Sloshed in the subway after a night of partying? Bored during a virtual meeting? Dozing off at a red light? Companies are building software that uses AI to monitor people’s behavior and interpret their emotions and body language in real life, virtually and even in the metaverse. But to develop that AI, they need fake data, and startups are stepping in to supply it.
Synthetic data companies are providing millions of images, videos and sometimes audio data samples that have been generated for the sole purpose of training or improving AI models that could become part of our everyday lives in controversial forms of AI such as facial recognition, emotion AI and other algorithmic systems used to keep track of people’s behavior.
While in the past companies building computer vision-based AI often relied on publicly available datasets, now AI developers are looking to customized synthetic data to “address more and more domain-specific problems that have zero data you can actually access,” said Ofir Zuk, co-founder and CEO of synthetic data company Datagen.
Synthetic data companies including Datagen, Mindtech and Synthesis AI represent a corner of an increasingly compartmentalized AI industry. They produce AI parts that will eventually be assembled to build software, features in applications or systems used in vehicles. They serve customers such as computer vision engineers and data scientists working for big tech giants, automakers, gaming companies or mobile phone makers.
Like so much polyester, synthesized datasets are intended to mimic the real thing. Synthetic data does not just replicate actual photo and video data; it enhances it by adding dimensions and details that help AI-based systems learn. Sometimes the synthetic stuff fills serious data gaps where real data does not exist or is difficult to obtain. It might depict dangerous highway situations used to train autonomous vehicle AI, or include facial images representing people of multiple ethnicities or ages needed to help ensure AI makes fair and accurate decisions.
Many of these companies tout synthetic data as a panacea for the lack of diverse AI training datasets that has contributed to discriminatory AI, particularly facial recognition. “We help customers reduce AI bias by providing synthetic data spanning a wide range of age, gender, BMI and ethnicity,” said Yashar Behzadi, CEO of Synthesis AI.
Training AI to spot drunks and cheats
“One company came to us recently needing a solution to detect cheating in exams,” said Steve Harris, CEO of Mindtech, a company that offers a platform for designing and rendering images based on photorealistic computer graphics. Harris would not name the virtual testing technology customer, but said that like many other virtual testing tech companies, the customer wanted to incorporate AI to monitor whether test takers are showing signs of cheating such as looking away from the computer screen or interacting with a person or a phone.
For an AI model to pick up on all the different possible signs of cheating in multiple environments involving a variety of people, it would need a large corpus of imagery showing hand, eye and body movements to learn from — the sort of images that could be too expensive to purchase, or force violations of privacy to obtain, even if there were enough of them. “It becomes even more complex when you throw in facial key point data and skeleton pose data to train systems to understand which way the student’s gaze is going, which way their body is about to turn or which direction their hands are facing,” Harris said.
The Center for Democracy and Technology reported that AI-based systems used to detect cheating on virtual testing pose risks to student privacy and mental health, and can discriminate against disabled people.
Harris said customers are also interested in using synthetic data to build AI for use in public places like transportation hubs. Some have sought data to train AI systems to prevent an impaired person from getting hurt in a train station, for instance, “if someone is drunk and moving too close to the line,” he said.
Mindtech recently unveiled a package of pre-built images designed specifically for retail environments that could be used to track customers’ interest in specific products, keep track of stock on store shelves or predict traffic flow in parking lots. In the earlier days of the pandemic, Mindtech customers wanted synthetic data to train AI to monitor for compliance with face mask rules, Harris said.
As companies continue to develop AI for a never-ending array of applications, investors see a bright future for the synthetic data makers supplying the raw materials. Mindtech collected $3.7 million in funding in March. Synthesis AI received $17 million in series A financing in April, and Datagen gathered $50 million in a series B funding round in late March.
Facial expression data detects meeting boredom and driver distraction
For now, a lot of the AI that synthetic data helps build is used in mundane, real-world situations such as driving. Synthetic data is training AI models for driver monitoring systems that capture driver images through dashboard cameras and use computer vision AI to detect distraction, such as in delivery vehicles.
“We have a number of customers in that area,” Behzadi said. For example, Synthesis AI provides synthetic facial data to Affectiva, a company that offers AI-based systems to estimate people’s emotions and cognitive states in real time to detect driver distraction and behavior such as road rage, according to Rana el Kaliouby, founder of Affectiva.
“The data we provide Affectiva and others in driver monitoring is focused on improving driver safety — actions like falling asleep, not wearing a seatbelt or being distracted. We believe synthetic data can be a net positive in reducing fatalities and improving overall safety,” Behzadi said.
AI-based systems for assessing driver behavior are bound to be more commonplace in coming years. The European Commission has made distraction and drowsiness recognition features mandatory for new vehicles this year. Meanwhile in the U.S., the 2022 Infrastructure Act set aside funds to study the use of driver-monitoring systems to minimize driver distraction.
Synthetic data is also fueling AI for workplace monitoring. Datagen’s latest product lets customers build annotated, 30-frame-per-second images for use in office, meeting and conference technologies. Zuk said the new data might be used to train AI that detects whether someone is bored during a work meeting; it includes image data showing people holding their heads in their hands, for instance.
Already, AI-based features for assessing people’s emotional states are showing up in virtual classroom platforms and even sales meeting software, although synthetic data has not been used to train many of these systems yet.
Human rights advocates are fighting potential use of emotion AI in everyday tech. More than 25 organizations including the American Civil Liberties Union, Electronic Privacy Information Center and Fight for the Future sent a letter on Wednesday to Eric Yuan, founder and CEO of Zoom, demanding the company end plans to incorporate emotion AI in its software features. The letter and previous efforts by Fight for the Future were prompted by reporting in Protocol in April about Zoom’s potential plan to incorporate emotion AI into its products.
“This software is discriminatory, manipulative, potentially dangerous and based on assumptions that all people use the same facial expressions, voice patterns, and body language,” the groups wrote.
Mindtech’s Harris said companies building the next wave of office and meeting software are interested in monitoring human interactions to “flag things that look unusual.” He said he expects companies like Meta and Google to incorporate this type of tracking AI in the virtual environments they create, but he added, “It’s some way off.”
For AI to pick up on whether people are paying attention to the road — or to the boss during a meeting — it often needs to recognize facial expressions. Synthesis AI’s datasets include minute distinctions among millions of images expressing as many as 150 facial “micromovements,” Behzadi said. Customers use the company’s digital system to submit requests for custom data, then it automatically renders what they ordered.
“They’ll say, ‘I need a million images that span all these different dimensions,’” Behzadi said. The result might be thousands of facial images with a variety of skin tones, hair styles or features like hats or glasses.
Diversity also manifests in the way images are lighted, Behzadi said. If Synthesis AI is making data for use in an augmented reality environment, the system will produce multiple versions of images of people or objects — a mug of coffee on a desk, for example — with variations in the direction that light emanates from. “So when I render this image in the scene it’s realistic,” he said.
A step removed from ethical implications
According to el Kaliouby, Affectiva has used synthetic data to increase the diversity of its dataset representing people across age ranges and ethnicities. Merely having lots of faces across cultures may not be enough to train AI that also needs to learn what people look like when they are wearing a ball cap, when they are alert or asleep or in environments with low or bright light. “It gets really complex and expensive super fast to scale this,” she said.
But as synthetic data companies push a diversity mission, their products may be used to build contentious forms of AI. The legitimacy of emotion AI has been questioned by researchers who say neither humans nor machines can accurately detect people’s emotions based on facial expressions. And in general, many also believe that algorithmic systems monitoring people’s facial expressions or how they walk or talk perpetuate unnecessary surveillance and could be used to unfairly penalize people.
However, in some cases, synthetic data suppliers remain a step removed from the products that will be manufactured using their data.
Because it is automatically produced, synthetic data comes with some baked-in metadata including details about what images and videos represent that are necessary to help AI models learn. In Datagen’s case, the company includes labels showing the intensity of facial expressions such as “slightly smiling” or “extremely happy,” but also lets customers add custom labels.
“Datagen is propelling advancements in AI by removing the need to source and manually annotate training data,” said Gil Elbaz, Datagen’s CTO and co-founder. The company did not say whether it has restrictions on the types of customers it will supply data to, other than to say it is “focused on commercial Computer Vision applications enabling AI teams to develop safe, human-centric use cases.”
But other synthetic data companies leave some of the data-labeling decisions to their customers and stay at arm’s length from the end products they help make. Instead of providing qualitative labels categorizing facial expressions as confused or bored, Synthesis AI only annotates facial images with technical information. An image label might include metadata stating that the left side of the mouth moved upwards 10 degrees, but would not come pre-labeled as “slightly happy,” for instance.
While Behzadi said Synthesis AI has turned down work with customers that wanted to use its data to identify people without their consent, he said the company has not turned down potential customers that want data to train emotion AI models.
Mindtech also leaves the labeling and end-product decisions to customers. “We understand that facial expressions are very subjective, so we allow a customer to determine how they want to use any labels,” said Chris Longstaff, Mindtech’s VP of Product Management. He said customer confidentiality prevents Mindtech from knowing details of products built with its data.
Expect more synthetic data creation in the near future as it forms the foundation of all sorts of AI built for emerging virtual environments. “There is the potential for synthetic data to be a prominent tool for metaverse companies,” said Harris.
Behzadi said he expects interest in metaverse-related uses for synthetic data to ramp up in the next year. Synthesis AI is working with customers that want to digitize sporting events in real time for metaverse environments.
In the future, Behzadi said, “I can watch the game from Tom Brady’s eyes.”
Correction: An earlier version of this story misstated Mindtech's funding round amount. This story was updated on May 13, 2022.