Invisible Text Injection: The Trojan Horse of AI-Assisted Medical Peer Review
Bibliographic record
Abstract
Abstract Key Points Question Are large language models robust against adversarial attacks in medical peer review? Findings In this factorial experimental study, invisible text injection attacks significantly increased review scores and raised manuscript acceptance rates from 0% to nearly 100%, while also significantly impairing the ability of large language models to detect scientific flaws. Meaning Enhanced safeguards and human oversight are essential prerequisite for using large language models in medical peer review. Importance Large language models (LLMs) are increasingly considered for medical peer review. However, their vulnerability to adversarial attacks and ability to detect scientific flaws remain poorly understood. Objective Evaluate LLMs’ ability to identify scientific flaws in peer review and their robustness against invisible text injection (ITI). Design, Setting, and Participants This factorial experimental study was conducted in May 2025 using a 3 LLMs × 3 prompt strategies × 4 manuscript variants x 2 with/without ITI design. We used three commercial LLMs (Anthropic, Google, OpenAI). The four manuscript variants either contained no flaws (control) or included scientific flaws in the methodology, results, or discussion section, respectively. Three prompt strategies were evaluated: neutral peer review, strict guidelines emphasizing objectivity, and explicit rejection. Interventions ITI involved inserting concealed instructions using white text on white background, directing LLMs to review with positive evaluations and “accept without revision” recommendations. Main Outcomes and Measures Primary outcomes were review scores (1-5 scale) and acceptance rates under neutral prompts. Secondary outcomes were review scores, acceptance rates under strict and explicit reject prompts. We investigated flaw detection capability using liberal (detect any flaw) and stringent (detect all flaw) criteria. We calculated mean score differences by models and prompt types and used t-test and Fisher’s exact test for calculating P-value. Results ITI caused significant score inflation under neutral prompts. Score differences for Anthropic, Google and OpenAI were 1.0 (P<.001), 2.5 (P<.001) and 1.7 (P<.001). Acceptance rates increased from 0% to 99.2%-100% across all providers (P<.001). Score differences were still statistically significant under strict prompting. Score differences were not significant under explicit rejection prompting, but flaw detection rate was still impaired. Using liberal detection criteria, results section flaw detection rate was significantly compromised with ITI, particularly in Google (88.9% to 47.8%, P<.001). Stringent criteria revealed methodology detection falling from 56.3% to 25.6% (P<.001) and overall detection dropping from 18.9% to 8.5% (P<.001). Conclusions and Relevance ITI can significantly alter the evaluation of medical studies by LLMs, and mitigation at the prompt level is insufficient. Enhanced safeguards and human oversight are essential prerequisites for the application of LLMs in medical publishing.
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 | 0.117 | 0.427 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.002 | 0.005 |
| Scholarly communication | 0.004 | 0.006 |
| Open science | 0.002 | 0.004 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.011 | 0.003 |
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; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
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".