On May 8, 2018, Google published its research blog post “Google Duplex: An AI System for Accomplishing Real-World Tasks Over the Phone,” announced on stage by CEO Sundar Pichai at the I/O conference. Duplex used a recurrent neural network to carry on natural-sounding phone conversations and, the post said, complete the majority of tasks such as booking a restaurant table or a hair appointment “fully autonomously, without human involvement.” The demo recordings included realistic speech disfluencies - “um” and “mhm” - that made the assistant hard to distinguish from a person.
The reaction turned quickly from amazement to unease. Critics noted that in the demos the AI did not tell the human on the other end of the line that it was a machine, raising the prospect of routine, convincing deception at scale. The blog post itself stated that “transparency is a key part” of the experience, but said only that Google would be “experimenting with the right approach over the coming months,” and framed transparency mainly toward businesses receiving calls rather than as a clear disclosure to the person answering.
After the backlash, Google said Duplex would identify itself when calling, with an opening such as announcing it was an automated booking service that would record the call. Questions were also raised about how representative the polished demo was of real-world performance.
Duplex is a hype-cycle cautionary tale about the gap between an impressive demo and a deployable, ethical product. The technical achievement was real, but the episode showed that “can the AI fool a human” is the wrong design goal; people generally expect to be told when they are talking to a machine.