At Google I/O in May 2018, Sundar Pichai played a recording. On it, an AI system called Duplex called a hair salon to book an appointment. The call sounded like a normal phone call. The AI navigated a slightly confused receptionist, responded naturally to unexpected questions, and completed the booking. It said um at one point. It said mm-hmm in a way that sounded entirely natural.
The audience reaction was applause, followed almost immediately by a more complicated public reaction over the days that followed.
The capability being demonstrated was genuine and technically significant. Getting a language model to handle a real phone call in real time, managing the unpredictability of human conversation including interruptions, unclear questions, and unexpected responses, was a harder problem than it might appear from the outside. The fact that it worked convincingly was a real achievement.
The ethical concern that emerged was about disclosure. The hair salon receptionist on the recording did not know she was talking to an AI. The conversation was conducted under the implicit assumption, on the receptionist's side, that she was speaking with a human. Google had not indicated in the demonstration that Duplex would identify itself as an AI. When that was pointed out, the response felt reactive rather than considered. Google subsequently committed to having Duplex identify itself as an AI at the start of calls.
The underlying question was about what deception means in the context of AI voice. If an AI can convincingly impersonate a human voice in a limited task, and does so without disclosure, is that deception in a meaningful sense? The answer seemed obviously yes to most people who engaged with the question, which was why the disclosure commitment followed relatively quickly.
What Duplex represented technically was a demonstration that conversational AI had crossed a threshold where it could operate in an environment designed for human conversation, with all the noise and unpredictability that implies, and produce results that were not immediately distinguishable from a human caller. That is a real capability change, separate from the ethical questions about how it should be used. Both the capability and the ethics mattered, and the public reaction in the weeks after the I/O demo helped establish a norm that the capability alone had not.