AI-assisted literature review needs an audit trail, not a disclaimer

‘Human in the loop’ is not a method

Many research services now promise AI speed with human oversight. The phrase sounds reassuring because it names the correct instinct: do not let a language model make unsupported research judgements. But it does not say what the human checks, when the check happens, or what evidence is preserved.

A person who approves a fluent final paragraph without seeing the search, exclusions and underlying papers is technically in the loop. They are not providing a meaningful control. A credible workflow defines checkpoints around the failure modes of each stage.

Where AI can help

Models and related tools can be useful for query expansion, deduplication support, prioritising records for screening, extracting candidate fields, classifying study characteristics, comparing terminology, and drafting structured summaries from verified notes. These uses can reduce mechanical work and help a researcher inspect a larger evidence space.

The benefit is strongest when the task is bounded and the output can be checked against a source. ‘Extract the sample size from this paper’ is a different task from ‘decide whether this literature proves the product works.’ The first has a reference answer. The second combines appraisal, synthesis and judgement.

Where it fails quietly

Language models can invent citations, misstate study details, collapse conflicting findings, miss negation, treat indirect evidence as direct, and produce a coherent explanation that is stronger than its sources. Research on LLM use in medical systematic reviews has highlighted both potential efficiencies and substantial risks, including hallucination and limitations in tasks that require methodological judgement.

The dangerous errors are not always dramatic. A fabricated paper is detectable. A real paper cited for a claim it only partly supports may survive casual review. So can an exclusion that silently removes inconvenient evidence. Reliability depends on preserving the chain of decisions, not only checking whether citations exist.

Design checkpoints around the pipeline

• Question check: a researcher confirms the decision question, scope and definitions before searching.
• Search check: queries, databases, dates and limits are recorded; known key papers test whether retrieval is plausible.
• Screening check: inclusion rules are explicit, uncertain records are escalated, and a sample of exclusions is reviewed.
• Extraction check: important fields are verified against the full text, with page or section references where useful.
• Appraisal check: study quality and directness are assessed using an appropriate framework, not inferred from tone.
• Synthesis check: every material claim is source-linked; disagreement and missing evidence remain visible.
• Conclusion check: the strength of the recommendation is compared with the strength and relevance of the evidence.

If you need these checkpoints applied to your own evidence base, you can commission a structured literature review with agreed search coverage, methodological appraisal and source-linked conclusions.

Keep three layers separate

A strong brief distinguishes observation, estimate and interpretation. Observation reports what a source directly states or measures. Estimate combines evidence under explicit assumptions. Interpretation explains what the result may mean for the decision.

Models often blur these layers because fluent prose rewards continuity. Keeping them separate makes review easier. A reader can challenge an interpretation without disputing the observation, or inspect the assumptions behind an estimate without dismantling the entire report.

Questions to ask a provider or internal team

• Can you show the search strategy and explain its limits?
• Are citations verified against full texts or only abstracts and snippets?
• How are exclusions recorded and sampled for error?
• What quality or bias framework is used?
• Can each important conclusion be traced to evidence?
• How are conflicting results reported?
• Which tasks are automated, and which require researcher judgement?
• What would make you lower confidence or recommend a fuller review?

If these questions cannot be answered, the issue is not whether AI was used. The issue is that the research process is not inspectable.

The right promise

AI can make evidence work faster and more systematic. It does not remove the need to define a question, judge relevance, assess methods, preserve uncertainty and take responsibility for the conclusion.

The credible promise is therefore modest and useful: a transparent, source-linked workflow in which automation handles bounded tasks and a researcher owns the decisions. That is more informative than ‘human reviewed,’ and far easier to audit.