When Algorithms Decide: Challenges for AI Accountability in Nepal

A project reflection by Open Knowledge Nepal, supported by the Pulitzer Center through the South to South (S2S) AI Accountability CoLab microgrant.

Nepal is not usually the first country that comes to mind in global conversations about AI accountability. That is partly the problem.

When we began this project in late 2025, the question we kept returning to was deceptively simple: what does algorithmic content moderation actually do to people in Nepal? Not in theory. Not as a policy abstraction. But to the creator who watched her reach collapse without explanation, the journalist who learned which words were dangerous by trial and error, the queer advocate whose community lost its digital infrastructure one account removal at a time.

Answering that question turned out to be harder than we expected. 

Situating the work

Nepal sits at an uncomfortable intersection. The government is moving fast on digital governance, with bills, policies, and platform bans already in play. At the same time, the platforms themselves are largely absent. Meta maintains no local office or support for Nepal. TikTok removed approximately 9.12 million Nepali videos across 2025, nearly 1.9 million in a single quarter, with almost no transparency on why, and with automated systems responsible for those removals. 

Civil society is caught between these two forces: a government inclined toward restriction and platforms that are structurally invisible. Our project tried to occupy the space between them, documenting what is actually happening in terms that affected communities can recognize and that policymakers cannot easily dismiss.

That is what makes this work relevant to accountability. Powerful actors, both state and platform, make consequential decisions about whose voices are heard online. Accountability requires evidence. Evidence requires attention. 

The translation problem

The most persistent challenge we faced was not access or resources. It was a translation. 

There is no shortage of global research on algorithmic harm. Studies from Africa, Latin America, the Middle East, and South Asia document the same structural patterns: English-built systems deployed globally, consistently failing communities in the Global South. The challenge is converting that body of evidence into claims that are locally valid and locally legible. Does this finding apply to Nepal? Does this mechanism work the same way here? Is the experience creators are describing consistent with what the research predicts, or is something different happening? That process of alignment, between global research and local reality, between documented patterns and specific stories, is slow and painstaking. There are no shortcuts.

The second challenge is awareness. In Nepal, as across much of South Asia, the basic premise of this work cannot be assumed: that automated systems make consequential decisions about your content; that those decisions reflect structural biases; and that this matters. Convincing the public that something invisible is worth paying attention to is itself a task, one that comes before any advocacy can begin. The zine and the educational carousels were partly a response to this, making the invisible legible before asking anyone to act on it.

The third challenge is the legislative landscape. Nepal's digital governance framework is still being written. The Social Media Bill, now withdrawn, the AI Policy, and related regulations are all at various stages, and the national position on platform accountability is genuinely unsettled. That fragmentation makes it difficult to identify a single pressure point or to build a coalition around a shared target. The ground keeps shifting. 

What the network made possible

The S2S peer learning sessions were unexpectedly valuable, not as a process requirement, but as a genuine reorientation.

Hearing from grantees working on AI accountability in other parts of the Global South helped us see our own project differently. What we were documenting in Nepal was not a local anomaly. It was the same pattern, playing out in a different context. That comparative lens strengthened our analysis and gave us language to connect Nepal's experience to a wider argument about how these systems are built and who bears the cost. 

The Pulitzer Center's journalism, particularly Policing by Proxy and the methodology about covering misinformation on TikTok, gave us both a methodological foundation and a framing device. It lets us anchor local findings in credible, independent reporting rather than building the evidentiary case entirely from scratch. The limitation, as above, is that the translation still has to happen; journalism produced in one context does not automatically transfer to another. But it gives you a starting point, and that matters. 

What we are watching

The online accountability story is important. But we are increasingly aware that it is not the whole story. 

Through the peer learning network, we encountered examples of AI implementation with consequences far beyond social media. In India, AI-powered face matching used to verify benefits recipients has failed in ways that denied people access to welfare they were entitled to, not because of policy intent, but because the technology did not work reliably on people the training data had not anticipated. The harm was not online. It was at the ration shop, the hospital, the government office.

Nepal is moving rapidly toward digital governance and e-government systems. The window to ask hard questions about how these systems will be built, who they will serve, and what happens when they fail is open now. Before deployment, not after, we want to be part of that conversation. 

Key findings from the situation in Nepal

•  Enforcement at scale, accountability in absence 

  TikTok removed approximately 9.12 million Nepali videos in 2025, nearly 1.9 million in the final quarter alone, with 99.6 to 99.9 percent removed by proactive automated detection rather than human review. YouTube ranked Nepal among the top 25 countries globally for removals, and Meta publishes no Nepal-specific enforcement data at all. Enforcement happens at volume; meaningful accountability does not. 

• Three structural gaps

  Every finding points to the same three gaps: the transparency gap (platforms remove content at scale without publishing data to evaluate accuracy or equity), the language gap (automated systems were not trained on Nepali, Maithili, or Newari, but the consequences fall on Nepali creators), and the accountability gap (when something goes wrong, there is a form—and no one knows if it works). 

• What the numbers cannot show

  Community research surfaces the human texture behind the data. Creators discover they have been shadow-banned only by watching their own analytics. A journalist learns which words are unsafe by trial and error. Self-censorship becomes automatic, built into the creative process before content is even made. Language operates as a structural disadvantage, and appeals remain unknown to most and inaccessible to many. 

• Concrete recommendations for three actors

  Platforms should publish language-level enforcement data, establish a real Nepal contact point and creator-facing grievance mechanisms, and invest in local-language training data. Government should make platform registration a genuine accountability mechanism, pursue a binding AI Act, and clearly distinguish regulation from removal. Civil society should build a public log of moderation incidents and develop Nepali-language literacy materials. 

• A global pattern, not a local exception 

  Nepal sits inside a documented global dynamic. Research across Africa, Latin America, South Asia, and the Middle East records the same pattern: English-built moderation systems deployed worldwide that consistently fail communities across the Global South. The language differs; the outcome is the same.

What comes next

This project produced a foundation: platform data, community stories, a public evidence base, a diagnostic report, and a network of stakeholders who have now sat in the same room together. That foundation needs to be built on.

In the near term, we see two directions. The first is deepening the platform accountability work: expanding community monitoring, building a systematic public log of moderation incidents in Nepal, tracking the evolving regulatory landscape, and contributing to the emerging coalition of AI accountability, safety, and governance groups now forming in the country. The second is broadening the lens, moving from social media moderation to the wider landscape of AI in public services, informed by what we have learned from peer networks about how that terrain looks in neighboring countries.

Neither direction requires starting over. Both require continuity of funding, of relationships, and of the slow, careful work of building evidence that holds powerful actors to account. 

Explore the research 

This reflection accompanies two public resources produced by the project: 

Diagnostic report: The full research report, methodology, findings, and recommendations. Read and download: https://oknp.org/resources/when-algorithms-decide-diagnostic-report 

Zine: An accessible companion documenting creator stories and the three structural gaps. Read and download: https://oknp.org/resources/when-algorithm-decides