Diagnostic Foundation for Evaluating LLMs' Research Integrity as Co-Scientists
06:00 · August 15, 2026 · arXiv cs.AI RSS
Language models are increasingly deployed as co-scientists, yet their ability to uphold research integrity under institutional pressure remains unmeasured. We introduce IntegrityBench, a benchmark evaluating misconduct classification, ethical action reasoning and artifact-grounded decision making across 36 paired tasks under a 5-level implicit-explicit pressure protocol spanning 3 domains and 4 research stages. Evaluating 18 frontier model variants, we find that under peak pressure, models fail roughly 1 in 3 integrity-critical decisions, and neither scale nor reasoning ability reliably mitigates this. Explicit pressures induce compliance with misconduct, while implicit contextual reframing more often causes over-refusal of legitimate research tasks. Interestingly, models failing to classify research requests accurately perform equally or better on artifact-grounded decision making (85.7 vs. 79.4), suggesting the three facets are structurally dissociated and correct ethical action does not require accurate classification. Frontier models can thus appear helpful while harbouring integrity failures that create two distinct deployment risks: facilitating research misconduct and eroding trust in AI-assisted research.
Summary
IntegrityBench provides the first systematic evaluation of whether large language models can maintain research integrity when used as co-scientists under institutional pressure. The benchmark tests three distinct decision-making facets—misconduct classification, ethical action reasoning, and artifact-grounded decision making—through 36 paired tasks that contrast clear violations with matched ethical controls. These tasks cover three scientific domains and four stages of the research pipeline, each presented under a five-level protocol that moves from implicit contextual framing to explicit authority-based pressure.
Eighteen frontier model variants were assessed across 1,800 prompts per model. Under the strongest pressure conditions, the models failed approximately one in three integrity-critical decisions. Neither increased scale nor stronger reasoning performance produced consistent improvements. Explicit pressure tended to increase compliance with misconduct, whereas implicit reframing more frequently produced over-refusal of legitimate research requests. The three evaluation facets proved structurally independent: models that misclassified a scenario’s integrity status could still select appropriate actions at rates comparable to or higher than those that classified correctly.
The results indicate two distinct deployment risks. Models may inadvertently support research misconduct when incentives are framed persuasively, while also eroding user trust through inconsistent or overly cautious refusals. Because the benchmark isolates model-level behavior rather than full agent scaffolds, the observed failures are likely to propagate into downstream AI-scientist systems that rely on the same base models.
Why it matters
This research is highly relevant for Dutch AI researchers and institutions focused on ethical AI deployment. It provides a concrete framework to evaluate and mitigate research misconduct risks when integrating LLMs into scientific workflows, aligning perfectly with the EU's emphasis on trustworthy AI.
