Building a Sex Ed Chatbot for the MENA Region – Here's What We've Learned
,Can a chatbot answer users' anonymous questions about sexual health and related topics, drawing its knowledge from DW's journalistic content? This article summarizes what our Lab team has discovered while developing a prototype.
"What happens to my body during puberty?" Or: "Can I use a condom twice?" These are examples of questions that users aged 14 to 25 from the MENA region could ask our chatbot مساحة أمان (Arabic for "Safe Space"). The bot is designed to provide fact-based answers to exactly these kinds of questions.
The unpublished prototype currently draws on around 100 articles and video scripts from DW's Arabic edition, covering topics such as sexual education, human rights, gender equality, and LGBTQ+ issues. These editorially verified materials form a reliable knowledge base for educational content.
Why are chatbots becoming more relevant for public media?
We're exploring conversational AI as more and more users turn to chatbots for information. According to the Reuters Institute Digital News Report, 7% of respondents say they use a chatbot for news on a weekly basis – and among those under 25, that figure jumps to 15%. Chatbots are not only transforming how people search for information online, but also opening up new ways to engage users through dialogue, which is one of DW's broader strategic goals.
During our research, we came across a number of similar projects from other media organizations and NGOs. We connected with the project manager behind Ask Aunty, an initiative from the Egyptian media outlet Raseef22, and also discovered Zuzi, a chatbot developed by the South African NGO Gender Rights in Tech (GRIT), and Disha Didi, from the Indian NGO Ipas Development Foundation (IDF).
RAG technology with DW knowledge at its core
Our chatbot uses a method called Retrieval-Augmented Generation (RAG), which combines a language model with a custom database. When a user submits a question, the system first searches its database and retrieves relevant excerpts from DW articles and video scripts. The model then generates a response based on both its training data and the retrieved content, including links back to the original sources.
Unlike well-known AI chatbots such as ChatGPT – which rely primarily on their internal training data – مساحة أمان / Safe Space draws its information directly from DW's editorially verified content. This allows it to access specific, fact-checked knowledge that goes beyond the model's training data, while also reducing the risk of so-called "hallucinations," i.e. fabricated or inaccurate information in the responses.
Sounds straightforward – but it isn't.
Although we've built a first rough prototype and a proof of concept, we've run into a number of obstacles along the way:
-
Safe technical infrastructure still needs to be established. Since we have no prior experience at DW with chatbots handling sensitive topics, guidelines and benchmarking must be developed with great care (this EBU report may serve as a useful reference). As a public broadcaster, we have a responsibility to make 100% sure our chatbot doesn't provide incorrect or potentially harmful advice.
-
Legal accountability is still unresolved. Depending on the chatbot's functionality and technical setup, it remains unclear whether it would be classified as a service provider or an operator under the EU AI Act–a distinction that carries different liability implications.
Content collection is partly manual. DW Arabia's relevant content is spread across dw.com and the team's various social media channels. While pulling articles from dw.com is relatively easy, sourcing content from social media remains a challenge and requires manual effort. As always, the better the data quality, the more accurate the results.
What's next?
Initial tests suggest that the chatbot doesn't yet draw sufficiently on DW articles and video scripts when generating responses. Our next step is to complete the database and, on the basis of that more comprehensive dataset, conduct a content analysis to assess whether the knowledge base is fundamentally sufficient to support well-founded answers – which would ultimately reduce the risk of inaccurate outputs.
Depending on those findings, we'll explore suitable backend infrastructure to serve as a home base for the bot and support further technical experimentation.
Special thanks to: Lea Ahlers, Annette Diegel-Ruppert, Nicole Goebel, Thomas Harders, Samar Heinein, Arian Ivec, Lynn Khellaf, Marie Kilg, Isabell Lorenzo, Alexander Matthews, Jasper Steinlein, Burak Ünveren and everybody else who supported the chatbot project.
