When Google built Smart Reply, the feature that suggests short email responses, its early prototype had a quirk the team found funny and instructive. As the launch blog post recounts, the model kept proposing “I love you” as a reply - to all sorts of messages, appropriate or not.
The reason was statistical. Short, generic responses like “Thanks,” “Sounds good,” and “I love you” are extremely common across email in general, so when the sequence-to-sequence model was uncertain what to say, it leaned on whichever responses were most common overall as a safe bet. The result was a system that confidently told colleagues and strangers that it loved them.
The fix was to stop rewarding responses for merely being common. The team normalized each candidate reply’s likelihood by its overall prior probability, which pushed the model toward responses that fit the specific incoming message rather than ones that were popular everywhere. Only then did the suggestions become genuinely useful.
Why business readers should care: this is a small, early version of a problem that recurs throughout generative AI. Models trained to predict likely text gravitate toward bland, high-frequency output, and turning that tendency into something useful takes deliberate engineering. The same instinct that produced “I love you” in 2015 is a cousin of the safe, generic, sometimes wrong answers that later large language models had to be tuned away from.