Qureshi et al. 2023 — Are ChatGPT and LLMs the answer to SR automation?

A commentary in Systematic Reviews documenting the PICO Portal webinar of 6 February 2023, where the authors tested ChatGPT (GPT-3.5, with minor GPT-4 comparison) against systematic-review tasks: review-question formulation, eligibility criteria, PubMed search strategy, meta-analysis code outline, and summarisation of three abstracts.

Why it matters here

The earliest widely cited commentary taking the position that LLMs are not yet a drop-in replacement for any SLR stage and will need content experts as an always-on safety net. A grounding counterweight to over-optimistic framings of concepts/llm-assisted-literature-review. Uses Kitchenham-style SR phases implicitly (via PICO, eligibility, search, extraction, synthesis).

Findings

Where ChatGPT was acceptable as a starter:

Formulating a structured review question and contextualising PICO.
Drafting eligibility criteria (needs refinement).
Screening three abstracts for relevance (promising but with errors).
Outlining Python/R code for a meta-analysis (needs SE expertise to debug).

Where ChatGPT failed:

PubMed search strategy — fabricated MeSH/controlled-vocabulary terms that would silently break recall.
Factual referencing — cannot verify sources; hallucinates citations.
Synthesis / summary of multiple studies — the most essential product of an SR, and the one with the least trustworthy output.

The "uncanny valley" diagnosis

From a distance the output mimics expert SR writing; on inspection it is not expertly formed. Reviewers without domain expertise cannot tell the difference — which inverts the value proposition of automation.

Connections