{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,9,24]],"date-time":"2026-09-24T07:20:04Z","timestamp":1790234404278,"version":"4.1.0"},"reference-count":45,"publisher":"Oxford University Press (OUP)","issue":"5","license":[{"start":{"date-parts":[[2026,5,18]],"date-time":"2026-05-18T00:00:00Z","timestamp":1779062400000},"content-version":"vor","delay-in-days":18,"URL":"https:\/\/creativecommons.org\/licenses\/by-nc\/4.0\/"}],"funder":[{"name":"Initiative on the Digital Economy at the Sloan School of Management at the Massachusetts Institute of Technology"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"published-print":{"date-parts":[[2026,4,30]]},"abstract":"<jats:title>Abstract<\/jats:title>\n                  <jats:p>Large language models (LLMs), initially developed for generative AI, are now evolving into agentic AI systems, which make decisions in complex, real-world contexts. Unfortunately, while their generative capabilities are well-documented, their decision-making processes remain poorly understood. This is particularly evident when testing targeted decision-making: for instance, how models handle exceptions, a critical and challenging aspect of decision-making made relevant by the inherent incompleteness of contracts. Here, we demonstrate that LLMs, even ones that excel at reasoning, deviate significantly from human judgments because they adhere strictly to policies, even when such adherence is impractical, suboptimal, or even counterproductive. We then evaluate three approaches to tuning AI agents to handle exceptions: ethical framework prompting, chain-of-thought (CoT) reasoning, and supervised fine-tuning. We find that while ethical framework prompting fails and CoT prompting provides only slight improvements, supervised fine-tuning\u2014specifically with human explanations\u2014yields markedly better results. Surprisingly, in our experiments, supervised fine-tuning even enabled models to generalize human-like decision-making to novel scenarios, demonstrating transfer learning of human-aligned decision-making across contexts. Furthermore, fine-tuning with explanations, not just labels, was critical for alignment, suggesting that aligning LLMs with human judgment requires explicit training on how decisions are made, not just which decisions are made. These findings highlight the need to address LLMs\u2019 shortcomings in handling exceptions in order to guide the development of agentic AI toward models that can effectively align with human judgment and simultaneously adapt to novel contexts.<\/jats:p>","DOI":"10.1093\/pnasnexus\/pgag144","type":"journal-article","created":{"date-parts":[[2026,5,18]],"date-time":"2026-05-18T08:31:37Z","timestamp":1779093097000},"source":"Crossref","is-referenced-by-count":1,"title":["Teaching AI to handle exceptions: Supervised fine-tuning with human-aligned judgment"],"prefix":"10.1093","volume":"5","author":[{"ORCID":"https:\/\/orcid.org\/0009-0005-9247-5947","authenticated-orcid":false,"given":"Matthew","family":"DosSantos DiSorbo","sequence":"first","affiliation":[{"name":"Harvard Business School , Harvard University, 20 N Harvard Street, Cambridge, MA 02163 ,","place":["USA"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0003-1904-1753","authenticated-orcid":false,"given":"Harang","family":"Ju","sequence":"additional","affiliation":[{"name":"Carey Business School, Johns Hopkins University , 100 International Drive, Baltimore, MD 21202 ,","place":["USA"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-2762-058X","authenticated-orcid":false,"given":"Sinan","family":"Aral","sequence":"additional","affiliation":[{"name":"Sloan School of Management , Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, MA 02139 ,","place":["USA"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"286","published-online":{"date-parts":[[2026,5,18]]},"reference":[{"key":"2026052605144429600_pgag144-B1","author":"Singla","year":"2024"},{"key":"2026052605144429600_pgag144-B2","author":"Bick","year":"2024"},{"key":"2026052605144429600_pgag144-B3","author":"Anthropic","year":"2024"},{"key":"2026052605144429600_pgag144-B4","author":"Coignion"},{"key":"2026052605144429600_pgag144-B5","doi-asserted-by":"publisher","author":"Martin","year":"2024","DOI":"10.48550\/arXiv.2401.16212"},{"key":"2026052605144429600_pgag144-B6","doi-asserted-by":"publisher","author":"Street","year":"2025","DOI":"10.3389\/fnhum.2025.1633272"},{"key":"2026052605144429600_pgag144-B7","doi-asserted-by":"crossref","first-page":"20230254","DOI":"10.1098\/rsta.2023.0254","article-title":"GPT-4 passes the bar exam","volume":"382","author":"Katz","year":"2024","journal-title":"Philos Trans R Soc A"},{"key":"2026052605144429600_pgag144-B8","doi-asserted-by":"publisher","author":"Kasai","year":"2023","DOI":"10.48550\/arXiv.2303.18027"},{"key":"2026052605144429600_pgag144-B9","author":"DeepMind","year":"2025"},{"key":"2026052605144429600_pgag144-B10","doi-asserted-by":"publisher","author":"Masterman","year":"2024","DOI":"10.48550\/arXiv.2404.11584"},{"key":"2026052605144429600_pgag144-B11","author":"Accenture","year":"2024"},{"key":"2026052605144429600_pgag144-B12","first-page":"1","article-title":"A survey on hallucination in large language models: principles, taxonomy, challenges, and open questions","volume":"43","author":"Huang","year":"2025","journal-title":"ACM Trans Inf Syst"},{"key":"2026052605144429600_pgag144-B13","author":"Chen"},{"key":"2026052605144429600_pgag144-B14","doi-asserted-by":"crossref","first-page":"1097","DOI":"10.1162\/coli_a_00524","article-title":"Bias and fairness in large language models: a survey","volume":"50","author":"Gallegos","year":"2024","journal-title":"Comput Linguist"},{"key":"2026052605144429600_pgag144-B15","doi-asserted-by":"publisher","author":"Ji","year":"2024","DOI":"10.48550\/arXiv.2404.17833"},{"key":"2026052605144429600_pgag144-B16","author":"Jia","year":"2024"},{"key":"2026052605144429600_pgag144-B17","unstructured":"Chen Z, Chen J, Gaidhani M, Singh A, Sra M \u00a02023. Xplainllm: a QA explanation dataset for understanding LLM decision-making [preprint], arXiv, arXiv:2311.08614, 10.48550\/arXiv.2311.08614"},{"key":"2026052605144429600_pgag144-B18","doi-asserted-by":"crossref","first-page":"2613","DOI":"10.1038\/s41591-024-03097-1","article-title":"Evaluating and mitigating limitations of large language models in clinical decision making","volume":"30","author":"Hager","year":"2024","journal-title":"Nat Med"},{"key":"2026052605144429600_pgag144-B19","author":"Lamparth"},{"key":"2026052605144429600_pgag144-B20","author":"Jain","year":"2024"},{"key":"2026052605144429600_pgag144-B21","author":"Ryseff","year":"2024"},{"key":"2026052605144429600_pgag144-B22","doi-asserted-by":"publisher","author":"Dijkman","year":"2017","DOI":"10.48550\/arXiv.1706.08255"},{"key":"2026052605144429600_pgag144-B23","author":"Russell"},{"key":"2026052605144429600_pgag144-B24","first-page":"187","article-title":"Incomplete contracts and the theory of contract design","volume":"56","author":"Scott","year":"2005","journal-title":"Case W Res L Rev"},{"key":"2026052605144429600_pgag144-B25","doi-asserted-by":"crossref","first-page":"233","DOI":"10.2307\/2526538","article-title":"Costly contract contingencies","volume":"26","author":"Dye","year":"1985","journal-title":"Int Econ Rev (Philadelphia)"},{"key":"2026052605144429600_pgag144-B26","first-page":"316","article-title":"Markets and hierarchies: some elementary considerations","volume":"63","author":"Williamson","year":"1973","journal-title":"Am Econ Rev"},{"key":"2026052605144429600_pgag144-B27","doi-asserted-by":"crossref","first-page":"432","DOI":"10.2307\/2555872","article-title":"Incomplete contracts and signalling","volume":"23","author":"Spier","year":"1992","journal-title":"Rand J Econ"},{"key":"2026052605144429600_pgag144-B28","doi-asserted-by":"crossref","first-page":"115","DOI":"10.1111\/1467-937X.00080","article-title":"Foundations of incomplete contracts","volume":"66","author":"Hart","year":"1999","journal-title":"Rev Econ Stud"},{"key":"2026052605144429600_pgag144-B29","doi-asserted-by":"publisher","author":"Jiang","year":"2025","DOI":"10.48550\/arXiv.2512.16301"},{"key":"2026052605144429600_pgag144-B30","author":"Rao","year":"2023"},{"key":"2026052605144429600_pgag144-B31","first-page":"24824","article-title":"Chain-of-thought prompting elicits reasoning in large language models","volume":"35","author":"Wei","year":"2022","journal-title":"Adv Neural Inf Process Syst"},{"key":"2026052605144429600_pgag144-B32","author":"Dong","year":"2024"},{"key":"2026052605144429600_pgag144-B33","first-page":"27730","article-title":"Training language models to follow instructions with human feedback","volume":"35","author":"Ouyang","year":"2022","journal-title":"Adv Neural Inf Process Syst"},{"key":"2026052605144429600_pgag144-B34","doi-asserted-by":"crossref","first-page":"eadq2852","DOI":"10.1126\/science.adq2852","article-title":"AI can help humans find common ground in democratic deliberation","volume":"386","author":"Tessler","year":"2024","journal-title":"Science"},{"key":"2026052605144429600_pgag144-B35","doi-asserted-by":"publisher","author":"Hu","year":"2021","DOI":"10.48550\/arXiv.2106.09685"},{"key":"2026052605144429600_pgag144-B36","author":"Lin"},{"key":"2026052605144429600_pgag144-B37","first-page":"28458","article-title":"When to make exceptions: exploring language models as accounts of human moral judgment","volume":"35","author":"Jin","year":"2022","journal-title":"Adv Neural Inf Process Syst"},{"key":"2026052605144429600_pgag144-B38","doi-asserted-by":"publisher","author":"Ju","year":"2025","DOI":"10.48550\/arXiv.2511.13979"},{"key":"2026052605144429600_pgag144-B39","doi-asserted-by":"crossref","first-page":"1378","DOI":"10.1038\/s42256-025-01109-4","article-title":"Aligning generalization between humans and machines","volume":"7","author":"Ilievski","year":"2025","journal-title":"Nat Mach Intell"},{"key":"2026052605144429600_pgag144-B40","doi-asserted-by":"crossref","first-page":"rdaf084","DOI":"10.1093\/restud\/rdaf084","article-title":"Algorithmic recommendations and human discretion","author":"Angelova","year":"2025","journal-title":"Rev Econ Stud"},{"key":"2026052605144429600_pgag144-B41","doi-asserted-by":"crossref","first-page":"538","DOI":"10.1287\/mnsc.2024.05684","article-title":"Roles of artificial intelligence in collaboration with humans: automation, augmentation, and the future of work","volume":"72","author":"F\u00fcgener","year":"2026","journal-title":"Manage Sci"},{"key":"2026052605144429600_pgag144-B42","doi-asserted-by":"publisher","author":"Ju","year":"2025","DOI":"10.48550\/arXiv.2503.18238"},{"key":"2026052605144429600_pgag144-B43","first-page":"80079","article-title":"Jailbroken: how does LLM safety training fail?","volume":"36","author":"Wei","year":"2023","journal-title":"Adv Neural Inf Process Syst"},{"key":"2026052605144429600_pgag144-B44","doi-asserted-by":"crossref","first-page":"2997","DOI":"10.1007\/s13042-024-02432-9","article-title":"Performance evaluations of large language models for customer service","volume":"16","author":"Li","year":"2025","journal-title":"Int J Mach Learn Cybern"},{"key":"2026052605144429600_pgag144-B45","doi-asserted-by":"publisher","author":"Neuman","year":"2025","DOI":"10.48550\/arXiv.2501.08951"}],"container-title":["PNAS Nexus"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/academic.oup.com\/pnasnexus\/advance-article-pdf\/doi\/10.1093\/pnasnexus\/pgag144\/68318444\/pgag144.pdf","content-type":"application\/pdf","content-version":"am","intended-application":"syndication"},{"URL":"https:\/\/academic.oup.com\/pnasnexus\/article-pdf\/5\/5\/pgag144\/68318444\/pgag144.pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"syndication"},{"URL":"https:\/\/academic.oup.com\/pnasnexus\/article-pdf\/5\/5\/pgag144\/68318444\/pgag144.pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,5,26]],"date-time":"2026-05-26T09:15:01Z","timestamp":1779786901000},"score":1,"resource":{"primary":{"URL":"https:\/\/academic.oup.com\/pnasnexus\/article\/doi\/10.1093\/pnasnexus\/pgag144\/8684211"}},"subtitle":[],"editor":[{"given":"Frederick","family":"Chang","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"editor"}]}],"short-title":[],"issued":{"date-parts":[[2026,4,30]]},"references-count":45,"journal-issue":{"issue":"5","published-print":{"date-parts":[[2026,4,30]]}},"URL":"https:\/\/doi.org\/10.1093\/pnasnexus\/pgag144","relation":{},"ISSN":["2752-6542"],"issn-type":[{"value":"2752-6542","type":"electronic"}],"subject":[],"published-other":{"date-parts":[[2026,5]]},"published":{"date-parts":[[2026,4,30]]},"article-number":"pgag144"}}