Alignment techniques work by training and deploying models differently

The article presents a conceptual framework for understanding a class of AI alignment techniques. Rather than treating steering vectors, inoculation prompting, and post-hoc honesty fine-tuning as separate methods, the author groups them under a single principle: train-deploy mismatch. This unifying lens is valuable because it reveals a shared structural constraint that each method must navigate. The core challenge in AI alignment — the difficulty of specifying what we want — makes researchers rely on proxy training signals (labels and reward functions on chosen data distributions) and hope the trained model will perform as intended in the real world. The train-deploy mismatch pattern appears to be one response to this challenge: deliberately create a gap between how the model is trained and how it is deployed, accepting a tradeoff between how well the training data matches the actual deployment task and how effective the alignment technique proves to be. Understanding this pattern may help future work identify which techniques are appropriate for which scenarios.