Project summary

Most of AI safety testing happens in English, one prompt at a time and is often translated. But that’s not how people who actually use these systems talk or write. They use own languages and dialects, over longer-turn conversations.

So through the PRISM AI Safety Research Fellowship, our team built LocalTurn: three-turn conversations written by native speakers in Emirati Arabic, Bangla, Sinhala and Yoruba, mainly around religious and political harms. We tested 8 different commercial AI systems with them.

Our paper got accepted as a poster at the LP4FM workshop at NeurIPS 2026 in Sydney and I'm the chosen fellow to present our project. I’m asking for help with the part of the trip that can't be covered by me.

🔗 Project Web Page: https://prism-research.org/projects/multilingual-safety-evals

🔗 Paper: https://openreview.net/forum?id=xTdevd1UGw&referrer=%5BAuthor%20Console%5D

What are this project's goals? How will you achieve them?

I’m aiming to connect with the brilliant people who’ll be at LP4FM, especially researchers working on low-resource and African languages and on agentic AI. As a team, we want our project to become something people actually use for more honest safety reporting, e.g., red-teaming and safety evaluations, in the languages that get left out, and we’re also looking for collaborators to grow it into an open resource with more languages and dialects.

How will you know it worked?

I present our poster at Sydney, get valuable feedback from the researchers who will all get to be in the same place that day, have at least some follow-up conversations with people who want to participate with us & lastly we get to move closer to releasing our paper data LocalTurn as an open-source to help the AI safety community.

How will this funding be used?

I reached the minimum goal of $1,000. This gracefully covers my hotel accommodation for 4 nights near the venue for $979.

Reaching the full $2,215 would also cover my food for 5 days (~$500), airport rides and light rail (~$160), a top-up for the nonstop Emirates flight (~$415), travel insurance (~$60), and printing the poster (~$100).

Who is on your team? What's your track record on similar projects?

Our accepted paper was built by our team in the PRISM AI Safety Research Fellowship, led by Naiyarah Hussain, with fellow researchers Abdullah Ibne Hanif Arean, Fathima Raihaan Ihsan, Victor Tolulope Olufemi, and Fatma Egal (me), plus the help of 31 native-speaker annotators and 30 evaluators across four languages to make the dataset come to life.

I’m a co-author of the paper and I wrote the native Emirati Arabic prompts (51 multi-turn conversations), wrote the abstract, introduction and conclusion, and led outreach to recruit and onboard our Emirati Arabic volunteers. I also led our team’s progress presentations at PRISM, and I’m the voice on our project video. [Video here: https://www.youtube.com/watch?v=ZJAMmZ7Zt2E]

Outside this project, I’m a core team member of AI Safety UAE, where I handle creative and events, helping build the UAE AI Safety community.

What are the most likely causes and outcomes if this project fails?

If it fails, I don't get to present the hard work we have done in the 16 weeks of the fellowship. I also don't get to be in the presence of lots of other brilliant researchers who are going to be there at the workshop conference to connect & get feedback.

But the project itself will keep going. It doesn't end here for us.

How much money have you raised in the last 12 months, and from where?

PRISM (Women Who Do Data) is covering registration and partially some of the flight funds. I applied for NeurIPS Financial Assistance, the NeurIPS volunteer programme, and a BlueDot Rapid Grant, but none came through.