Bounding Box is Dead: Data Annotators Find New Purpose in the Physical AI Age

In the early days of machine learning, the human labour behind artificial intelligence was largely confined to data labelling, drawing bounding boxes around cars, pedestrians, and street signs so a computer-vision model—especially those for self-driving technology—could learn to recognise them. 

According to Steve Larson, Chief Operating Officer of Digital Divide Data (DDD), one of the industry's largest impact-sourcing employers, that job barely exists anymore.

Founded in 2001, New York City-based DDD is a non-profit social enterprise that provides digital data solutions, AI training data, and IT outsourcing services to large enterprises. Its data labelling work spans geographies including the US, Kenya, Cambodia, and Laos, and serves roughly 100 clients across the AI development lifecycle—from assembling training datasets to those refining mature, widely used consumer AI models.

More than 95% of its workforce comes from economically and socially marginalised backgrounds.

"The industry in general has grown well past that stage where a human needs to draw a bounding box," Larson tells AIM on the sidelines of the launch of the company’s first wholly-owned global capability Centre (GCC) in Dehradun.

Even annotation has been superseded by Visual Language Models (VLMs) and Visual Language Action (VLA) models, which don't merely detect objects in an image but interpret what they mean and how to act on them.

The shift matters beyond one company's internal workflow. It signals where the estimated hundreds of thousands of data annotation jobs are heading next—from teaching machines to identify to physical AI, where companies are teaching models to understand and act for applications ranging from robots, self-driving cars, automated healthcare, and manufacturing equipment.

However, Larson says humans remain central to that work because as models grow more sophisticated, someone still has to ensure the outputs are accurate, safe, consistent, and representative of different cultures and communities.

“We help our clients to make their models even better, and that's where humans are increasingly important,” says Larson.

DDD is now betting a large chunk of its India expansion on that shift. It will focus its Dehradun GCC around robotics and physical AI data work, a combination it says is uncommon among GCCs operating in India. The centre will be essential to collecting and annotating egocentric video, sensor readings, and other multimodal data that trains machines to operate safely in real-world settings such as hospitals, factory floors, and roads.

The centre has already onboarded around 200 employees in the six weeks since it became operational on August 1, and the company expects that number to swell to between 500 and 600 by December, and close to 1,000 within a year. 

To the Himalayan Foothills

The Dehradun facility, according to the company’s senior executives, is expected to become one of Uttarakhand's largest concentrations of specialist AI talent. The centre supports a broad slate of services: multimodal data annotation, data curation, large-language-model fine-tuning support, low-resource language services, AI model evaluation, and data preparation and validation.

DDD's expansion into Dehradun also follows the same impact sourcing model the company has run since its founding—hiring fresh graduates and local candidates for aptitude rather than prior experience, then putting them through structured training before they begin live production work. 

Larson frames that the goal is to let talented young people build AI careers close to home rather than having to relocate to metros such as Bengaluru or Delhi, enabling them to eventually start their own AI companies or move to established Indian IT majors.

Most of the 200 employees hired so far are natives of Uttarakhand, with a small number who have family ties in Nepal.

Digital Divide Data has not yet engaged directly with the Uttarakhand state government but plans to do so, and any government work it undertakes in India would be processed and stored entirely within the country to meet data localisation requirements.