LLMs are known to have a lot of linear structure that controls their output styles and semantics. Is sounding like an AI one of those?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-is-there-a-2026,
author = {Holtzman, Ari},
title = {Is there a 'sounds like AI' direction in the residual stream?},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/TgNhdzCkjiUs99PjVwuk}
}Please sign in to comment on this idea.
One potential follow-up question: To identify the 'sounds like AI' vector, is it a good idea to examine the training dynamics of alignment/RLHF?