{"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":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003124023,0.0001830547,0.0002522785,0.0005099373,0.0003473445,0.0004046147,0.0005263761,0.0000641557,0.0000338907],"category_scores_gemma":[0.00000700123,0.0001656436,0.00019602,0.001384025,0.00008684117,0.0007610102,0.000007839843,0.0001477223,0.000005815684],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006136636,"about_ca_system_score_gemma":0.00007010898,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000554248,"about_ca_topic_score_gemma":0.001280415,"domain_scores_codex":[0.9984851,0.00005224461,0.0004092153,0.0005790569,0.0002021878,0.0002721318],"domain_scores_gemma":[0.9989909,0.0002122229,0.00006823392,0.0005105145,0.0001127728,0.0001053783],"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":[0.00007091895,0.001159123,0.00002885022,0.00002667363,0.000340641,0.000003317542,0.0003507708,0.009535636,0.00009118643,0.3208893,0.00003612695,0.6674674],"study_design_scores_gemma":[0.00006701911,0.0003013537,0.0005099814,0.00002385566,0.0002406668,4.376327e-7,0.0001897507,0.8805674,0.003738596,0.1115757,0.002520428,0.0002647371],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.00670283,0.00001242182,0.9900174,0.0006974326,0.0004029541,0.0003889238,0.00009581237,0.000129759,0.001552443],"genre_scores_gemma":[0.9732093,0.00005104388,0.02571322,0.0003539005,0.00001976066,0.00004993852,0.00001853661,0.000005919591,0.0005783654],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9665065,"threshold_uncertainty_score":0.6754751,"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."}}