According to the TechCrunch AI report, the core advancement here is a shift from older models like GPT-3.5 to a mixture of frontier models, which ostensibly allows for better tone mimicry and contextual understanding. The feature's success, however, appears to hinge on a critical, unaddressed variable: user psychology. Even a 40% send rate for drafts, as cited by the co-founder, suggests a majority of AI-generated text still goes unused.
The industry-wide challenge isn't just technical fidelity; it's overcoming the user's internal editor who questions whether an AI draft truly represents their intent. Superhuman's approach of offering multiple variations is a clever workaround, but it arguably highlights the assistant's uncertainty more than its confidence.