Insights from Auditing AI Chatbots for Survivors of Domestic Violence

You have probably heard and interacted with general-purpose chatbots like ChatGPT, Claude, and Gemini.

But, how about Sophia? Aimee Says? Ruth? Hope Chat?

Artificial Intelligence (AI)-powered chatbots are emerging to help domestic violence survivors, and various advocacy organizations have already integrated them into their websites.

To have a better understanding of the current state of this growing ecosystem, my collaborators and I conducted a research project where we systematically identified and audited AI chatbots designed for survivors of domestic violence.

We found there are currently 36 domestic violence chatbots deployed across 22 countries, and the majority of these chatbots are initiatives led by nonprofit organizations.

There are two main types of domestic violence chatbots we identified: rule-based (26) and LLM-powered (10). In rule-based chatbots, survivors can interact with some predefined options (e.g., buttons with common questions), while LLM-powered chatbots allow survivors to provide any text as input, and a few even support voice input. These chatbots are accessible via multiple mediums, including via websites (22), WhatsApp (14), Telegram (2), and Viber (2).

These chatbots have promising features. Several educate people on what a healthy relationship looks like and help survivors identify different forms of abuse. Others offer interactive safety planning, guidance on storing digital evidence securely, and practical digital safety advice such as adjusting privacy settings or detecting spyware. Many act as gateways to support networks by directing users to helplines. They offer survivors another route to information and support.

Other findings were more mixed. Thirteen of the 22 web-based chatbots implemented quick exit buttons, a good practice adopted from domestic violence advocacy groups' websites. However, only five effectively remove the chatbot visit from the browsing history, which an abuser might monitor. Only a few chatbots asked follow-up questions to contextualize a survivor's situation, an important safeguard.

Other shortcomings are worth addressing too. For instance, various chatbots do not clearly communicate that they are an AI and not a human. Live chats staffed by human advocates have been used in this space for years, so some people may reasonably assume they are talking to a real person.

Also, 16 of 36 chatbots have no privacy policy at all; only 7 have one specific to the chatbot itself. As these chatbots' deployment often involves several organizations, each with different privacy policies, it is unclear who is responsible for survivors' data.

We developed a list of questions advocacy organizations can reflect on when considering integrating a domestic violence chatbot into their website (see Appendix F in the paper):

We also created a checklist for developers working on these initiatives (see Appendix G in the paper):

If you are interested in more of our findings, I invite you to read our paper: https://dl.acm.org/doi/abs/10.1145/3805689.3812260

Image at the top of this post

It is a screenshot taken from https://www.thehotline.org/

Correction

Table 9 in Appendix I transposes two values between adjacent rows: it lists 6 ad-tech companies for Ruth and 0 for SafeChat, when the correct values are 0 for Ruth and 6 for SafeChat. A sentence in Section 4.4 repeats the incorrect figure. All other values in both rows are correct, and the correction does not affect the paper's findings. A corrigendum request has been pending with ACM since July 14, 2026.