A key risk associated with using LLMs in evidence synthesis is hallucination and loss of traceability—where the model generates confident but unsupported outputs, or introduces information that cannot be linked back to the source study. This is especially problematic in systematic reviews, where transparency, reproducibility, and auditability are critical.
EasySLR has been specifically designed to mitigate this risk. Across screening and full-text stages, we apply strong context and prompt constraints so the LLM operates only within the boundaries of the relevant citation or PDF, rather than drawing from general knowledge. In the Full-Text stage in particular, all AI outputs are explicitly linked back to source passages in the document, ensuring that no information is generated without a verifiable reference. This substantially reduces hallucinations and allows reviewers to quickly validate AI suggestions.
From a performance perspective, with the continuous improvement of LLMs over the past few years, we now regularly observe ~85–90% accuracy at the screening stages. However, rather than relying solely on raw model capability, EasySLR focuses on engineering safeguards—prompt optimisation, context control, and source grounding—to ensure stable and reliable behaviour.
What’s novel compared to many other platforms is that EasySLR was built as an AI-first system from day one, not as a traditional review platform with AI added later. This has allowed us to deeply integrate AI into the workflow, rather than treating it as an external add-on. In addition, we’ve invested heavily in usability and infrastructure, ensuring the platform remains responsive and easy to use even with large citation volumes—addressing everyday reviewer pain points such as speed, scalability, and setup complexity.