Fei-Fei Li, known as the "Godmother of AI" for creating ImageNet, is now leading World Labs to build spatial intelligence systems that understand how the physical world works. These world models represent a fundamental shift in AI development, moving beyond text-based language models to systems that can reason about objects, space, and movement in the real world. However, this next frontier requires training data at a scale that raises significant questions about how companies collect and use information from everyday devices like phones, doorbells, and car cameras .

What Are World Models and Why Do They Matter?

World models are AI systems designed to build persistent, updateable representations of physical space and how objects behave within it. Unlike large language models (LLMs), which process and generate text, world models need to understand the three-dimensional world, how things move, and how physical systems interact. This capability is considered essential for machines that can truly reason about the physical environment rather than just process information .

Fei-Fei Li's work on ImageNet two decades ago quietly became the data foundation for the computer vision revolution. ImageNet was a large visual database containing 14 million images across more than 20,000 categories, which enabled rapid advances in computer vision and object recognition . Now, through World Labs, she is building the infrastructure for spatial intelligence, the next major leap in AI capability.

Yann LeCun, Meta's former Chief AI Scientist who founded AMI Labs earlier this year, has been one of the most consistent voices arguing that world models, not larger language models, are the path to machines that can actually reason about the physical world . Both Li and LeCun represent the cutting edge of where AI research is heading.

How Are Companies Gathering Data for World Models?

• Crowdsourced Location Data: Niantic collected 30 billion images from Pokémon GO and Ingress players who submitted location scans for in-game rewards. The company later spun out a geospatial AI division called Niantic Spatial, which now offers a commercial visual positioning system achieving centimetre-level accuracy in mapped areas .
• Device Fingerprinting: LinkedIn has been silently scanning user devices for more than 6,000 specific browser extensions without disclosing this practice in its privacy policy. The platform assembles 48 hardware and software characteristics into a device fingerprint attached to every user action, collecting sensitive data about neurodivergent conditions, religious practice, and political interests .
• Ambient Sensors: The cameras in phones, doorbells, cars, and smart glasses represent potential training data sources for world models, though users are often unaware of how this data might be repurposed .

What's the Consent Problem with World Model Training Data?

The core issue is a gap between what users think they are consenting to and what actually happens with their data. When Pokémon GO players submitted location scans, they understood they were contributing to improved in-game maps. They did not anticipate that their images would become the foundation for a commercial geospatial AI platform licensed to robotics firms, construction companies, and industrial inspection services a decade later .

LinkedIn's situation is even more troubling because there is no opt-in at all. The company began scanning for 38 browser extensions in 2017, but by April 2026, that number had grown to 6,222, with no notification to users and no option to opt out. LinkedIn is a Microsoft subsidiary, and its vast dataset of professional identity, employment history, and device-level browsing behavior sits at the center of Microsoft's AI ambitions .

"The next frontier after large language models is what researchers are calling world models: systems that do not just process text but build persistent, updateable representations of physical space, objects, and how they behave," according to reporting on the AI ethics landscape.

Montreal AI Ethics Institute, The AI Ethics Brief #188

The consent frameworks that failed to anticipate geospatial AI platforms will not automatically anticipate what world model companies are building either. As Fei-Fei Li and others develop spatial intelligence infrastructure, the question of how that training data is collected, used, and repurposed becomes increasingly urgent .

Why Should You Care About This Shift in AI Development?

World models represent a genuine leap in AI capability. Systems that understand physical space could enable advances in robotics, autonomous vehicles, industrial automation, and countless other applications. However, the data requirements for training these systems are enormous, and the history of AI development shows that companies often repurpose data in ways users never anticipated .

The distance between "players chose to submit scans" and "players understood they were contributing to a commercial geospatial AI platform" is substantial. As world models become more central to AI development, understanding what data is being collected, how it is being used, and what consent actually means becomes essential for anyone using connected devices .

Fei-Fei Li's work has always been about making AI more accessible and understandable. Her ImageNet project democratized computer vision research. Now, as she builds World Labs, the question is whether the data collection practices that enable spatial intelligence will be as transparent and ethically grounded as her earlier contributions to the field.