones, and a foundation model trained on billions of hand-held, street-level captures is exactly what a drone that cannot trust its satellites needs. That is why Vantor is buying rather than building from scratch, and it is why any company sitting on a large corpus of user-collected 3D scans is now holding something with a very different price tag than it was two years ago.

The honest caveat is that Vantor reportedly would neither confirm nor deny whether the AI model was trained on Pokémon Go scans, and Niantic Spatial says that following the game's move to Scopely, Pokémon Go data is no longer shared with it, a denial that carefully sidesteps what was already ingested. What the reporting does not resolve is whether the original consent language covers military-adjacent derivative use, or what European regulators will make of it once the Dutch newspaper Trouw's original story circulates further. That is the part worth watching, because the answer will set precedent for every consumer-scan corpus that already exists.