AI · 1h ago
AI users can spot model drift with a simple 'RES' token trick
A developer shares a tip to detect when AI assistants stray from initial instructions: ask the model to start every response with a chosen word like 'RES'. If the word disappears, it signals the AI may have forgotten constraints. The method is inspired by cybersecurity canary tokens but monitors internal model drift rather than external attacks.
Meridian48 take
The trick is clever but limited—it only catches one type of drift and relies on the model's compliance with the instruction itself.
ai-promptingmodel-drift