Teaching AI to handle exceptions: Supervised fine-tuning with human-aligned judgment
Matthew DosSantos DiSorbo, Harang Ju, Sinan Aral, Frederick Chang · 2026
Read Teaching AI to handle exceptions: Supervised fine-tuning with human-aligned judgment by Matthew DosSantos DiSorbo, Harang Ju, Sinan Aral, Frederick Chang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/pmc-13202207.txt.
Description
Matthew DosSantos DiSorbo and co-authors study how AI agents handle exceptions to rules, in a paper published in 2026. Contracts and policies can never cover every case, so human judges bend rules when strict use would be wasteful or unfair. The team found that language models, even strong reasoners, tend to follow policy to the letter, unlike people.
They tried three fixes. Giving the model an ethical frame did not help, and step-by-step reasoning helped only a little. Training on human decisions worked far better, above all when the training included the humans' explanations, not just their verdicts. Trained models even carried the human style to new kinds of cases. The authors say teaching the reasoning, not just the answer, is the key to aligned judgment.
Catalog record
| CREATOR | Matthew DosSantos DiSorbo, Harang Ju, Sinan Aral, Frederick Chang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |