Field Notes

source

You Didn’t Play Pokemon GO. It Played YOU.

StatusStableUpdatedStands on2 sourcesSupportsnotes

A 15-minute narrated video uploaded 2026-06-12 by The Infographics Show, arguing that Pokemon Go was a data-collection operation disguised as a game and that Niantic is positioned to become the landlord of augmented reality. The timestamped transcript carries [seconds] marks from video start. The video is narrated by Josh Risser, lists its sources only as a Pastebin link, and carries no sponsor read.

It is the weaker of the two Pokemon Go videos imported on 2026-07-28, and it is kept mainly to document how the claim circulates. Its analytical frame is worth something; most of its quantities are not.

The one defect that governs the rest

The video was published on 2026-06-12, more than a year after Scopely closed its purchase of Niantic’s games business on 2025-05-29, and it never mentions the sale. Its entire closing chapter at [784]-[907] describes Niantic weighing whether to keep Pokemon Go running now that it has the data it needs, and warns that players may wake up to find their accounts erased. Niantic did not own Pokemon Go when this was published. Scopely did, and had for twelve months.

That is not a detail. The sale is the event that separates the game from the mapping company, and it is the fact that makes the custody question interesting at all. A video that misses it cannot describe the arrangement it is warning about. See Niantic Spatial § The split and the defense turn.

The video also never reaches the defense angle. It was published one week after the Trouw investigation that made this story current, mentions neither Vantor nor any military application, and speculates instead about advertisers and police locating fugitives at [829]-[841].

Quantities that do not check out

Claim Timestamp Assessment
“30 billion images” of the world [70] A frame count across several Niantic games, not distinct captures
$6–10 million to retrain the model [133] No visible source
Surveying these spaces would have cost over $15 billion [196] No visible source; the framing figure of the chapter
“Net profit of over $6 billion as of 2026” [208] Conflates net profit with gross revenue; lifetime revenue is estimated near $8 billion
Players worked “for the wage of roughly $1 an hour” [404] An unexplained calculation presented as a finding
34.2 million active users, “just over 15% of its peak” [787] Unsourced; Scopely reported over 20 million weekly active players and 100 million unique players in 2024

The “50 million neural networks” figure at [70] is accurate and comes from Niantic’s own announcement. It sits in the same sentence as the mistaken image count, which is a fair summary of the video’s reliability: the sourced numbers are right and the load-bearing ones are not.

Where the analysis outruns the mechanism

At [515]-[526] the video claims that “for the first time, a company is trying to patent reality itself,” then identifies the patent as “Generating 3D Models of Real-World Objects.” That is an ordinary patent on a reconstruction technique. Nothing in it forecloses anyone else from mapping the world.

At [574]-[604] it claims Niantic “has access to their location and can find the person in real space,” tracking users “to the exact centimeter,” and concludes that the company “knows exactly where you are right now. Down to the step.” This misdescribes what a visual positioning system is. The system localizes a device whose owner is running the software and pointing a camera at the world. It is a reference map, not a live feed of people, and holding it confers no ability to find someone who is not using it.

At [291] it presents In-Q-Tel as the funding that started Keyhole. In-Q-Tel invested in February 2003, after money from Sony’s venture group and NVIDIA, into a company that was running out of cash.

What it gets right

The reverse-labour frame is the video’s real contribution and it is fairly argued. At [389]-[490] it observes that players performed time-consuming physical scanning tasks paid in a currency redeemable only inside the company’s own store, and compares this to company scrip. The specific hourly figure is invented, but the structural observation — that the compensation loop returned to the issuer, and that the resulting asset was worth far more than the scrip — is sound and is not made as clearly elsewhere.

At [159]-[176] its account of “the last meter” is correct and identifies the genuinely distinguishing property of the resulting dataset: Street View cars reach public roads, while pedestrians reach parks, trails, campuses, and interiors.

At [110]-[143] it makes the point that a model cannot selectively unlearn one contributor’s data without retraining, which is the correct mechanism even though the cost estimate attached to it is unsourced. See Training data provenance.

At [689]-[745] its platform argument — that the advantage lies in holding the largest reference map, so competitors end up licensing rather than rebuilding — matches the competitive logic of the category.

Built on 2 sources (1 archived here, 1 external).

Working out connections…