Meta Scattered Its Responsible AI Team Across Other Groups

Meta has quietly broken up its Responsible AI (RAI) team, according to a report from The Information. Most of the team’s members are being shifted into the company’s generative AI product group, while others will work on AI infrastructure. The news broke on November 18, 2023, and Meta confirmed the change in a statement to The Verge.

On paper, this sounds like a simple reorganization. Meta’s communications rep, Nisha Deo, said the move is meant to help develop AI features faster, and that the company will “continue to prioritize and invest in safe and responsible AI development.” She also claimed the RAI members now in the generative AI org will “continue to support relevant cross-Meta efforts on responsible AI development and use.”

But let’s be real here. This is the same team that, earlier this year, was already described as “a shell of a team” after layoffs, according to Business Insider. The RAI team, which had existed since 2019, had little autonomy to begin with. Its initiatives had to go through lengthy stakeholder negotiations before they could actually do anything. Now they’re being absorbed into the very product teams they were supposed to be policing.

Meta’s RAI team was created to catch problems in the company’s AI training approaches—things like whether models are trained with diverse enough data, or whether they’ll cause moderation issues. And we’ve seen plenty of those issues. A Facebook translation error led to a false arrest. WhatsApp’s AI sticker generator spat out biased images. Instagram’s algorithms helped people find child sexual abuse material. The RAI team was supposed to be the guardrail for all that.

Now those guardrails are being moved inside the race to ship generative AI products. That’s not exactly a vote of confidence in responsible development.

Meta isn’t alone here. Microsoft did something similar earlier this year, shuffling its responsible AI folks into other teams. This all comes as governments scramble to regulate AI. The US has agreements with companies and Biden’s executive order on AI safety. The EU is still struggling to pass its AI Act. But while regulators are still figuring things out, the companies that built these systems are quietly dismantling the internal teams meant to keep them in check.

I’m not saying Meta is going to throw caution to the wind. But when you dissolve the team whose job is to say “no” and put them inside the team whose job is to ship, the incentives are pretty clear. The RAI team was already weak. Now it’s just gone.

This article was updated on November 25, 2023 with a statement from Meta.