{"id":"W4415300695","doi":"10.1038/s41746-025-02008-z","title":"When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior","year":2025,"lang":"en","type":"article","venue":"npj Digital Medicine","topic":"Topic Modeling","field":"Computer Science","cited_by":52,"is_retracted":false,"has_abstract":true,"ca_institutions":"Artificial Intelligence in Medicine (Canada)","funders":"Google Research; National Center for Advancing Translational Sciences; European Commission; Harvard Catalyst; National Cancer Institute; National Institutes of Health; Harvard University; Patient-Centered Outcomes Research Institute","keywords":"Helpfulness; Vulnerability (computing); Consistency (knowledge bases); Suspect; Risk assessment; Benchmark (surveying); Baseline (sea); Sophistication","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":[0.0006060585,0.0001053924,0.0002654073,0.0001371302,0.00006007484,0.00008081345,0.0006079687,0.00005838003,0.00001671304],"category_scores_gemma":[0.001618671,0.00006139327,0.00003026378,0.0002791077,0.0002002399,0.0006595148,0.0003426172,0.0001329977,0.00001294669],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001637326,"about_ca_system_score_gemma":0.0000772593,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001327986,"about_ca_topic_score_gemma":0.00001107278,"domain_scores_codex":[0.9986818,0.00003974123,0.0004298757,0.0001637675,0.0005312233,0.0001535871],"domain_scores_gemma":[0.9988757,0.0003807034,0.00009682086,0.0004275253,0.0001014563,0.0001177904],"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":[0.0000437114,0.00003920423,0.003276942,0.0000934445,0.00003170869,0.00001393644,0.004446061,0.00002496055,0.00001004242,0.06713468,0.001124657,0.9237607],"study_design_scores_gemma":[0.03703792,0.001827235,0.1705061,0.007823034,0.0007135108,0.0005800953,0.006255245,0.5138574,0.002142129,0.2124201,0.0450846,0.001752569],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.3220237,0.0002366501,0.6641446,0.008768124,0.0005347563,0.0003905976,0.000005501395,0.00005671152,0.003839391],"genre_scores_gemma":[0.9978627,0.00001755085,0.0008495898,0.001061558,0.00005662641,0.00003561003,0.000003037383,0.000002641578,0.0001106909],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9220081,"threshold_uncertainty_score":0.2503545,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.008338394089495093,"score_gpt":0.2471250392254192,"score_spread":0.2387866451359241,"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."}}