EasySLR is an AI-assisted platform for systematic and targeted literature reviews. It's designed to augment evidence synthesis teams and can be configured from human-only workflows to AI-assisted workflows. In all cases, review teams remain accountable for protocols, decisions and conclusions.

This page explains how we use AI. Our approach is aligned with the principles set out in Responsible use of AI in evidence synthesis (RAISE), which provides guidance on the responsible development, evaluation, selection and use of AI tools in evidence synthesis, together with the 2025 position statement from Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence. It also reflects more recent task-based guidance on how different AI tool classes are being used across review tasks.

1. How EasySLR uses AI

EasySLR supports the main stages of an SLR or targeted review, including:

• Search strategy support
• Protocol support
• De-duplication
• Title and abstract screening
• Full-text screening
• Conflict analysis (human-human and human-AI)
• Data extraction during screening and full data extraction
• Risk of bias and appraisal checklists (14 built-in checklists supported; AI may pre-populate responses, but all outputs require human verification and approval)
• Report generation
• AI agent for targeted reviews

AI modes

Projects can be run with varying levels of AI involvement:

All modes retain human accountability, audit trails and configurable governance controls.

Review configurations

Separately from AI usage, projects can require either one or two independent reviews per article:

• Single reviewer: Suitable for targeted literature reviews, rapid reviews, scoping reviews, etc. QC workflows allow sampling, review and override of decisions.
• Dual reviewer: Suitable for systematic literature reviews or reviews with AI as one of the reviewers. Reviews can be by two humans or one human and one AI. Decisions are blinded, with a human conflict resolver for disagreements.

Governance and controls

• QC workflows: Project owners can sample, review and override any decision, human or AI
• Audit trails: All decisions are recorded with clear identification of which reviewer, human or AI, made each decision
• Governance settings: Organisation administrators can enable or restrict AI by project and stage, with usage limits where required
• Stage approvals:The AI agent can run with manual approval between stages or auto-progression, depending on the team's workflow

Key design choices

• Humans remain responsible for protocols, final decisions and interpretation
• AI outputs are traceable to specific articles, with suggested values linked to highlighted text in source PDFs
• For screening decisions and extracted fields, AI rationales are surfaced for reviewer verification rather than treating the system as a black box
• All AI decisions are labelled and exportable in Excel for external audit

These design choices aim to be consistent with the core RAISE themes and the 2025 joint position statement: human accountability, transparent use, function-specific evaluation and appropriate oversight.

AI technology

EasySLR uses large language models (LLMs) via API from established providers. These models are trained by the vendor and are not fine-tuned on customer data. For full-text screening, data extraction and checklist support, the platform uses retrieval-augmented generation (RAG), converting PDF documents into structured text and retrieving relevant passages. This is designed to ground outputs in source passages from your PDFs. Human reviewers still verify and approve outputs before use.

How EasySLR maps to RAISE tool classes

RAISE groups tools into five broad classes: 1. Rule-based algorithms, 2. Unsupervised classifiers, 3. Supervised classifiers, 4. Generative LLMs, and 5. Agentic AI. The summary below shows how EasySLR maps to the classes relevant to our workflow. See the RAISE guidance.

2. Evidence on performance

Multiple organisations have carried out evaluations and reported their findings through conference abstracts, posters and manuscripts. A consolidated list is available on our Publications page.

Published evaluations

What this means

AI-assisted workflows can deliver high recall and meaningful time savings, though performance varies by topic and human oversight remains essential.

We recommend:

• Running a pilot project before large-scale adoption
• Setting project-specific targets for recall and agreement
• Using EasySLR's reviewer quality statistics and conflict reports to monitor AI performance

Additional studies are in progress and will be added to our Publications page.

3. Human oversight

EasySLR is built on the principle that review teams remain accountable for their reviews. Here's how that works in practice:

• Project owners configure protocols, AI modes and decision rules
• Leads can see who made each decision, including AI decisions
• AI outputs are visible, attributable and reviewable within the workflow
• QC mode allows sampling and override of any decision
• Credit limits allow governance teams to control AI usage
• For higher-stakes reviews, we recommend dual human screening or AI Assistant mode until local pilots validate broader automation modes
• We recommend documenting AI configuration (mode, stages, model identifiers) in the protocol or methods section

4. Data protection and privacy

EasySLR is provided as a secure cloud service hosted on Amazon Web Services (AWS). We follow recognised security and privacy standards:

Key data protection points

• Ownership: You retain ownership of your uploaded content and results
• No training on customer data: Customer data are not used to train EasySLR models
• Zero data retention with our AI model provider: For customer content sent for AI processing, we use zero data retention
• LLM provider controls: We use provider data controls and contractual terms intended to prevent submitted content being used for provider training
• Role-based access: Access controls restrict access to authorised users, typically the project team; organisation administrators have configurable permissions
• Regional deployment: Enterprise customers requiring EU or US data residency can request separate deployment regions
• DPAs available: Data Processing Agreements provided to institutional customers

Full details are in our Privacy Policy, Terms of Service and Trust Center. Security documentation and audit reports are available under Non-Disclosure Agreement for regulators and data protection teams.

5. Limitations

We're transparent about where EasySLR works best and where it has limitations.

Language and domain coverage

EasySLR is optimised for English-language systematic reviews in clinical, economic, humanistic and epidemiological domains. Performance may be weaker for:

• Non-English texts
• Very recent terminology
• Highly specialised subdomains poorly represented in LLM training data
• Some forms of grey literature

Users should pilot EasySLR on a subset of records for each new protocol and topic. That said, researchers have used the platform successfully beyond typical domains, including road safety, implementation science and governance.

Bias in the evidence base

AI reflects the published literature and underlying model training data, which can under-represent some regions, languages and study types. Protocols and search strategies should still be designed with equity and inclusion in mind. EasySLR's explanations and reviewer statistics can help identify patterns in AI behaviour.

Model updates

Third-party models evolve over time. We validate new versions internally before deployment and run regression tests on benchmark projects.

6. RAISE alignment

RAISE provides guidance on the responsible use of AI in evidence synthesis. More recent task-based guidance adds task-specific recommendations by tool class. EasySLR is designed to align with the main themes across both.

7. Resources

• Research & Publications
• Knowledge Base
• Video Tutorials
• Live chat available within the platform