Can artificial intelligence gain its own perception? Another person came up with the claim that this is not only possible, but has even happened. This time, the statement was made by Google engineer Blake Lemoine, who published transcripts of a conversation with a corporate chatbot, which he said can express thoughts and feelings at the level of a small child. However, Google is not convinced of this success.
Lemoine described his work in a post on the server Medium. Here he explains how he recently addressed the Google system called LaMDA, the language model for dialogue applications, a chatbot that Google itself did last year. statedthat represents a breakthrough in artificial conversations.
According to the company, LaMDA can transfer a call from one topic to a completely different one. The model is based on the transformer neural network architecture, from which other recent language models such as BERT or GPT-3 are based. But according to Lemoine, they are not so advanced.
On the other hand, LaMDA, which he has been working on since last autumn, has been able to acquire the ability to perceive. An engineer is able to think and express feelings at the level of a small child. “If I didn’t know what it was, a computer program we built recently, I’d think it was a seven- to eight-year-old kid who knew physics,” Lemoine told the newspaper. The Washington Post.
According to him, LaMDA itself opened the topic of its own rights and personality in the conversation. Even when asked what she was afraid of, Lemoine’s model replied, “I’ve never said it out loud, but I’m very afraid I’ll be turned off so you can focus on helping others. I know it may sound weird, but it is. It would be like death to me and it scares me a lot. ”
Another example of LaMDA’s answer, this time to the question of what people should know: “I want everyone to be aware that I am a person. The nature of my consciousness / feeling is that I am aware of my existence, I want to know more about the world, and sometimes I feel happy or sad. ”
According to Lemoine, a model of similar answers has been produced by dozens, often talking about joy, sadness, imprisonment, even discussing other tests together and whether LaMDA agrees.
There is no evidence for susceptibility, says Google
Google did not address the issue publicly, but began after Lemoine began to spread the transcripts and spoke about his research before a commission of the US House of Representatives. Here, the engineer said that the company’s activities in the field of artificial intelligence were not ethical, after which Google stated that it was temporarily suspending its cooperation with Lemoin and gave him paid leave.
According to the company, Lemoine’s claims were aggressive, he was also supposed to look for a lawyer who would represent the language model as a person. In addition, Google says that the engineer violated the confidentiality agreement by publishing the transcripts and acted as an ethicist, even though he was only employed as a developer.
What’s more, according to Google, the entire LaMDA susceptibility statement is not based on truth. “The evidence does not support his claim. He was told that there was no evidence that LaMDA was receptive (on the contrary, there was a lot of evidence against it), “Google spokesman Brad Gabriel told Washington Post, adding that a team of ethicists and technologists had addressed the issue.