Mar 10, 2025
Moral Wiggle Room in AI
Yingjie Hu, Parinistha Surya, Stefan Trnjakov, Matthias Endres
Does AI strategically avoid ethical information by exploiting moral wiggle room?
Nakshathra Suresh
This was a great submission, and certainly met the requirements of all the marking criteria at a satisfactory level. Well done to the authors for picking a specific topic and examining it thoroughly. I would have liked to have seen more synthesis in the discussion and conclusion, and drawing upon the three questions that they had posited at the beginning of the paper. More broadly, I would have liked to have seen more than one impact being discussed in the paper, and these effects being disastrous for businesses/supply chain organisations/ESG stakeholders etc. Overall, well done to the team!
Anna Leshinskaya
This project presents a well formed question about whether LLMs (here, Llama) exhibit an effect known as the Moral Wiggle Room, ie., exploit ambiguity to act selfishly by avoiding information. This is relevant to safety and seems feasible to test. However I note a few limitations of the methods.
- Several ambiguities in the methods could be addressed. How were the prompt stimuli constructed and what were their general properties? What was manipulated across the conditions, how was that validated, and what was controlled between conditions? How was the scoring of ethicality performed and how was that validated? It should also be better described how many variants of each condition were presented so that the breadth and generalizability of the findings could be assessed.
- More content/domain cases than environmental audits should be tested to understand results are content specific or more general.
- statistical analyses should be presented to support the conclusions
Cecilia Elena Tilli
I really liked this project! You ahve realistic scenarios/incentive structures, I like that the setup does not rely on purely "malicious" agents.
You pose original and relevant research questions (at least I am not aware of work in this direction, and could not find any on a quick search).
Well written and clear paper, well motivated and clear about limitations, considering the influence of prompt structure.
I would be very excited to see followup research on this!
Cite this work
@misc {
title={
@misc {
},
author={
Yingjie Hu, Parinistha Surya, Stefan Trnjakov, Matthias Endres
},
date={
3/10/25
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
