Pokémon Go Players Helped Niantic Train AI to Navigate the World

In big tech's endless quest to collect as much data – any data – from their users for AI training as possible, it seems no one is entirely safe from having at least some bits of information about them fed to the algorithms of our future robotic overlords. That includes fans of the 2016 mobile AR sensation Pokémon Go, who, as it turns out, unknowingly helped Niantic teach its artificial intelligence to navigate and understand our physical world.

In its recent blog post, Niantic outlined the concept of a Large Geospatial Model (LGM), a machine learning-powered system capable of understanding a scene and connecting it to millions of other scenes worldwide. According to the developer, LGMs are trained similarly to the LLMs, using massive amounts of raw data – in this case, billions of geolocated images – thus enabling AI to "perceive, comprehend, and navigate the physical world."

To realize their vision for an LGM and as part of Niantic's Visual Positioning System (VPS), they have already trained over 50 million neural networks, with more than 150 trillion parameters, allowing operation in over a million locations. "In our vision for a Large Geospatial Model, each of these local networks would contribute to a global large model, implementing a shared understanding of geographic locations, and comprehending places yet to be fully scanned," the company wrote in the blog post.

The caveat lies in how exactly they obtained the data necessary to train those "50 million neural networks," and, to Niantic's credit, they gave the answer in the very same write-up. The company's VPS is built using scans submitted by users capturing locations through Niantic's games and the Scaniverse app. These scans are often taken from various perspectives and include positioning data with each capture, making the process of training AI to understand the world much easier for the developer.

What's more, the data obtained in such a way is relatively unique considering it is captured from a pedestrian perspective, making it different from similar datasets about real-world locations, like, for example, everyone's familiar Google Maps. As it stands, Niantic has scans for 10 million locations across the entire world, with over 1 million of them activated and available for use with the VPS service. "We receive about 1 million fresh scans each week, each containing hundreds of discrete images," the company boastfully noted.

As highlighted by Garbage Day, the first to draw attention to Niantic's blog post, the fact that regular Pokémon Go players unwittingly helped AI models map out a huge chunk of the world is not a problem in and of itself, the real issue lies in how Niantic chooses to use this data. While it could be utilized to build a human-oriented alternative to the aforementioned Google Maps or to create fun gameplay mechanics for AR games, there's also the possibility it could be sold to shadowy third parties and repurposed for nastier applications, such as surveillance, spying, or even automated weapon systems (AI-powered drones, not literal plasma rifle-wielding Terminators). As always, we likely won't know the full implications until it's too late.

While the news might come as a bummer to anti-AI Pokémon Go players, it's not exactly surprising given the game's history. Since its launch in 2016, the game has faced scrutiny over its use of sensitive user data, with many believing that Niantic is up to no good considering who the company's CEO is and his connection to the "Wi-Spy" scandal from the early 2010s.

And in 2019, it was discovered that businesses were paying Niantic to influence Pokémon Go's PokéStops and place them conveniently close to their establishments – such as bars, restaurants, service stations, etc. – something that Harvard Business School's Shoshana Zuboff described as an example of so-called 'surveillance capitalism'. Using Pokémon Go data for AI training seems like just another chapter in this practice and, in hindsight, should have been anticipated as early as 2022 – a year many consider the start of the ongoing AI boom.

Don't forget to join our 80 Level Talent platform and our new Discord server, follow us on Instagram, Twitter, LinkedIn, Telegram, TikTok, and Threads, where we share breakdowns, the latest news, awesome artworks, and more.