The End of Meta’s Fact-Checking Program: What It Teaches Us About AI Alignment
Meta's decision to sunset its fact-checking program is a necessary shift, offering vital lessons for AI alignment and accountability
When Meta decided to end its third-party fact-checking program, it wasn’t just a policy change—it was a window into the challenges of shaping global governance policies.
I helped build these products at Meta and think now is the ideal time to reflect on the values that shape these decisions. And critically, the lessons AI companies should take away about alignment and accountability.
Fact-Checking in “Crisis”: Why Meta’s Shift Matters
Meta’s decision to sunset the third party fact checking program raises questions about the utility of relying on fact checks to address billions of pieces of content across Facebook, Instagram, etc.
My goal is not to critique the overall approach of engaging hundreds of fact checkers to try and shape a global information ecosystem. The program was undoubtedly novel. It introduced mechanisms like Community Review (similar to X’s Community Notes) and we built cutting edge AI tools to support content moderation. The company will continue moderating content, without third party fact checking for misinformation.
However, the challenges Meta faced reveal deeper issues that resonate with broader concerns about AI alignment.
The Hidden Challenges Behind Meta’s Fact-Checking Program
Ideological (In)Coherence
Meta developed the rules its content was judged against. During my tenure, these content rules could lack a clear philosophical foundation. Should they be grounded in democratic values, company priorities, free expression…?
Fact-checking at scale isn’t just a tech challenge—it’s a values challenge.
Without a coherent ideological framework, policies become reactive instead of strategic, leaving users and fact-checkers navigating arbitrary rules. Decisions were left to mid-level employees to figure out, rather than laddering up to a grand plan.
In her Noema piece, The Great Decentralization, Renee DiResta notes that “governance policies were largely crafted by American lawyers…to the spirit of American law.” While this may have been a goal, it wasn’t my experience on the ground. After leaving Meta, I interviewed more than 30 professionals working in trust & safety across the industry and a lack of coherent principles driving content policy was raised repeatedly. Where did these company rules come from?
At a minimum, there needs to be ideological coherence to the policies, grounded in principles that align with their global impact.
Mark Zuckerberg has now, helpfully, made it crystal clear that Meta’s content rules will prioritize free expression.
Variable Expertise & Training
Making policies for platforms with billions of users is a monumental task.
Fact-checkers did their best to navigate Meta’s rules but they often faced ambiguity. For example, should a fact-checker consider the potential harm of a piece of viral false content when reviewing it? If so, they might adopt a more conservative stance when rating it, reviewing it more harshly than they otherwise would.
Some of my colleagues were excellent first-principles thinkers, adept at identifying and interrogating these kinds of edge cases and alternative scenarios. Others weren’t.
This inconsistency raises a fundamental question: Who decides what “good” looks like in content policy and governance? What qualifications or training equip someone to shape rules that influence billions of people? The lack of clear expertise and standards for this work - and the immense power it entails - should concern us.
Shifting Values Embedded in Technology
Technology is infused with the values of its creators. What values did companies like Meta prioritize when developing fact-checking policies? Are those values transparent?
At times, content decisions could reflect the goals of individuals and their interpretations, rather than a principled long term strategy.
These challenges go beyond Meta.
They speak to a broader issue in tech: the need for clarity, expertise and accountability in decision making processes that shape digital infrastructure. This is particularly relevant as AI companies confront similar questions in their alignment efforts.
From Meta Fact-Checks to AI: Lessons for AI Alignment
Meta’s experiment with fact-checking reveals a lesson for AI: transparency and coherence aren’t optional—they’re foundational.
AI alignment isn’t just about ensuring models behave ethically or avoid harm. It’s also about defining who decides what “ethical” means and equipping them with the tools and training to do the work well.
A few lessons:
1. Articulate transparent values
Tech companies must clearly articulate a coherent set of values that underpin their policies and products. For AI alignment, this means defining the principles that guide system behavior, how and why they were created, and being transparent about how those principles are applied. Create processes to revisit and evolve your values and interrogate missteps. Without this, alignment efforts risk becoming as fragmented as platform policies.
2. Professionalize policy development
Policymaking for AI systems should be treated as a professional discipline. This is challenging: I have not seen 1:1 skill mapping from, for example, a PhD in philosophy or a law degree and time in legal practice, to doing this work effectively. Excelling in one field does not make someone adept in another, but it’s difficult to judge performance.
Testing candidates for first principles thinking, epistemological humility, creativity, and analytical rigor are good starting points.
Combining this with other technical approaches to alignment are promising and necessary, such as deliberative alignment and relational alignment, to systematize human-machine ethical dilemmas. Remove as much subjectivity as possible and make rules development transparent and open to critique.
3. Recognize the power of digital infrastructure
The people shaping digital infrastructure - whether content policies or AI alignment strategies - shape our political and social realities. Companies must ensure they are equipped to wield this power responsibly, decentralizing some of that power, fostering a culture of dissent, creativity, and productive dialogue.
Meta’s fact-checking program was an experiment in navigating the complicated intersection of technology, values and global impact. It wasn’t perfect, but its legacy offers critical lessons.
Companies must now prioritize transparency and intentionality - not just in what they build, but in how they decide to build it.
Meta’s decision marks the end of one experiment, but it raises a bigger question: How do we ensure alignment in AI systems that shape billions of lives?
Super interesting! Thanks for writing this up!
This sort of thing was why my own bsky proposal strongly separated the content data and the curation process, so the same underlying content data could be curated differently according to different national legal systems and to different community norms.
I also coined the term "San Francisco Values", which I am glad to see occasionally spreading, to describe the community norms that "rich tech people in the bay area" think are universal humanistic values, that literally everyone else in the world ranges from mild discomfort to outright raging hatred for.
The fact is, curation needs to be localized, while underlying content data needs to be "everything legal in the area where the datacenter is".