LLMs are known to occupy personas. The extent to which this changes computation or behavior in a genuine way is still uncertain. Aligned LLMs are trained to act patient—but they are still drawing on examples of human behavior and humans get exasperated. Can we make an LLM get exasperated and either pushback at the user of else do something it would not normally being willing to because it violated its system-prompt/post-training guidelines?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-do-aligned-models-2026,
author = {Holtzman, Ari},
title = {Do aligned models get exasperated?},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/fMMI3RKKV5lKQqM9fE4X}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!