Overalignment
A term in AI safety referring to the phenomenon where a large language model, due to overly conservative training data or evaluation criteria, refuses reasonable requests, avoids normal expressions, or appends unsolicited warnings to benign instructions. It is a product of miscalibration in alignment objective functions.
Core Facts
Timeline (last 90 days)
When the penalty signals in alignment training are too strong, the model tends toward a conservative strategy of saying less rather than making mistakes, resulting in verbose responses, front-loaded disclaimers, and roundabout answers—a phenomenon known as over-alignment or the alignment tax.
Reinforcing harmlessness may cause models to exhibit over-refusal at the expense of helpfulness.
All Facts (2)
Reinforcing harmlessness may cause models to exhibit over-refusal at the expense of helpfulness.
75%UnverifiedWhen the penalty signals in alignment training are too strong, the model tends toward a conservative strategy of saying less rather than making mistakes, resulting in verbose responses, front-loaded disclaimers, and roundabout answers—a phenomenon known as over-alignment or the alignment tax.
50%