Content moderation in African languages is thinly resourced relative to the number of people who speak them, and the cost of that gap falls on users in two directions at once: legitimate speech is removed because a system misreads it, and genuinely harmful content stays up because no system is reading it at all.
Both failures are invisible in the metrics platforms publish, which is part of why the gap persists.
Why the content moderation gap exists
Automated moderation depends on training data. Training data depends on large volumes of labelled text, which depends in turn on commercial incentive and academic attention. Languages that carry a great deal of human conversation but comparatively little advertising revenue attract neither.
The result is structural rather than malicious. A classifier trained overwhelmingly on English performs badly on Wolof, Tigrinya, Fulfulde or Cameroonian Pidgin — and worse still on the code-switching that is normal in actual speech, where a single sentence may move between three languages. Human review, the fallback when automation fails, requires reviewers who speak the language, understand the local political context, and are available in the relevant time zone. That is expensive and it is where budgets get cut first.
Platforms publish some of this in their transparency reporting — Meta's transparency centre is the most detailed — but language-level breakdowns of enforcement accuracy generally are not included, which makes independent assessment difficult by design.
The two failures, and who absorbs them
Over-removal hits ordinary speech. Political criticism, reporting on violence, reclaimed slurs, satire and religious language are all routinely misread when the reviewer or the model lacks context. For a journalist or an activist, an account suspension in the middle of a crisis is not an inconvenience; it is the loss of their distribution at the moment it matters most.
Under-removal hits targets of coordinated harm. Incitement, organised harassment and dehumanising language in a language nobody is monitoring simply remains. Where that language is tied to ethnic or political conflict, the consequences are not confined to the platform.
These failures are not symmetrical in who they affect, and neither shows up in a headline enforcement figure.
What actually moves the needle
Complaint volume alone does not. Documentation does. The organisations that have won changes have done so by producing evidence of a specific, repeated, categorisable failure.
The components of a usable case:
- A defined language and dialect, named precisely rather than as "local language".
- Multiple instances of the same failure type, captured with screenshots, timestamps, account context and the enforcement notice received.
- The correct reading — what the content actually meant, explained by someone who speaks it, so the error is legible to a reviewer who does not.
- Demonstrated pattern, showing this is systemic rather than a single bad call.
Cases built this way have a route. The Oversight Board takes appeals on Meta's decisions and publishes reasoned determinations. ARTICLE 19 and CIPESA pursue policy engagement, and Global Voices has documented language-specific moderation failures across multiple regions.
What organisations should do in the meantime
Assume your distribution can vanish without warning, and plan accordingly.
Keep an owned channel — a mailing list, a website, an SMS list — that no platform decision can switch off. Archive your own content, because you cannot appeal what you cannot produce. Document every enforcement action against you at the time it happens, with the notice text, rather than reconstructing it later. Where your audience is reachable by radio, treat that as infrastructure rather than a legacy channel; our guide to community radio partnerships covers how those relationships are built.
The measurement problem underneath
The deeper issue is that moderation quality is measured in aggregate while it is experienced locally. A system that is 99% accurate globally can be close to useless in a language representing a fraction of a per cent of the corpus, and the global figure will never show it.
Until enforcement accuracy is reported by language, external assessment depends on exactly the kind of ground-level documentation described above. That work is slow, unglamorous and currently the only thing that produces evidence at all.
Common questions
Why not just hire more moderators?
It is necessary but not sufficient. Reviewers need language, local political context and workable conditions — and reviewing violent content at volume carries a documented psychological cost that thin contracting arrangements rarely account for.
Does AI translation solve this?
Not reliably. Machine translation degrades on low-resource languages and on code-switching, which is precisely where moderation decisions are hardest. It can assist triage; it cannot substitute for comprehension.
What can an individual user do?
Appeal every wrongful action, and keep the record. Individual appeals rarely succeed alone, but they are the raw material from which a documented pattern is built.
This connects directly to shutdown advocacy, where the same documentation discipline applies — see digital rights advocacy after the shutdown ends.