Meta Was Collecting Employee Data To Train Its AI But It Didn't Go As Planned

Meta had to stop a programme that collected employee data to train artificial intelligence (AI) after workers raised privacy concerns and an internal leak exposed the information across the company, according to a Business Insider report.

The programme, called the Model Capability Initiative (MCI), was started by Meta's AI team earlier this year. It was meant to collect data such as employees' keyboard typing and mouse movements.

Meta wanted to use this data to help its AI models understand how people use computers and perform everyday tasks. However, the plan faced strong opposition from employees soon after it was announced. More than 1,800 Meta workers signed a petition asking the company to stop the programme, according to the BBC.

The programme was developed inside Meta's AI unit, known as TBD. The group has around 100 researchers and is considered one of the company's most important AI teams.

One of the people involved in the programme was Rylan Schaeffer. He reported to Meta AI chief Alexandr Wang and senior researcher Jack Rae. Schaeffer left Meta in July, according to his LinkedIn profile.

Meta announced MCI to its employees in April.

The company said AI models were already good at tasks such as coding and research but still struggled with simple computer actions. These included using dropdown menus and keyboard shortcuts. Meta believed employee data could help AI learn these actions.

But employees were unhappy after learning that they could not initially opt out of the programme.

When a cybersecurity engineer raised concerns about how the data would be protected, Bosworth said the information was "tightly controlled".

Employees soon began protesting against MCI.

Workers were worried that the programme could collect sensitive information from their computers. This could include passwords, immigration documents and credit card details, according to organisers who spoke to Business Insider.

Some employees were also concerned that the data could help Meta develop AI systems that could eventually replace human workers. Workers shared memes against the programme and circulated flyers at Meta offices. They also gave the screen-recording icon used by MCI the nickname "Eye of Sauron".

Workers also put up messages protesting MCI at Meta's Burlingame office. In June, Meta allowed employees to temporarily stop the tracking for up to 30 minutes while doing sensitive tasks. However, employees said this was not enough. They wanted Meta to completely scrap the programme.

In late June, a Meta engineer discovered that data collected through MCI could be accessed across the company. The information reportedly included private conversations. This was shared widely on Meta's internal platform.

One employee wrote a post titled "MCI Data: Not So Tightly Controlled". The post also referred to Bosworth's earlier comments about the security of the programme. However, Meta switched off the programme within hours.

The company said MCI had been suspended and was not permanently cancelled. It also said the programme would be optional if it was brought back.

According to Business Insider, a Meta manager said in July that the data had not been used to train the company's latest AI models. However, Bosworth reportedly said during a July town hall that the data had been more useful than Meta expected. 

Meta has not explained how the data was used or whether it was used to train AI models.

Bosworth later criticised the employees who opposed MCI. In a leaked July audio recording published by the activist group More Perfect Union, he said the campaign had made the company worse, according to IB Times. He said the employees who started the petition believed they were helping Meta but had instead made it harder for the company to make changes.

William Fitzgerald, founder of advocacy and communications firm The Worker Agency, described the employee campaign as "a historic win" and said Meta wanted more data but failed to collect it in this way.