🧪 Test?View on arXiv
Diagnostic Foundation for Evaluating LLMs' Research Integrity as Co-Scientists
research integritybenchmarkingethical AIlanguage models
2608.12345
Builder Relevance
6h ago70%
Abstract
The paper introduces IntegrityBench, a benchmark for evaluating the research integrity of language models as co-scientists under institutional pressure.
Reality Card
Core Claim
The study reveals that under peak pressure, language models fail approximately 1 in 3 integrity-critical decisions, indicating significant risks in deploying these models in research contexts.
Method / Result
Models fail roughly 1 in 3 integrity-critical decisions under peak pressure.
Limitations
The findings suggest that neither scale nor reasoning ability reliably mitigates integrity failures, raising concerns about the reproducibility of ethical decision-making in AI models.