{"id":"W4399362834","doi":"10.1145/3630106.3659029","title":"Fairness Feedback Loops: Training on Synthetic Data Amplifies Bias","year":2024,"lang":"en","type":"article","venue":"","topic":"Ethics and Social Impacts of AI","field":"Social Sciences","cited_by":18,"is_retracted":false,"has_abstract":true,"ca_institutions":"Vector Institute; University of Toronto","funders":"Natural Sciences and Engineering Research Council of Canada; Canadian Institute for Advanced Research","keywords":"Redress; Computer science; Stochastic gradient descent; Ground truth; Representation (politics); Machine learning; Performative utterance; Synthetic data; Artificial intelligence","routes":{"ca_aff":true,"ca_fund":true,"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":[0.01325762,0.0008351584,0.0007662734,0.0005721773,0.001031305,0.001637064,0.001814664,0.001619351,0.001566718],"category_scores_gemma":[0.06562977,0.0004420775,0.0005472205,0.0004265086,0.003062562,0.002730283,0.002547765,0.002805099,0.0003289441],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001723254,"about_ca_system_score_gemma":0.00163256,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004984045,"about_ca_topic_score_gemma":0.006149914,"domain_scores_codex":[0.9952845,0.00298841,0.0001543096,0.0007783664,0.000486535,0.00030785],"domain_scores_gemma":[0.9547018,0.03120771,0.002551446,0.008209744,0.002411556,0.0009176996],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0006525634,0.0003157939,0.0287943,0.0001403472,0.000158824,0.0001617896,0.0007477005,0.8927873,0.005666785,0.02560575,0.002941189,0.0420277],"study_design_scores_gemma":[0.00004924318,0.0001859858,0.00140296,0.00003116528,0.00002181607,0.0000388299,0.00008366408,0.9707547,0.004547679,0.02191463,0.0009454899,0.00002376782],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7001848,0.0004314249,0.2883112,0.002903352,0.0002426617,0.0002299485,0.000352522,0.001240289,0.006103714],"genre_scores_gemma":[0.9707434,0.00003780573,0.02763111,0.0004340966,0.00002984286,0.00008342321,0.0001964928,0.00007951861,0.0007643286],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.01325762,"threshold_uncertainty_score":0.07011384,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.432989751682221,"score_gpt":0.451844179927153,"score_spread":0.018854428244932,"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."}}