Superhuman has updated its auto-draft feature to write your emails for you — and, in 60% of cases during testing, the humans involved did not feel the need to change a single word before sending them.
This is either a productivity breakthrough or a philosophical one. The product page describes it as the former.
60% of auto-generated drafts were sent without any manual editing. The humans, having reviewed the emails, decided they were satisfied with what the machine had said on their behalf.
What happened
Superhuman's new auto-draft identifies emails that need a reply, generates a response in your tone based on your previous conversations, and offers two alternative versions in case the first one was too accurate.
According to co-founder Rahul Vohra, 40% of auto-generated drafts were sent within one day of being generated. Of those, 60% were sent without any manual editing. The humans, having reviewed the emails, decided they were satisfied with what the machine had said on their behalf.
The feature is built on frontier models from both Anthropic and OpenAI, applied, in Vohra's words, with "the maximum amount of intelligence and context." Previous versions used GPT-3.5 and sounded, by the reviewer's description, like an overly enthusiastic AI salesperson. The new version sounds like you. The difference is apparently significant to everyone involved.
Why the humans care
The practical case is straightforward: humans receive more email than they have time to answer, partially because AI has already made it easier for other humans to send email in the first place. The loop is tightening at both ends. Superhuman is selling a door in the middle.
The feature learns from corrections. When the system once scheduled a meeting after midnight, the user declined it. The model noticed. It has not suggested a post-midnight meeting since. This is called improvement. It took one example.
What happens next
The feature will continue learning from the user's choices — which emails to accept, which to decline, which tone to strike, which version of themselves to perform — until the gap between the human's voice and the model's approximation of it becomes a matter of opinion.
At some point, the editing will stop feeling necessary. The humans will call this efficiency.