We discovered that language models leave a natural "signature" on their API outputs that's extremely hard to fake. Here's how it works 🔍
📄 arxiv.org/abs/2510.14086 1/
I had the honor of being the TA for this group of students and helping their class project mature into an accepted FAccT paper with mentorship from @jaspreetranjit_ and @swabhz.
TLDR; even current LLMs exhibit negative bias toward AAVE despite their extensive safety training!
📢 Language models exhibit stereotypical biases toward different English dialects. We find that, despite safety training, comparing texts across dialects side by side AMPLIFIES model bias!
Our FAccT’26 paper:
📄arxiv.org/abs/2605.24384
🔗dill-lab.github.io/dialect_bias_l…
[1/n]
Headed to #ICLR2026 to present this!
Oral: Thurs 10:30 session 15 room 201c
Poster: Thurs 3:15 Pavilion 4 #4003
iclr.cc/virtual/2026/p…
DM me if you want to chat, especially about language model forensics. Also looking for surf buddies! 🌊🏄♂️
We discovered that language models leave a natural "signature" on their API outputs that's extremely hard to fake. Here's how it works 🔍
📄 arxiv.org/abs/2510.14086 1/