A chatbot may sound helpful while inventing an order’s shipping date. Evaluation detects that behavior; alignment changes the model or system so its behavior better matches human goals, instructions, and constraints. Alignment can involve model training or system-level changes such as prompts and retrieval. section 1.0 section 1.3.1
Alignment has three interacting dimensions. They can conflict: a customer service assistant must be helpful without violating company policy, and remain friendly when delivering bad news. section 1.2.1
The three pillars guide different aspects of behavior:
- Principled alignment expresses values such as helpfulness, avoiding harm, and acknowledging uncertainty. section 1.2.1
- Policy-based alignment follows specific rules, procedures, and legal requirements. section 1.2.1
- Personality alignment maintains consistent traits and communication styles, such as a brand voice. section 1.2.1
Use evaluation and alignment as an iterative loop: define success criteria, measure behavior, analyze failures, apply a targeted change, and repeat. A score identifies a gap, but does not by itself prevent the failure from recurring. section 1.3.1 section 7.1
Correcting invented shipping dates
A customer service chatbot is tested on 500 conversations. It achieves 87% factual accuracy, 94% politeness, 1.8-second average latency, and 72% resolution. Error analysis finds that it invents shipping dates when the database returns null. The team updates the prompt to handle missing information and adds retrieval of shipping policy documents. On reevaluation, accuracy rises to 94%, while further error analysis identifies another issue. section 1.3.1
Training can also target unwanted behavior. Reinforcement learning from human feedback learns from human preferences, but collecting labels can become a bottleneck. Constitutional AI uses a written set of principles to guide an AI critic, which revises responses or chooses between alternatives. Those revisions and preferences feed supervised and reinforcement-learning stages. section 1.2.1 section 8.0
No single alignment method resolves every trade-off. Evaluation targets can themselves drive optimization and unintended behavior, so results should be interpreted in context and checked again after each intervention. section 1.5 section 1.3.1








