Dataset for: Hidden Instructions in Submitted Manuscripts Bias AI-Assisted Peer Review Toward Acceptance: A Controlled Experiment

This dataset contains the complete experimental data for a controlled study examining whether prompt injection attacks embedded in manuscript text can manipulate AI-assisted peer review toward more favourable recommendations. The study exposed 50 published open-access biomedical articles to two large language models under four conditions (baseline, injection at beginning / middle / end of manuscript) and two reviewer prompt styles (standard and injection-aware), yielding 800 peer reviews with structured outcome extraction and psycholinguistic scoring. See accompanying metadata file for additional information.