Unknowingly, Pokémon Go players are feeding data into a geographic AI model

The developer of Pokémon GoNiantic, is hard at work building and training an AI that can automatically fill real-world locations with a limited amount of information. To do this, it uses data collected by gamers. Pokémon Gonothing less.

According to several sites, including IGN, in an official blog spotted by Garbage Day and reported by 404 Media, Niantic revealed that it is building what is called a “large geospatial or geographic model (GMG).” It is the equivalent of a large language model, like ChatGPT, but applied to the physical environment around us.

A GMG is trained on the appearance of real-world locations (buildings, parks, houses, etc.). But despite the billions of photos available, there are still some missing for rarely photographed sections. This is where GMG comes in to artificially generate images of the corners of the world that have not yet been digitized.

Niantic suggests the example of a church whose facade is widely photographed, but whose rear section the GMG model has not yet seen. So, based on thousands of other photos of churches around the world, a GMG can use the common characteristics of churches to create the missing parts.

Additionally, Niantic claims that this will be useful for technologies like augmented reality glasses, robotics, content creation, and other things.

Tons of data… contributed by players of Pokémon Go

But for this to work, Niantic needs a lot of data to train the AI, and it can’t do it all on its own. The giant Google has been collecting location data for Maps for years from its camera-packed cars that it uses to scan public roads, but that’s not enough in this case. Cars can only drive on roads.

Niantic needs geographic information elsewhere than on the roads. Fortunately, tech has thousands of people around the world pointing their phones at objects and sending that information back through its various projects and apps, including Pokémon Go.

Virtual positioning system… centimeter!

More specifically, Niantic indicated in its post that it had developed a visual positioning system (VPS), a technology that uses the image of a phone to determine the position and orientation of a place on a 3D map. This technology is supposed to allow users to position themselves in the world with “centimeter precision,” allowing them to see digital content superimposed on the physical world “precisely and realistically.”

But all this technology exists because users are constantly scanning the world with their phones when using Niantic’s apps, including Pokémon Goand has been doing so for years now. Niantic said it currently has 10 million digitized locations around the world, a million usable with its VPS service, and receives a million new geographic files every week. That’s a lot of terabytes.

For now, Niantic said it uses this data explicitly to develop its own technologies, which it then integrates into its existing products. However, in recent years there has been widespread concern about how companies collect data to train their services’ artificial intelligence.

If today Niantic’s work in GMG is limited to allowing us to release cute Pokémon models into the world for other people to find, tomorrow their uses could become more and more complex.