Q: What is the first step when assessing AI tools for systematic reviews? A: When assessing AI tools for systematic reviews, start by using the responsible handover framework from RAISE 3 to clarify the tool’s purpose, provenance of training and validation data, validation status, usability, and transparency/licensing. This assessment can rely on public information and may include contacting developers for missing details before you decide to proceed. Q: What red flags should make me decide not to use an AI tool? A: When assessing AI tools for systematic reviews, watch for red flags such as lack of relevant validation, validation that is not replicable, performance claims based only on developer‑led studies, legal or policy conflicts, terms that permit reuse of your content for training without opt‑out, inadequate human oversight, or poor developer transparency and support. If you see any of these issues, the guidance is to stop and not proceed with the tool. Q: How can I demonstrate an AI tool will not compromise the methodological rigour of my review? A: When assessing AI tools for systematic reviews you should use evidence gathered via the responsible handover framework and then carry out human verification or validation as needed. RAISE guidance categorises tool uses from acceptable to not acceptable and recommends disclosure, human verification, or validation within the review (a SWaR) depending on the category, with current advice to proceed with mitigations for large language models. Q: What performance thresholds should teams set before validating an AI tool? A: When assessing AI tools for systematic reviews, set performance thresholds prospectively by defining a futility boundary (minimum acceptable) and a non‑inferiority margin (target or aim). CESAR gives example thresholds such as screening sensitivity stop if under about 80% and aim for an upper confidence limit near 95%, full‑text specificity stop near 50% and aim around 60%, data‑extraction sensitivity stop near 92% and aim near 97%, major extraction errors stop if above about 3% and aim for about 2% on upper limits, and usability (SUS) stop if in the high‑50s and aim for mid‑70s. Use these examples as guides, justify your own thresholds for the task and risk level, and adhere to them during validation. Q: What should I include in my review to transparently report AI use? A: When assessing AI tools for systematic reviews, fully disclose the AI system name, version, date of use, developer, specific purpose in the review, any customisations or parameters, and the degree of human oversight. Also report how you judged it methodologically sound, any validation or calibration performed in your context, known limitations or biases, and links to protocols or supplementary methods. Q: Is human oversight always required when using AI in evidence synthesis? A: Yes — expectation four in Cochrane’s guidance is that AI must be used with human oversight, because people remain accountable for research decisions and outcomes. When assessing AI tools for systematic reviews you should document monitoring, verification or override steps and be prepared to stop use if outputs fail your checks or thresholds. Q: When is it acceptable to proceed with an AI tool that has incomplete validation or evidence gaps? A: When assessing AI tools for systematic reviews, you may proceed with mitigations if a tool is promising but has evidence gaps or moderate, monitorable risks, provided you plan additional verification or validation. Such mitigations can include human verification for critical outputs or running a study within a review (SWaR) to validate the tool’s performance in your context. Q: What practical signs should make teams stop using an AI tool even if it passes validation? A: A tool should be stopped if it fails to meet predefined performance thresholds, harms usability by slowing workflows or increasing confusion after standard onboarding, or lacks adequate support or auditability. When assessing AI tools for systematic reviews, these practical usability and support failures are valid reasons to discontinue use even if prior validation existed.