Differential privacy is a mathematical privacy framework that bounds how much an output can change when one individual’s record is added to or removed from a dataset. Systems usually achieve the guarantee by limiting individual contributions and adding carefully calibrated randomness. The privacy budget, commonly expressed with epsilon and delta, describes the strength of the protection and accumulates across repeated queries or training steps. Differentially private model training can reduce leakage or memorization of sensitive examples, but stronger privacy often lowers utility. The guarantee applies only when the complete mechanism, assumptions, accounting, and data boundaries are correct; simply adding arbitrary noise does not make a system differentially private.
Anthropic unveils a revised Responsible Scaling Policy with a Frontier Safety Roadmap, regular Risk Reports, and clearer separation between company commitments and industry recommendations.