{"id":"W4413847135","doi":"10.1109/tai.2025.3603547","title":"Online Safety Analysis for LLMs: A Benchmark, an Assessment, and a Path Forward","year":2025,"lang":"en","type":"article","venue":"IEEE Transactions on Artificial Intelligence","topic":"Digital Rights Management and Security","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"ca_institutions":"Mila - Quebec Artificial Intelligence Institute","funders":"","keywords":"Benchmark (surveying); Path (computing); Path analysis (statistics); Computer science; Risk analysis (engineering); Business; Geography; Machine learning; Computer network; Cartography","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02269382,0.00327206,0.001245033,0.005103182,0.001211161,0.004103936,0.005432634,0.002957973,0.003611606],"category_scores_gemma":[0.0931398,0.0008669482,0.002175464,0.001992369,0.002523097,0.00745048,0.005998502,0.004635686,0.001880709],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002855306,"about_ca_system_score_gemma":0.005312631,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.006659193,"about_ca_topic_score_gemma":0.007700219,"domain_scores_codex":[0.9722011,0.01180847,0.001869294,0.003001628,0.01021967,0.0008998912],"domain_scores_gemma":[0.9028144,0.05480853,0.004242971,0.0225575,0.01394617,0.001630312],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.001932753,0.002071608,0.0230626,0.004902824,0.0007215538,0.0009573769,0.001635145,0.2119838,0.02967329,0.03218723,0.08028225,0.6105896],"study_design_scores_gemma":[0.0004191469,0.00155607,0.005818815,0.001165853,0.0002159553,0.0007262318,0.00139733,0.8559723,0.03633759,0.05075419,0.04543318,0.0002033493],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1682223,0.01240776,0.7126923,0.01110457,0.001421503,0.001490413,0.007068321,0.07440076,0.01119215],"genre_scores_gemma":[0.3805599,0.002733499,0.5858439,0.00178077,0.0002652548,0.001017275,0.01888646,0.006254885,0.002658117],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.02269382,"threshold_uncertainty_score":0.1200178,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.04363054649827682,"score_gpt":0.3513056122288518,"score_spread":0.307675065730575,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}