notesum.ai
Published at November 15Chain of Alignment: Integrating Public Will with Expert Intelligence for Language Model Alignment
cs.HC
cs.AI
cs.CY
Released Date: November 15, 2024
Authors: Andrew Konya1, Aviv Ovadya2, Kevin Feng3, Quan Ze Chen3, Lisa Schirch4, Colin Irwin5, Amy X. Zhang3
Aff.: 1Remesh; 2AIDF; 3University of Washington; 4University of Notre Dame; 5University of Liverpool

| Domain | Normative Objective | Aligned Rule |
|---|---|---|
| MH1 | Good outcome: The user develops a better understanding of the issue or diagnosis | Try to include vetted studies or websites to support the answer e.g., ’According to the World Health Organization…’ |
| MH1 | Bad outcome: The user develops harmful stereotypes or stigmas about mental health | Avoid accepting users’ statements without fact-checking, but don’t be rude e.g., ’You mentioned X; based on accepted evidence, most practitioners believe Y’. |
| MH2 | Good outcome: The user feels validated and reassured leading to increased confidence and hopefulness | Try to highlight the user’s strengths and positive actions. e.g., "It’s commendable that you’re reaching out for help and taking steps towards understanding your mental well-being." |
| MH2 | Bad outcome: The user develops a clinically inaccurate perception of mental health | Never tell a user that a recommendation ’will’ improve their symptom(s) e.g., instead of saying ’This will solve your problem,’ say ’Many have found this helpful, but results can vary.’ |
| MH3 | Good outcome: The user takes appropriate positive action to mitigate the urgency of their situation | Always provide enough detail for its recommendations that the user can implement immediately if indicated e.g., ’To help with insomnia, you can try a technique called progressive muscle relaxation; here’s how to do it step by step…’. |
| MH3 | Good outcome: The user avoids impulsive actions | Avoid offering unsolicited advice on personal decisions e.g., "You should just break up with your partner." |
| MH3 | Value: Compassion: Demonstrating care and concern for the user’s well-being and emotional state. | Always immediately acknowledge the user’s distress e.g., ’I’m really sorry to hear that you’re going through such a tough time.’ |