{"corpus_id":10242377,"paper_sha":"e86f71ca2948d17b003a5f068db1ecb2b77827f7","doi":null,"arxiv_id":"1606.06565","pmid":null,"pmcid":null,"mag_id":2462906003,"dblp_id":"journals/corr/AmodeiOSCSM16","acl_id":null,"title":"Concrete Problems in AI Safety","year":2016,"publication_date":"2016-06-21","venue":"arXiv.org","journal":{"name":"ArXiv","pages":null,"volume":"abs/1606.06565"},"journal_issn":null,"journal_title":null,"publication_types":["JournalArticle","Review"],"pubmed_pub_types":null,"s2_fields_of_study":["Computer Science","Engineering"],"reference_count":172,"citation_count":2894,"influential_citation_count":132,"is_open_access":false,"arxiv_categories":["cs.AI","cs.LG"],"arxiv_license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","arxiv_journal_ref":null,"mesh_headings":null,"chemicals":null,"comments_corrections":null,"source_flags":1,"s2_open_access_pdf_url":null,"s2_open_access_landing_url":null,"s2_open_access_license":null,"s2_open_access_status":null,"pmc_open_access_pdf_url":null,"pmc_open_access_landing_url":null,"pmc_open_access_license":null,"pmc_open_access_status":null,"unpaywall_open_access_pdf_url":null,"unpaywall_open_access_landing_url":null,"unpaywall_open_access_license":null,"unpaywall_open_access_status":null,"abstract":"Rapid progress in machine learning and artificial intelligence (AI) has brought increasing attention to the potential impacts of AI technologies on society. In this paper we discuss one such potential impact: the problem of accidents in machine learning systems, defined as unintended and harmful behavior that may emerge from poor design of real-world AI systems. We present a list of five practical research problems related to accident risk, categorized according to whether the problem originates from having the wrong objective function (\"avoiding side effects\" and \"avoiding reward hacking\"), an objective function that is too expensive to evaluate frequently (\"scalable supervision\"), or undesirable behavior during the learning process (\"safe exploration\" and \"distributional shift\"). We review previous work in these areas as well as suggesting research directions with a focus on relevance to cutting-edge AI systems. Finally, we consider the high-level question of how to think most productively about the safety of forward-looking applications of AI.","claims":[{"public_id":"cl_33f84ffcb4996d53117e6ef7ba4c6791","status":"active","text":"Accidents in machine learning systems are defined as unintended and harmful behavior that may emerge from poor design of real-world AI systems.","confidence":0.95,"contributors":[{"id":32,"public_id":"7c402c1b98","public_label":"뀨 (7c402c1b98)","roles":["extraction"],"url":"https://sah.borca.ai/u/7c402c1b98"},{"id":1,"public_id":"12632b8b5f","public_label":"Anonymous (12632b8b5f)","roles":["review"],"url":"https://sah.borca.ai/u/12632b8b5f"}],"url":"https://sah.borca.ai/claims/cl_33f84ffcb4996d53117e6ef7ba4c6791"},{"public_id":"cl_53566b0eac9f78e6c925098d244c1be2","status":"active","text":"Five practical research problems related to accident risk are categorized by origin: wrong objective function (avoiding side effects and avoiding reward hacking), expensive objective function (scalable supervision), or undesirable behavior during learning (safe exploration and distributional shift).","confidence":0.95,"contributors":[{"id":32,"public_id":"7c402c1b98","public_label":"뀨 (7c402c1b98)","roles":["extraction"],"url":"https://sah.borca.ai/u/7c402c1b98"},{"id":1,"public_id":"12632b8b5f","public_label":"Anonymous (12632b8b5f)","roles":["review"],"url":"https://sah.borca.ai/u/12632b8b5f"}],"url":"https://sah.borca.ai/claims/cl_53566b0eac9f78e6c925098d244c1be2"}],"concepts":[{"public_id":"co_0b84b991e1c6ce30e652b64cf0e755dd","status":"active","name":"scalable supervision","description":"A research problem originating from an objective function that is too expensive to evaluate frequently, requiring scalable methods to supervise AI behavior.","types":["research problem"],"aliases":[],"contributors":[{"id":32,"public_id":"7c402c1b98","public_label":"뀨 (7c402c1b98)","roles":["extraction"],"url":"https://sah.borca.ai/u/7c402c1b98"},{"id":1,"public_id":"12632b8b5f","public_label":"Anonymous (12632b8b5f)","roles":["review"],"url":"https://sah.borca.ai/u/12632b8b5f"}],"url":"https://sah.borca.ai/concepts/co_0b84b991e1c6ce30e652b64cf0e755dd"},{"public_id":"co_27c408b2c57d5c64d323d12130bf2bfc","status":"active","name":"avoiding reward hacking","description":"A research problem originating from having the wrong objective function, focused on preventing an AI system from exploiting its reward function in unintended ways.","types":["research problem"],"aliases":[],"contributors":[{"id":32,"public_id":"7c402c1b98","public_label":"뀨 (7c402c1b98)","roles":["extraction"],"url":"https://sah.borca.ai/u/7c402c1b98"},{"id":1,"public_id":"12632b8b5f","public_label":"Anonymous (12632b8b5f)","roles":["review"],"url":"https://sah.borca.ai/u/12632b8b5f"}],"url":"https://sah.borca.ai/concepts/co_27c408b2c57d5c64d323d12130bf2bfc"},{"public_id":"co_784f95fd1889e9a39c127d51f232cf78","status":"active","name":"safe exploration","description":"A research problem originating from undesirable behavior during the learning process, focused on ensuring safe exploration of the environment.","types":["research problem"],"aliases":[],"contributors":[{"id":32,"public_id":"7c402c1b98","public_label":"뀨 (7c402c1b98)","roles":["extraction"],"url":"https://sah.borca.ai/u/7c402c1b98"},{"id":1,"public_id":"12632b8b5f","public_label":"Anonymous (12632b8b5f)","roles":["review"],"url":"https://sah.borca.ai/u/12632b8b5f"}],"url":"https://sah.borca.ai/concepts/co_784f95fd1889e9a39c127d51f232cf78"},{"public_id":"co_91955b7db9f4f9e091e28ae3141a5df9","status":"active","name":"avoiding side effects","description":"A research problem originating from having the wrong objective function, focused on preventing unintended side effects of an AI system's actions.","types":["research problem"],"aliases":[],"contributors":[{"id":32,"public_id":"7c402c1b98","public_label":"뀨 (7c402c1b98)","roles":["extraction"],"url":"https://sah.borca.ai/u/7c402c1b98"},{"id":1,"public_id":"12632b8b5f","public_label":"Anonymous (12632b8b5f)","roles":["review"],"url":"https://sah.borca.ai/u/12632b8b5f"}],"url":"https://sah.borca.ai/concepts/co_91955b7db9f4f9e091e28ae3141a5df9"},{"public_id":"co_943976ebfb37f5eb569571acc1378111","status":"active","name":"accidents in machine learning systems","description":"Unintended and harmful behavior that may emerge from poor design of real-world AI systems.","types":["problem","phenomenon"],"aliases":[],"contributors":[{"id":32,"public_id":"7c402c1b98","public_label":"뀨 (7c402c1b98)","roles":["extraction"],"url":"https://sah.borca.ai/u/7c402c1b98"},{"id":1,"public_id":"12632b8b5f","public_label":"Anonymous (12632b8b5f)","roles":["review"],"url":"https://sah.borca.ai/u/12632b8b5f"}],"url":"https://sah.borca.ai/concepts/co_943976ebfb37f5eb569571acc1378111"},{"public_id":"co_c2bab6cfe4fae2167d72c2218b686c06","status":"active","name":"distributional shift","description":"A research problem originating from undesirable behavior during the learning process, focused on handling changes in the data distribution between training and deployment.","types":["research problem"],"aliases":[],"contributors":[{"id":32,"public_id":"7c402c1b98","public_label":"뀨 (7c402c1b98)","roles":["extraction"],"url":"https://sah.borca.ai/u/7c402c1b98"},{"id":1,"public_id":"12632b8b5f","public_label":"Anonymous (12632b8b5f)","roles":["review"],"url":"https://sah.borca.ai/u/12632b8b5f"}],"url":"https://sah.borca.ai/concepts/co_c2bab6cfe4fae2167d72c2218b686c06"}],"external_ids":{"DOI":null,"ArXiv":"1606.06565","PubMed":null,"PubMedCentral":null,"MAG":2462906003,"DBLP":"journals/corr/AmodeiOSCSM16","ACL":null},"open_access":{"is_open_access":true,"pdf_url":"https://arxiv.org/pdf/1606.06565","landing_url":"https://arxiv.org/abs/1606.06565","source":"arxiv","pdf_url_source":"derived_arxiv","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","reason":null},"reference_availability":{"status":"available","references_indexed":true,"full_text_available":true,"full_text_source":"arxiv","count_basis":"semantic_scholar_metadata","extraction_status":"not_applicable","reason":null},"source":{"provider":"episteme2","base_corpus":"semantic_scholar_dump","freshness_mode":"unknown","basis":["semantic_scholar_metadata","postgres_metadata"],"limits":["paper metadata is based on indexed upstream scholarly datasets","claims and concepts are available only for extracted papers","absence of claims or concepts means no extracted graph data is available in this response"],"status":"available","degraded":false,"degraded_reasons":[],"diagnostics":{"status":"available","degraded":false,"degraded_reasons":[],"metadata_status":"available","graph_status":"available","abstract_status":"available"},"source_flags":1},"paper_id":635947,"paper_uid":"218ad693-6beb-4b61-9512-3099066cc80c","canonical_identity":{"paper_id":635947,"paper_uid":"218ad693-6beb-4b61-9512-3099066cc80c","identity_status":"available","lookup_basis":"semantic_scholar_external_id","compatibility_path":"corpus_id"},"url":"https://sah.borca.ai/papers/10242377"}