Evaluation of a large language model (ChatGPT) versus human researchers in assessing risk-of-bias and community engagement levels: a systematic review use-case analysis
Bibliographic record
Abstract
Large language models (LLMs) like OpenAI's ChatGPT (generative pretrained transformers) offer great benefits to systematic review production and quality assessment. A careful assessment and comparison with standard practice is highly needed. Two custom GPTs models were developed to compare a LLM's performance in "Risk-of-bias (ROB)" assessment and "Levels of engagement reached (LOER)" classification vs human judgments. Inter-rater agreement was calculated. ROB GPT classified a slightly higher "low risk" overall judgments (27.8% vs 22.2%) and "some concern" (58.3% vs 52.8%) than the research team, for whom "high risk" judgments were double (25.0% vs 13.9%). The research team classified slightly higher "low risk" total judgments (59.7% vs 55.1%) and almost double "high risk" (11.1% vs 5.6%) compared to "ROB GPT" (55.1%), which rated higher "some concerns" (39.4% vs 29.2%) (P = .366). With regards to LOER analysis, 91.7% vs 25.0% were classified "Collaborate" level, 5.6% vs 61.1% as "Shared leadership", and 2.8% as "Involve" vs 13.9% by researchers, while no studies classified in the first two engagement level vs 8.3% and 13.9%, respectively, by researchers (P = .169). A mixed-effect ordinal logistic regression showed an odds ratio (OR) = 0.97 [95% confidence interval (CI) 0.647-1.446, P = .874] for ROB and an OR = 1.00 (95% CI = 0.397-2.543, P = .992) for LOER compared to researchers. Partial agreement on some judgments was observed. Further evaluation of these promising tools is needed to enable their effective yet reliable introduction in scientific practice.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.632 | 0.751 |
| Meta-epidemiology (narrow) | 0.003 | 0.003 |
| Meta-epidemiology (broad) | 0.008 | 0.023 |
| Bibliometrics | 0.013 | 0.009 |
| Science and technology studies | 0.002 | 0.004 |
| Scholarly communication | 0.008 | 0.009 |
| Open science | 0.005 | 0.008 |
| Research integrity | 0.005 | 0.003 |
| Insufficient payload (model declined to judge) | 0.004 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; the direct Gemma label and the distilled Codex classifier agree on what is shown here.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".