An outsourcing firm cut dozens of contractors who were training the language model behind OpenAI's ChatGPT

• Invisible Technologies laid off 31 contractors hired to train OpenAI's GPT, Insider has learned.
• The contractors were hired to improve the AI model's abilities like creative writing and coding.

Dozens of contractors who helped train the OpenAI language models that power ChatGPT were laid off in March, according to a person familiar with the matter and documentation of internal communications.

San Francisco-based firm Invisible Technologies laid off 31 contractors as of March 16, according to internal Slack screenshots that Insider obtained. The layoffs come as OpenAI's ChatGPT takes the world by storm, with users flocking to the bot in hopes of making their lives easier. OpenAI is still hiring throughout its business.

Hundreds of Invisible contractors known as "advanced AI data trainers" work with OpenAI to train its GPT bots, internal Slack screenshots show. Invisible's AI data trainers are responsible for tasks like improving the models' coding skills, enhancing their creative writing capabilities, or training them to stop saying certain things, said an Invisible contractor familiar with the matter, who requested to remain anonymous because he signed a non-disclosure agreement. Insider verified his identity and employment.

Kamron Palizban, vice president of operations at Invisible, addressed the layoffs during an all-staff meeting in March. He said OpenAI wanted to reduce its ranks of contractors due to changing business needs, according to a recording of the meeting Insider obtained. Many of the laid-off contractors worked on projects that didn't provide a high enough return on investment for OpenAI, Palziban said in the meeting.

OpenAI and Invisible Technologies did not respond to requests for comment.

OpenAI slashed some of its contractors after hiring 1,000 globally

Invisible's relationship with OpenAI provides a glimpse into the ChatGPT-maker's data-training practices, which it has largely kept secret from the public.

The adjustment in OpenAI's contract with Invisible follows a six-month staffing ramp-up first reported by Semafor. As of January, OpenAI had hired close to 1,000 data-labeling contractors in places like Eastern Europe and Latin America, sources with knowledge of the matter told Semafor.

Invisible's layoffs came just two months after Microsoft poured $10 billion into OpenAI. But Invisible isn't the only contracting firm that has worked with OpenAI.

In February 2022, contracting firm Sama — also based in San Francisco — ended its partnership with OpenAI after learning its data labelers in Kenya were reviewing harmful content like sexual abuse, hate speech, and violence, according to a Time investigation.

In a statement to Time, an OpenAI spokesperson said, "Classifying and filtering harmful [text and images] is a necessary step in minimizing the amount of violent and sexual content included in training data and creating tools that can detect harmful content."

A day in the life of an AI trainer

According to the Invisible contractor, data trainers' most basic duties include reviewing conversations between AI and its users to identify messages that are potentially illegal, private, offensive, or riddled with errors. The contractor who spoke to Insider explained a daily routine like this:

They start their shift by opening an internal work browser and checking their teams' task lists. They might click on a task like: "Have a conversation about a random topic with browsing disabled," then enter a query into a message box.

Once the query is submitted, the model generates four responses. Contractors evaluate each response by opening a drop-down menu and selecting the types of errors present, such as factual inaccuracies, spelling, grammar, or harassment. Then, they rank the severity of the errors on a scale of one to seven — with seven indicating a "basically perfect" answer, according to a demo the contractor gave to Insider.

Next, contractors must craft what a perfect response might be, and submit it to complete the task. The result is sent to OpenAI and quality checkers at Invisible, the contractor said. The cycle repeats for each following task.

"They're in a stage where they're on the cusp of getting a lot more clarity on where they're going," Palizban said in reference to OpenAI during the meeting.

Invisible laid off contractors based on performance metrics like "quality" and "throughput," Grace Matelich, a partner and operations manager at Invisible, said during the recorded meeting.

Some contractors who underperformed — as well as those who were being onboarded but didn't "hit their bar for certification" — were laid off, though many were given the option to move to a different OpenAI team, per the meeting. "If you're still here today, I want you to know it's because we have faith and trust in your ability to operate with excellence," Matelich said.

If you work at OpenAI or have a story to share, contact this author at amok@insider.com.