Meta’s Oversight Board has a warning for AI companies building their own watchdogs

Members of Meta’s Oversight Board, the longest-running experiment of independent oversight in the tech industry, have some advice for AI companies looking to launch their own independent safety committees — give boards real power, or they will not mean much.

The board has spent six years ruling on the company’s content moderation decisions. Members said in interviews that the independent safety committees that AI companies, including Google, Anthropic, OpenAI, Meta, XAI and Nvidia, pledged to create last week could become little more than window dressing unless they’re given a clear mandate, real independence and the ability to push back.

“If it’s going to be genuine oversight, there’s going to be friction between the company and the body,” Paolo Carozza, a member of the Oversight Board and a professor of law and political science at the University of Notre Dame, said about the board’s recommendations for AI company oversight.

“If that friction doesn’t exist, it’s a sure sign that there’s no effective independent oversight going on,” he said.

The Oversight Board, an entity legally independent from Meta that has final say on certain social media issues at the company and offers nonbinding advice on a plethora of other topics, issued an open letter in response to the new voluntary White House Accord on Super Intelligence to share its recommendations for the new AI governance endeavors.

“With a tool that has as much potential and as much risk as AI, we will need a layered approach to governance that includes company policies, industry-adopted standards, independent oversight of those standards from a range of perspectives as well as both regulation and international coordination,” the group wrote in the recent statement.

Members of the group said they hoped that any new committees would meaningfully address concerns about AI, but they also expressed skepticism that increasingly powerful AI companies would give up the control necessary to bring about actual change.

“I think an AI company should be ready to open itself to scrutiny, even when it’s uncomfortable for them, because it is precisely in that discomfort that the company will gain even more trust from the public,” said Julie Owono, an Oversight Board member and executive director of the international digital rights nonprofit Internet Sans Frontières (Internet Without Borders).

The Oversight Board released six recommendations in its recent letter about how independent entities should exert control over AI companies, including through establishing a concrete mandate, ensuring structural stability and independence, involving cross-sector expertise and focusing on transparency. “Transparency is paramount,” Owono said.

The White House Accord on Super Intelligence — which does not bind OpenAI, Anthropic, Google, Meta, Nvidia or XAI to any particular action — clarifies that the envisioned independent committees overseeing AI companies’ safety efforts should be composed of members from the companies’ boards of directors.

Meta’s Oversight Board does not include Meta board members and is instead populated by independent experts, including journalists and human rights legal scholars, as well as a Nobel Peace Prize laureate and a former Danish prime minister.

Nighat Dad, an Oversight Board member and executive director of the Pakistan-focused Digital Rights Foundation, said independent safety committees at AI companies need to have real access to nonpublic information that might be unflattering for the companies, along with financial independence.

“Is the committee separate from the company? Does it control its own budget and its own membership? Who are the members who are making these decisions, and are those decisions binding?” Dad asked.

“If the answer is no, that isn’t oversight. And if the answer is yes, it can genuinely change how companies behave,” she said.

Meta controlled some of the board’s membership selection process until 2023, and it provided the board with funding — over $300 million since 2019, held in an independent trust.

Evelyn Douek, a law professor at Stanford University who has studied the Oversight Board, said efforts to incorporate independent oversight of AI companies have much to learn from the board’s experience. “I think the Oversight Board should be more of a cautionary tale rather than a sort of promising model for what independent oversight should look like in these kinds of situations,” Douek said.

She pointed out that the board was given only limited powers to govern a small subset of Meta’s decisions — specifically on content moderation — instead of having binding say on Meta’s wider policy and safety issues.

Carozza agreed that the board could have been more effective had it been constructed differently — and that such hindsight is what drove it to speak out before AI companies cement their own oversight plans.

“Nobody at the Oversight Board is saying we are the model or that we are the solution,” Carozza said, instead highlighting that the board has six years of experience to draw upon to understand what could be improved.

Ronaldo Lemos, a law professor at Rio de Janeiro State University and a Meta Oversight Board member, said any effective mechanism for ensuring AI companies comply with safety measures requires those involved to have true separation of power from any of a company’s financial or political incentives. OpenAI and Anthropic are on track to join public stock exchanges within the next few months, with estimated valuations of over a trillion dollars each.

“Oversight needs independence. It needs to be transparent, and it needs to have a clear mandate,” he said.

The Oversight Board was established in 2019 by Meta — then called Facebook — to provide external guidance for thorny free speech issues. At the time, Facebook was being slammed for its alleged role in hosting fake news and content that incited violence that was part of the Rohingya genocide in Myanmar, according to the Oversight Board.

“The board will be an advocate for our community — supporting people’s right to free expression, and making sure we fulfill our responsibility to keep people safe,” Facebook co-founder and CEO Mark Zuckerberg wrote at the time, pledging to give the Oversight Board enough information and resources to make its decisions.

Douek of Stanford pointed to the Oversight Board’s unique structure — reliant on Meta for funding — as a potential reason for its limited efficacy. “At the end of the day, it’s very hard for people to bite the hand that feeds them in any meaningful way,” she said. “The Oversight Board’s continuing existence was reliant on Meta continuing to fund it.”

Some experts point to OpenAI as another example of a company whose attempts to oversee safety efforts have fallen short, despite its reliance on an independent oversight committee.

Because of OpenAI’s unique corporate structure, OpenAI’s Safety and Security Committee has the ability to pause or delay product releases over safety concerns, though the state authorities overseeing the company have recently expressed dissatisfaction with the committee’s efficacy. OpenAI has said the committee works diligently to ensure the company’s systems and products are safe.

Dad, the Digital Rights Foundation director, also highlighted that AI companies’ global reach would require concerted global oversight efforts. “A standard based on any one country’s law or written by the company itself will never feel legitimate to users everywhere,” she said, highlighting differences in cultural context even between India and Pakistan. Dad said AI companies’ oversight bodies should turn to international human rights law as a way to ensure global norms are taken into account.

Many of the oversight mechanisms proposed in the White House accord are still unclear. As a result, Douek said, it is too early to tell whether the independent oversight bodies laid out in the accord will carry any weight, though she said the vast sizes of AI companies and their complex technical work could limit oversight groups’ visibility into knotty safety issues — an issue that the Oversight Board members said emerged with their original board.

“How much technical capacity and bandwidth does a body have to really dig into the questions and issues that need oversight?” she asked.

Emphasizing AI companies’ future economic and social power, Carozza of Notre Dame said: “How these mechanisms of oversight are designed now will have a critical impact two, five, 10 years down the road in terms of how effective they are and whether we need to undo things or redo things. It’s important to get it right from the start.”

NBC News Top Stories wp:paragraph