ALTO tests locally informed AI detection of online gender-based violence

A dataset built around local context

ALTO, short for African and Levantine Tech-Facilitated Gender-Based Violence Corpus, addresses gaps in automated detection for Levantine Arabic, African French and Swahili. The preprint, posted on November 4, 2026, is by Tianwei Zhang, Ivy Njoki Nyagwencha, Zahraa Dirani, Shalini Joshi and Scott A. Hale. Its findings are preprint results, not established through peer review.

The researchers use a framework that combines technology facilitation, gender-related targeting and harm. This reaches beyond explicit sexist insults to include covert hostility, sexual harassment, threats and doxxing: disclosing, or threatening to disclose, private identifying information. Crucially, mentioning women or gender does not automatically make content harmful.

How examples were collected and labelled

Civil society partners working in Jordan, Francophone West Africa and Kenya helped gather locally relevant material. Community members could submit potential abuse through messaging-based tiplines. Because participation varied across regions, researchers supplemented submissions with keyword searches on X, using terms curated by their partners.

The resulting pools contained 8,710 Arabic, 5,628 French and 4,798 Swahili items. Active learning selected subsets for annotation, combining examples chosen for diversity with those a developing classifier found uncertain. The approach concentrated annotation effort on selected material; the resulting dataset is not a representative sample of online content or a measure of abuse prevalence.

ALTO contains 3,022 annotated items: 954 Arabic, 1,026 French and 1,042 Swahili. Annotators were women with regional language expertise, recruited through partner organizations. The process included expert adjudication, training and discussions that refined the guidelines. Earlier examples were reannotated when definitions changed. Annotators also received training intended to mitigate trauma from exposure to harmful material.

Labels distinguish gendered references from harmful acts and organize harms into broad categories and finer subtypes. These include derogation, implicit animosity, objectification, violence threats and doxxing. Agreement varied across languages and annotation tasks, underscoring that these judgements were not straightforward or unanimous.

What the models detected

The experiments used 2,117 items retained under the study’s agreement and adjudication criteria. Researchers compared existing systems, including GPT-OSS-Safeguard, Perspective API and Learning from the Worst, with pretrained classifiers further trained using ALTO and existing benchmark data. Learning from the Worst received machine-translated English text; Perspective API was not evaluated for Swahili.

Fine-tuned models were evaluated on ALTO through five-fold cross-validation, with each item’s prediction coming from a model trained without that fold. The experiments reduced the detailed taxonomy to binary decisions about whether content constituted technology-facilitated gender-based violence.

Recall measures the share of labelled harmful examples a model catches. On ALTO, the strongest off-the-shelf recall was 68.6% for Arabic, 74.8% for French and 46.9% for Swahili. The jointly trained multilingual model reached 83.7%, 87.5% and 81.7%, respectively.

Higher recall did not mean uniformly better performance. Separate language-specific fine-tuning achieved higher F1 scores, which balance recall and precision, than joint multilingual training in each language on ALTO. Those scores were 0.818 for Arabic, 0.739 for French and 0.679 for Swahili. The Swahili-specific system itself used a multilingual pretrained model, so this comparison concerns the fine-tuning setup, not exclusively single-language architectures.

Subtle harms and false alarms

Implicit animosity, including stereotypes and condescending messages, was frequently missed by existing systems. Swahili also presented difficulties with objectification and doxxing. Fine-tuning reduced several of these errors, but gains came with substantial false alarms.

For gendered French examples labelled non-harmful, the joint multilingual model’s false-positive rate was 79%, compared with 45% for the separately fine-tuned model. The paper describes a French example in which a woman recounting a degrading characterization was incorrectly flagged. Recognizing gendered language is therefore not equivalent to recognizing an abusive act.

Performance also weakened on the separate benchmark collection, where positive examples were much rarer. Its Swahili test subset contained only three positive examples, none detected by the off-the-shelf systems. That tiny count limits conclusions about their wider Swahili performance.

What the evidence does not establish

The study demonstrates detection improvements on its evaluation data, not reduced violence or proven deployment safety. Scraped material came only from X, and image- and video-based submissions were excluded. Keyword searches could not guarantee exclusively Levantine or African French usage; Swahili examples also included mixed-language Sheng.

Other limitations include binary labels that flatten distinctions between harms, an all-women annotator pool that does not encompass all affected identities, and possible overlap with earlier model or benchmark data. Translation quality was not evaluated at scale.

For privacy, user tags and URLs were replaced, and doxxing-labelled content was excluded from the released dataset. The authors specify that moderation applications should operate alongside human review. ALTO offers evidence for locally informed training, while its errors show why detection scores alone cannot settle moderation decisions.