TThe speed at which AI is changing our lives is dizzying. Unlike previous technological revolutions (radio, nuclear fission, the Internet), governments are not leading the charge. We know that AI can be dangerous. Chatbots can advise teenagers about suicide and may soon be able to teach them how to make biological weapons. But there is no equivalent to the Federal Drug Administration, which tests new models for safety before they are released to the public. Unlike the nuclear industry, companies often There is no need to publicize dangerous violations or accidents. The strength of the tech industry lobby, the paralyzing polarization in Washington, and the complexity of such a powerful and rapidly changing technology have kept federal regulation at bay. European officials are facing a backlash against the rules, which some say will hurt the continent’s competitiveness. Several U.S. states have piloted AI laws, but they are operating in a temporary patchwork manner, and President Donald Trump is trying to repeal them.
Leaders of AI platforms such as OpenAI’s ChatGPT and Google’s Gemini say they prioritize safety. But taking control of the future of AI means making choices like spending billions of dollars on models that even their creators don’t fully understand, adding advertising that increases risk, and features that the Department of Defense currently wants from Anthropic. Anthropic, which bills itself as the most conscientious frontier AI company, says its models are trained by “sensible senior Anthropic employees” to weigh help against possible harm. The directive echoes criticism leveled years ago against Silicon Valley companies that have shaped the lives of users around the world from their closed boardrooms. Consumers don’t believe they are in good hands. A whopping 77% of Americans surveyed last year believe that AI could pose a threat to humanity.
We are not caught between the elusive hope of strong government regulation and the most powerful corporations in history doing their own policing. At least until legislators take action, independent oversight offers the possibility of determining the potential of AI and its dangers. By employing independent oversight, AI companies can demonstrate that they take public trust seriously enough to actively fight to protect it.
The logic behind independent supervision is simple. Regardless of corporate management’s good intentions, their obligations to shareholders and investors determine how they approach trade-offs between cost and safety and promote revenue and profit. Long-term considerations for a company’s reputation, customer loyalty, and ethics can act as a speed bump, but winning the AI race requires an appetite for risk. A belated investigation into how social media incites murder, conducts elections, and undermines the mental health of young people shows how the intoxicating power of technology can overshadow flashing warning signals.
Independent oversight of AI offers the potential to surface, analyze, and address AI risks, giving advocates and communities a little more control over how these technologies reshape society. Social media is an example. In 2020, Meta (then Facebook), hurt by accusations that it fueled Myanmar’s Rohingya crisis, set up an oversight committee to try to get the company out of a tough situation. Early the following year, the company adopted a policy pledging to comply with human rights law. The five-year-old board has fallen short of serving as the “Supreme Court of Facebook” that some had hoped, but its record offers important lessons about the prospects for effective independent oversight of AI and why it matters.
Directors are required to have diverse perspectives. Like other frontier AI companies, Meta has users on every populous continent. Deciding what can and cannot be posted from the safety of the Menlo Park campus has left blind spots and fueled outrage. The 21 members of the oversight board bring a wide range of cultural and professional expertise to adjudicating sensitive content moderation issues, such as whether a violent video should be shared as news or removed as an insult to the victim’s dignity. Board members reside in more than 27 countries and include conservatives and liberals, journalists, legal scholars, a former Danish prime minister, and a Nobel Peace Prize winner.
The oversight committee uses Meta’s own “community standards” to assess whether posts violate rules such as prohibitions on bullying and supporting terrorists. The Council upholds Meta’s commitment to uphold international human rights law, including article 19 of the International Covenant on Civil and Political Rights, which enshrines freedom of expression. AI companies need to take similar steps and establish oversight to protect them. Unlike the United States’ First Amendment or the European Union’s online “right to be forgotten,” human rights law provides a common currency across borders. The Code provides methods of reasoning to guide decisions about AI, such as whether a bot’s refusal to answer a question unreasonably denies a user’s right to information and whether reusing user data violates privacy rights.
Accessibility, consultation and transparency are key. The Oversight Board accepts complaints from the public, announces the cases it chooses to review, solicits public comment, and holds meetings with experts and relevant communities. Has made over 200 decisionss Detailed written opinions cited in courts around the world.
An autonomous supervisory authority is only as strong as the powers given to it by the company that founded it. The oversight board wants broader powers, but it allows Meta to go far beyond the lightweight advisory committees that other tech companies regularly convene and disband. Meta’s oversight board has the power to decide whether certain content should survive or cease, but exercising that power over individual posts can feel like fighting a wildfire by blowing out embers. Its more significant impact lies in selecting emblematic cases of erroneous content, providing public reasoning for decisions, and issuing recommendations to which the meta must respond. As reported in December, Meta has implemented 75% of the board’s 300+ recommendations, resulting in significant changes for billions of users.
These include providing notifications about what policies a user has allegedly violated when content disappears, ensuring that rhetorical mockery or satire is not removed as a threat, and surging corporate resources during crises such as natural disasters or armed conflict. The board also issued detailed advisory opinions on larger policy issues, including extending Meta’s leniency for policy violations by prominent posters and the extent to which coronavirus-related misinformation should be removed as the pandemic subsides. Although the board operates independently in making decisions and recommendations, it relies on Meta for important information, such as whether certain content decisions were made by humans or automation, and what exactly went wrong when content was removed by mistake. For surveillance to make sense, AI companies will need to provide at least the same degree of visibility.
As always, money is important. Mehta periodically places the Oversight Committee’s funds in a trust so that it does not run out of money overnight. However, more diverse and reliable resources would increase board independence. Monitoring cutting-edge technology is costly. It requires funding for specialized staff to support analysis and decision-making, as well as consultants who bring specific cultural and linguistic expertise. But given the hundreds of billions of dollars invested in AI, the cost of even strong surveillance is negligible.
AI is taking over our classrooms, universities, and businesses. Independent oversight is the least AI companies can do to ensure they don’t hijack our rights, intentionally or not.
