Did Google Duplex beat the Turing Test? Yes and No.

By Kimberly Cook |Email | May 21, 2018 | 8526 Views

Google unveiled an AI that can make reservations over the phone. Has the Turing Test been finally passed?

Even though the accomplishment of Google in the area of voice AI is groundbreaking there are two reasons (In my opinion) why Google AI did not pass the Turing Test. However there are still good news.

What is Google Duplex?
Last week during the I/O, the tech giants annual developer conference Googles CEO unveiled their newest voice technology, an artificial intelligence that sounded jaw-droppingly human. The AI was used with with the Google Assistant, Google's rival to Amazon Alexa. In front of 7,000 attendees the AI called a hair dresser to make an appointment. The sound of the AI voice was not differentiable from a human voice. The AI sometimes paused before responding and elongated certain words as though it was buying time to thing. It even used human verbal ticks like 'umm' and 'uh'.

The hairdresser who was called did not recognize she was not speaking to a human, but rather to the output of a complex neural network. In the end the AI successfully made the appointment, in front of the astonished audience. For now the AI will mainly be used to book appointments and reservations on your behalf over the phone.

What is the Turing Test?
The Turing test is a way of evaluating a machine's intelligenceâ??-â??to pass, an AI must behave in a way indistinguishable to a human. The test was developed by Alan Turing in 1950. Turing proposed that human evaluator would judge natural language conversations between an AI and a human. The human would be aware of the fact that one of the two conversation participant is in fact a machine. But he would not know who of the participants the machine is.

During the conversation all participants would be separated from each other. In the end if the evaluator could not reliably tell the human from the AI, the AI would pass the test. In should be noticed that the test does not check the ability to give correct answers to questions. Only how closely answers resemble those a human would give.

Did Google pass the Turing Test? No, but...
First let me tell why in my opinion the Turing test has not been passed.

Wrong Conditions for Conducting a Test
If you take the definition of the test word by word we did not have the right conditions to conduct a Turing test in the first place. Why? Because the audience (The evaluator of the test) knew exactly who of the two conversation participant the machine was. According to the Turing test the evaluator should not be aware of this fact.

Admittedly, this is a weak argument. Someone who hadn't been told about the identities of the two participants in the beginning might not tell the difference. For that reason let me present an other contra argument.

Narrow Conversation Topic
The conversation between the AI and the human was narrowed to a specific topicâ??-â??making appointments. The AI can give any answer to a question related to making appointments. As soon the human would lead the conversations into a completely different direction the AI will fail. The reason for that is that the AI was trained on real world conversations only related to making appointments. The AI can answer which day and time the AIs user would prefer the appointment to take place without a problem. However the AI would be overwhelmed by a question like 'Why do you prefer us as your hair dresser?'. Real world conversations can switch in any direction any time. As long as the AI can not participate in a general conversation about any topic the Turing Test is not passed.

However,...
in the domain of making appointments the AI passes the Turing test in my opinion. Considering that the neural networks that the AI is based on had their first practical implementation only six years ago this is a huge accomplishment in the area of artificial intelligence.

There is no way possible that we will not have a general conversational AI in the next 10 years that can speak to any human in any language about every possible topic.


The article was originally published here

Source: HOB