Fighting disinformation in today’s digital landscape is daunting, especially in linguistically diverse regions. For six years, I’ve worked across Africa, combating harmful digital narratives on platforms like those owned by Meta. While I’ve focused primarily on English-speaking nations, I’ve also tackled falsehoods in Swahili, Amharic, Oromo, Hausa and Zulu.

Meta’s automated filters and content moderators often fall short when dealing with non-English content. This isn’t just a technical limitation — it’s a dangerous oversight. In communities where English isn’t the dominant language, this gap leaves space for disinformation to flourish unchecked.

Two recent examples show the effects disinformation can have.

In Ethiopia, as tensions rise over access to Eritrea’s Port of Assab on the Red Sea, a vital strategic asset it lost following Eritrea’s secession in 1993, false Amharic-language Facebook posts claimed Ethiopian troops had seized the port. The images were digitally manipulated. But they spread widely and undermined diplomatic efforts to calm the tensions in the region.

Similarly, in Tanzania, ahead of the October general elections, a digitally manipulated video circulated on Facebook. It showed politician John Heche appearing to criticize fellow Chadema leader Freeman Mbowe. In truth, the unedited original clip captured Heche praising him.

As this kind of disinformation spreads, fact-checkers have taken dual roles — part journalist, part moderators — investigating and correcting false narratives. This is simply because major platforms like Meta and X struggle to flag or easily identify harmful content in local languages.

Since the advent of the COVID-19 pandemic, I’ve seen a sharp rise in falsehoods shared in African local languages. While at the wire service AFP, my team consistently raised this issue with Meta, urging the company to build tools or develop mechanisms to address the issue of flagging more harmful and toxic non-English content on the platform. For more than three years, no real progress was made. When I took a study break in August, the problem persisted.

Meta’s platforms cater to billions of users worldwide, a significant number of whom do not use English as their language of communication. Studies, including one from Harvard referenced below, have revealed that harmful content, such as misinformation, hate speech and violent or extremist material — more so in non-English languages, but also some in English — often evades moderation efforts. This challenge is especially pronounced in areas where Meta holds a dominant presence but lacks sufficient resources to address linguistic diversity.

Community Notes fall short

Meta’s replacement of professional fact-checkers with Community Notes in the U.S. raises serious concerns about its global strategy. A recent Harvard study showed that English-language posts are flagged at double the rate of their Spanish counterparts — 49% compared to 21%. That’s despite Spanish being spoken by more than 40 million Americans.

A leaked internal document from Meta posted on FBarchive.org — a repository of internal Facebook records — revealed that the tech company uses automated translations to convert non-English posts into English for moderation. The document warns that this method carries significant risks. Inaccurate translations can result in the removal of harmless posts or, conversely, permit harmful content to stay on the platform.

The risk goes beyond error. In one study, English speakers reviewing translated Mandarin content assessed it very differently from native speakers. This underscores the problem: automated translation isn’t good enough to moderate complex, context-driven speech.

FBarchive.org grants public access to leaked internal Meta documents provided by whistleblower Frances Haugen. These documents offer valuable insights into Meta’s difficulties in managing and moderating content on its platforms.

Linguistic inequity

Meta only reports statistics on content that it has removed from its platform, not what users report and it ignores. The lack of transparency is worse for non-English content, where moderation infrastructure is weak.

The issue becomes even more pronounced in non-English content due to inadequate infrastructure for language-specific moderation. While Meta claims that its hate speech detection tools can identify content in more than 100 languages, the company lacks transparency regarding the regional and language breakdown of its user base. Internal evidence shows variations in the accuracy of these detection tools, with some proving completely ineffective at identifying hate speech in languages or dialects that are not included in the programming.

The Harvard report highlights the impact of language coverage gaps in conflict regions, emphasizing how limitations in algorithmic moderation have allowed hate speech in languages like Amharic and Burmese to proliferate without adequate checks. This has contributed to real-world consequences, exacerbating tensions and challenges in countries like Ethiopia and Myanmar. Addressing these gaps is crucial for fostering safer online spaces and mitigating harm in vulnerable communities.

The findings of this study align with the on-the-ground experiences shared by fact-checkers working across multilingual settings.

Journalists across Africa have consistently sounded the alarm. For instance, Amharic — a language spoken by more than 57 million people in Ethiopia — is routinely exploited to spread harmful content related to the country’s ongoing conflicts. Similarly, Swahili, with over 150 million speakers across Africa, often serves as a vehicle for hate speech, propaganda and job scams. However, Meta’s response to these issues has been notably sluggish.

If Meta opts to replace fact-checkers with Community Notes in regions such as Africa, Latin America, Asia, or even parts of Europe where non-English languages are commonly used, the repercussions will be severe. Fact-checkers play a vital role in combating misinformation, especially in non-English-speaking areas where language-specific nuances are critical. Without targeted interventions and improved moderation strategies, such a shift could have dire consequences for regions already grappling with the harmful effects of unchecked content.

A journalist and researcher focused on disinformation in Ethiopia, whom I granted anonymity due to frequent harassment and threats targeting both him and his family, said that social media platforms in the country remain a vast hub for the spread of disinformation.

“Meta and X’s efforts to monitor and moderate harmful content in Ethiopian local languages have not been effective. Various empirical studies support this claim. Several reasons contribute to this issue, including the limited focus of content in Ethiopian local languages and the fact that major languages, such as Amharic and Afaan Oromoo, have not been adapted for machine learning techniques that could aid in monitoring harmful content. I question the expertise and capability of these platforms’ teams to address harmful content in these languages effectively,” the journalist said.

He added that this is concerning, given that Facebook in Ethiopia is viewed as “the internet” and hosts a significant amount of harmful and polarizing content.

“The Facebook tool appears to be ineffective in identifying harmful content that can be fact-checked in Amharic and Afaan Oromoo. Since I began monitoring the tool in 2021, the content flagged in these languages has primarily consisted of rumors or material that is too vague to be classified as misinformation or disinformation. Despite the widespread dissemination of harmful content on social media platforms in the country, the tool has rarely detected information related to the ongoing conflict in Ethiopia, disinformation, or other significant issues prevalent in these languages. This suggests a notable gap in the tool’s ability to effectively address content in local languages”.

Afaan Oromoo is spoken by more than 40 million people, predominantly in Ethiopia, where it serves as the native language of the Oromo ethnic group. It is also spoken by communities in northern Kenya, making it one of the most widely spoken languages in the Horn of Africa.

A fact-checker based in East Africa, who opted to remain anonymous due to her organization’s collaboration with Meta, expressed worry about the rise of divisive content posted in regional languages like Swahili, especially in Kenya and Tanzania. She pointed out a notable deficiency in content moderation for non-English languages common in the area, emphasizing how major platforms such as Meta and X often overlook or fail to identify such posts.

She shared that she has debunked multiple false claims in Swahili related to Tanzania’s upcoming elections and Kenya’s deeply divisive political climate. A recurring issue she’s identified in Tanzania involves the use of deepfakes — authentic videos digitally manipulated to include fake Swahili text.

Proactively, she regularly scours Facebook groups and X for toxic or harmful material shared in regional languages such as Swahili. When she detects misinformation, she promptly takes decisive steps to debunk it.