Alignment training aims to make an artificial-intelligence system’s behavior consistent with its creators’ goals, human values, and relevant constraints. For language models, this means reducing responses that are harmful, inappropriate, unsafe, or legally risky. section 3.20 section 3.5
Safety-oriented alignment work can combine several approaches:
- Post-processing detection algorithms identify problematic outputs after generation. section 3.20
- Content filtering and conditional pretraining limit exposure to or production of selected material. section 3.20
- Safety post-training improves how current models respond to particular queries. section 3.20 section 3.5
- Machine unlearning attempts to remove selected information or associations from models. section 3.20
Alignment remains partly theoretical when discussing superintelligent agents, while practical work addresses present-day model behavior. It does not by itself settle which human values should govern a system, how accountability should work, or whether safeguards prevent privacy leakage and concentrated power. section 3.5 section 11.6
