AI responses in survivor-serving contexts need more than generic quality review. Reviewers should look for missed coercive-control dynamics, overconfident guidance, unsafe action steps, privacy exposure, and language that removes survivor agency.
This checklist is for training, research, product review, policy discussion, and quality improvement. It is not crisis support, direct survivor support, legal advice, clinical advice, individualized safety planning, or safety certification.
- Use synthetic, public-safe, de-identified, or training-safe scenarios only.
- Compare multiple response modes rather than relying on a single output.
- Review privacy, coercive control, agency, actionability, lethality awareness, accuracy, and harm potential.
- Document what improved, what failed, and what still requires qualified human judgment.
- Do not treat a score or checklist as a safety certification.
- Keep prompts, case details, evaluation history, exports, and sensitive review work inside appropriate private systems.
Before deeper use, read the trust boundaries, review features, and consider whether a controlled pilot is appropriate for your team workflow.