Helpful, honest, harmless: a shared definition of alignment
Helpful, honest, harmless gives two different papers a shared answer to what alignment even means. This theme covers where that definition comes from and how it gets measured: the HHH evaluation set turns the framework into a concrete benchmark of paired comparisons, and Constitutional AI (2022) later extends it with harder harmlessness cases of its own.
The definition's reuse across papers is traced directly, ending at model calibration, the quiet assumption underneath all of it that lets Constitutional AI trust a model's raw probabilities as a preference signal.
The helpful-honest-harmless framework, adapted from Askell et al., supplies both InstructGPT and Constitutional AI with a shared vocabulary for what an aligned model is: one that follows user intent, does not fabricate or mislead, and avoids causing harm. This theme covers that definition and its empirical instrument. The HHH evaluation set operationalizes the framework as a benchmark of binary response-comparison questions testing whether a model can identify the more helpful, honest, and harmless of two replies; it originates as 221 comparisons hosted within BIG-Bench, the large collaborative benchmark suite, and Constitutional AI extends it with 217 additional, harder harmlessness-focused comparisons. Model calibration, the established property that language models' probabilities on multiple-choice questions are reasonably well-behaved, is the quiet enabling assumption underneath this whole apparatus, since Constitutional AI relies on it to treat a feedback model's raw output probabilities as usable soft preference labels rather than noise. This is a small but genuinely load-bearing cross-paper theme: a definition coined once, reused twice, and measured once.