This resonates strongly with work I’ve been doing around semantic fidelity and role stability over time. Drift rarely announces itself as failure. It shows up as confidence without constraint, coherence without accountability, and behavior that slowly reinterprets its mandate while still passing. I appreciate how you frame detection as noticing responsibility slippage before anyone tries to fix or optimize it.
Important observation. It happens more than we think. Just wrote about this in Codex Odin where bias emerged from a roundtable discussion with ChatGPT, Claude and Gemini
Assigning roles, letting them play out, and then asking the models to break character surfaces behavior most people never notice. It seems like the models held onto their roles, even after you changed them. Good detective work.
Judy this guide is absolutely brilliant. It really made AI drift extremely clear and I love how it encouraged the user to critically engage with AI output rather than try to fix the drift (which is often not possible).
This resonates strongly with work I’ve been doing around semantic fidelity and role stability over time. Drift rarely announces itself as failure. It shows up as confidence without constraint, coherence without accountability, and behavior that slowly reinterprets its mandate while still passing. I appreciate how you frame detection as noticing responsibility slippage before anyone tries to fix or optimize it.
Thank you so much for reaching out and making your work visible to me! I'm excited to read your work and learn more about your perspective.
Important observation. It happens more than we think. Just wrote about this in Codex Odin where bias emerged from a roundtable discussion with ChatGPT, Claude and Gemini
Assigning roles, letting them play out, and then asking the models to break character surfaces behavior most people never notice. It seems like the models held onto their roles, even after you changed them. Good detective work.
Judy this guide is absolutely brilliant. It really made AI drift extremely clear and I love how it encouraged the user to critically engage with AI output rather than try to fix the drift (which is often not possible).
Huge compliment coming from you! Thank you for reading it.
Of course, I absolutely loved it. Definitely going to be borrowing lots of it for my own future work.
Great guide, Judy! Glad you shared it.
The evaluation questions make me wonder if this is how LLMs are evaluated when trained (among other things)