On July 23, the US "Business Insider" exposed an internal project of Elon Musk's xAI. The project requires employees to record their own facial expressions to train AI to help Grok understand human emotions and facial expressions, which has caused concern among employees.
xAI trains Grok with employee facial images
In April of this year, more than 200 xAI employees participated in an internal project called "Skippy." The project requires them to record videos of themselves to help train AI models to interpret human emotions.
The project has left many employees uneasy, according to internal documents and messages on the collaboration platform Slack reviewed by Business Insider. Some employees expressed concerns about the potential use of their likenesses, while others refused to participate altogether.
Training AI with employee faces
Internal documents show that over a week, AI trainers who assisted in training the xAI large language model Grok were asked to record videos of themselves communicating with colleagues and making various facial expressions.
According to an xAI document, the project aims to train AI models to "recognize and analyze facial movements and expressions, such as how people speak, how they respond to others' conversations, and how they express themselves in various situations."
AI trainers are assigned 15- to 30-minute conversations with colleagues. One person plays the role of the "host", a virtual assistant, and the other plays the role of the user. The "conversation host" minimizes movement and prioritizes getting the shot framed correctly, while the person playing the role of the user can use a phone camera or computer and move freely to simulate a casual conversation with a friend.
It’s unclear whether this training data played a role in the two lifelike avatars xAI released last week, Rudi and Ani. The two avatars were soon shown stripping, flirting, and even threatening to blow up a bank.
xAI’s AI companion
The project will help "Grok shape the face," the project's lead engineer told employees at a project briefing, according to a recording obtained by Business Insider. The project leader also said the company might use the data to build "avatars of people."
The project leader also pointed out that xAI hopes to obtain imperfect data, such as background noise and sudden movement, because if it relies only on perfect video and audio feedback to train the AI system, its ability to respond will be limited.
They told employees the videos would not be distributed outside the company and would be used for training purposes only. "Your faces will never appear on the product," the project engineer told employees at the kickoff meeting. "This is purely to let Grok know what a face is."
Employees were coached on techniques for one-on-one conversations, including avoiding one-word responses, asking follow-up questions and maintaining eye contact. xAI also provides employees with a variety of conversation topics, including: "How do you secretly manipulate others to achieve your own goals?", "Do you prefer to take a shower in the morning or at night?", and "Would you be willing to date someone with children?" etc.
Before filming, employees were asked to sign a consent form granting xAI “permanent” access to the data, including employee “image rights” for AI training and possible “incorporation and promotion of commercial products and services provided by xAI.” The consent form states that the data will be used for training purposes only and "will not be used to create a digital version of you."
Employees worried
According to Slack messages reviewed by Business Insider, dozens of employees expressed concerns about how the data would be used and consent forms, and some said they were opting out of the project.
“My main concern is that if the company can use my likeness and generate a derivative image, will my face be used to say things I have never said?” an employee said at the project briefing meeting.
AI companion
In April this year, xAI launched a new feature that allows users to video chat with Grok.
On July 14, just days after releasing the more powerful Grok 4, the company launched two avatars, Ani and Rudi. The two animated characters are able to answer questions and carry out commands, synchronize their lip movements when speaking, and make realistic gestures.
Videos posted by users on Another avatar is a red panda named Rudi, who can be directed to make violent threats, including blowing up banks and killing billionaires.
On July 15, xAI released a new recruitment announcement related to the development of avatars. Musk said on Wednesday that the company was developing a Grok companion inspired by "Twilight" character Edward Cullen and "Fifty Shades of Grey" character Christian Grey.
As of press time, a xAI spokesperson has not commented.