Human-centred mechanism design with Democratic AI. Koster, R., Jan, B., Tacchetti, A., Weinstein, A., Zhu, T., Hauser, O., Williams, D., Campbell-Gillingham, L., Thacker, P., Botvinick, M., & Summerfield, C. Nature Human Behaviour, Nature Publishing Group, July, 2022. ![Human-centred mechanism design with Democratic AI [link]](https://bibbase.org/img/filetypes/link.svg "link")
Paper doi abstract bibtex Building artificial intelligence (AI) that aligns with human values is an unsolved problem. Here we developed a human-in-the-loop research pipeline called Democratic AI, in which reinforcement learning is used to design a social mechanism that humans prefer by majority. A large group of humans played an online investment game that involved deciding whether to keep a monetary endowment or to share it with others for collective benefit. Shared revenue was returned to players under two different redistribution mechanisms, one designed by the AI and the other by humans. The AI discovered a mechanism that redressed initial wealth imbalance, sanctioned free riders and successfully won the majority vote. By optimizing for human preferences, Democratic AI offers a proof of concept for value-aligned policy innovation.
@article{koster_human-centred_2022,
title = {Human-centred mechanism design with {Democratic} {AI}},
copyright = {2022 The Author(s)},
issn = {2397-3374},
url = {https://www.nature.com/articles/s41562-022-01383-x},
doi = {10.1038/s41562-022-01383-x},
abstract = {Building artificial intelligence (AI) that aligns with human values is an unsolved problem. Here we developed a human-in-the-loop research pipeline called Democratic AI, in which reinforcement learning is used to design a social mechanism that humans prefer by majority. A large group of humans played an online investment game that involved deciding whether to keep a monetary endowment or to share it with others for collective benefit. Shared revenue was returned to players under two different redistribution mechanisms, one designed by the AI and the other by humans. The AI discovered a mechanism that redressed initial wealth imbalance, sanctioned free riders and successfully won the majority vote. By optimizing for human preferences, Democratic AI offers a proof of concept for value-aligned policy innovation.},
language = {en},
urldate = {2022-07-26},
journal = {Nature Human Behaviour},
publisher = {Nature Publishing Group},
author = {Koster, Raphael and Jan, Balaguer and Tacchetti, Andrea and Weinstein, Ari and Zhu, Tina and Hauser, Oliver and Williams, Duncan and Campbell-Gillingham, Lucy and Thacker, Phoebe and Botvinick, Matthew and Summerfield, Christopher},
month = jul,
year = {2022},
keywords = {Economics, Science, technology and society},
pages = {1--10},
}
Downloads: 0
{"_id":"72tY24vL2Hgm4wbSH","bibbaseid":"koster-jan-tacchetti-weinstein-zhu-hauser-williams-campbellgillingham-etal-humancentredmechanismdesignwithdemocraticai-2022","author_short":["Koster, R.","Jan, B.","Tacchetti, A.","Weinstein, A.","Zhu, T.","Hauser, O.","Williams, D.","Campbell-Gillingham, L.","Thacker, P.","Botvinick, M.","Summerfield, C."],"bibdata":{"bibtype":"article","type":"article","title":"Human-centred mechanism design with Democratic AI","copyright":"2022 The Author(s)","issn":"2397-3374","url":"https://www.nature.com/articles/s41562-022-01383-x","doi":"10.1038/s41562-022-01383-x","abstract":"Building artificial intelligence (AI) that aligns with human values is an unsolved problem. Here we developed a human-in-the-loop research pipeline called Democratic AI, in which reinforcement learning is used to design a social mechanism that humans prefer by majority. A large group of humans played an online investment game that involved deciding whether to keep a monetary endowment or to share it with others for collective benefit. Shared revenue was returned to players under two different redistribution mechanisms, one designed by the AI and the other by humans. The AI discovered a mechanism that redressed initial wealth imbalance, sanctioned free riders and successfully won the majority vote. By optimizing for human preferences, Democratic AI offers a proof of concept for value-aligned policy innovation.","language":"en","urldate":"2022-07-26","journal":"Nature Human Behaviour","publisher":"Nature Publishing Group","author":[{"propositions":[],"lastnames":["Koster"],"firstnames":["Raphael"],"suffixes":[]},{"propositions":[],"lastnames":["Jan"],"firstnames":["Balaguer"],"suffixes":[]},{"propositions":[],"lastnames":["Tacchetti"],"firstnames":["Andrea"],"suffixes":[]},{"propositions":[],"lastnames":["Weinstein"],"firstnames":["Ari"],"suffixes":[]},{"propositions":[],"lastnames":["Zhu"],"firstnames":["Tina"],"suffixes":[]},{"propositions":[],"lastnames":["Hauser"],"firstnames":["Oliver"],"suffixes":[]},{"propositions":[],"lastnames":["Williams"],"firstnames":["Duncan"],"suffixes":[]},{"propositions":[],"lastnames":["Campbell-Gillingham"],"firstnames":["Lucy"],"suffixes":[]},{"propositions":[],"lastnames":["Thacker"],"firstnames":["Phoebe"],"suffixes":[]},{"propositions":[],"lastnames":["Botvinick"],"firstnames":["Matthew"],"suffixes":[]},{"propositions":[],"lastnames":["Summerfield"],"firstnames":["Christopher"],"suffixes":[]}],"month":"July","year":"2022","keywords":"Economics, Science, technology and society","pages":"1–10","bibtex":"@article{koster_human-centred_2022,\n\ttitle = {Human-centred mechanism design with {Democratic} {AI}},\n\tcopyright = {2022 The Author(s)},\n\tissn = {2397-3374},\n\turl = {https://www.nature.com/articles/s41562-022-01383-x},\n\tdoi = {10.1038/s41562-022-01383-x},\n\tabstract = {Building artificial intelligence (AI) that aligns with human values is an unsolved problem. Here we developed a human-in-the-loop research pipeline called Democratic AI, in which reinforcement learning is used to design a social mechanism that humans prefer by majority. A large group of humans played an online investment game that involved deciding whether to keep a monetary endowment or to share it with others for collective benefit. Shared revenue was returned to players under two different redistribution mechanisms, one designed by the AI and the other by humans. The AI discovered a mechanism that redressed initial wealth imbalance, sanctioned free riders and successfully won the majority vote. By optimizing for human preferences, Democratic AI offers a proof of concept for value-aligned policy innovation.},\n\tlanguage = {en},\n\turldate = {2022-07-26},\n\tjournal = {Nature Human Behaviour},\n\tpublisher = {Nature Publishing Group},\n\tauthor = {Koster, Raphael and Jan, Balaguer and Tacchetti, Andrea and Weinstein, Ari and Zhu, Tina and Hauser, Oliver and Williams, Duncan and Campbell-Gillingham, Lucy and Thacker, Phoebe and Botvinick, Matthew and Summerfield, Christopher},\n\tmonth = jul,\n\tyear = {2022},\n\tkeywords = {Economics, Science, technology and society},\n\tpages = {1--10},\n}\n\n\n\n\n\n\n\n\n\n\n\n","author_short":["Koster, R.","Jan, B.","Tacchetti, A.","Weinstein, A.","Zhu, T.","Hauser, O.","Williams, D.","Campbell-Gillingham, L.","Thacker, P.","Botvinick, M.","Summerfield, C."],"key":"koster_human-centred_2022","id":"koster_human-centred_2022","bibbaseid":"koster-jan-tacchetti-weinstein-zhu-hauser-williams-campbellgillingham-etal-humancentredmechanismdesignwithdemocraticai-2022","role":"author","urls":{"Paper":"https://www.nature.com/articles/s41562-022-01383-x"},"keyword":["Economics","Science","technology and society"],"metadata":{"authorlinks":{}},"downloads":0,"html":""},"bibtype":"article","biburl":"https://bibbase.org/zotero/warren.pettine","dataSources":["zYjFbyFmZWKpRCD4j"],"keywords":["economics","science","technology and society"],"search_terms":["human","centred","mechanism","design","democratic","koster","jan","tacchetti","weinstein","zhu","hauser","williams","campbell-gillingham","thacker","botvinick","summerfield"],"title":"Human-centred mechanism design with Democratic AI","year":2022}