Google Duplex Demo: Salon Booking, Disclosure, AI Voice Ethics

A recurrent network, a synthetic voice, and a lot of hmm

Google described Duplex as a recurrent neural network built with TensorFlow Extended and trained on anonymized phone conversation data. At call time it combined several inputs at once: the output of Google's speech recognizer, raw audio features, the history of the conversation so far, and the parameters of the task it was trying to complete. [1]

Speech came from a mix of concatenative and Tacotron and WaveNet based synthesis, and the system was tuned to insert disfluencies such as "hmm" and "uh" because Google's testing found they made the calls sound more familiar. Response timing varied deliberately: simple turns used a fast, lower confidence model to answer in under 100 milliseconds, while harder statements were allowed extra latency. [1]

The published examples were a hair salon booking, a restaurant reservation, and a call to check a business's holiday hours. Google said the system ran autonomously but monitored its own confidence, handing a call to a human operator when a conversation moved outside what it could handle, and that the model itself was refined by having operators guide live calls during training. [1]

Nobody on the other end of the line knew

The salon booking that Google played at its developer conference worked precisely because the person answering the phone treated the call as an ordinary one and never learned otherwise. That was the demo's whole point: a system convincing enough to pass without announcing itself. [1]

Atlas interpretation: That is also what made it uncomfortable to watch. A business owner who agrees to talk to a caller is agreeing to something different from being profiled for a model's training data or nudged through a script optimized to get a yes, and the demo gave the audience no way to tell which one had just happened. [1]

The reaction was fast enough that within two days Google told reporters the concern was fair. The Verge reported that Google committed to having Duplex calls identify themselves as automated to the person on the line, reversing the premise the I/O demo had just been built to show off. [2]

What was a concession in 2018 was a requirement by 2026

Atlas interpretation: Google's fix in May 2018 was a policy Google chose, adopted under public pressure and enforced by nobody but itself. The same principle, that an automated caller or chatbot has to say what it is, later became a legal requirement rather than a courtesy: the EU's AI Act transparency rules, enforced from August 2026, require chatbots to identify themselves in essentially the same way Duplex agreed to eight years earlier. [2]