You do not currently have permission to view the full text for this document. Please sign in to an account with an active subscription or contact support.
In the middle of 2022, the media reported a curious story about goings-on at Google, in which an employee working on AI, Blake Lemoine, was suspended from his job after claiming that the program he was working on, called LaMDA, was sentient. LaMDA, short for Language Model for Dialogue Applications, was Google's most advanced system for building chatbots. A chatbot is a program designed to simulate conversation with human users, which uses artificial intelligence to mimic speech by ingesting trillions of words from the internet, a procedure known as a large language model; Google uses such programs in its search engines, and said they planned to embed it in everything from Search to Google Assistant. Lemoine was engaged in testing if the program employed discriminatory or hate speech. Google was well aware that, in its own words, ‘Models trained on language can propagate … misuse - for instance, by internalizing biases, mirroring hateful speech, or replicating misleading information. And even when the language it's trained on is carefully vetted, the model itself can still be put to ill use.‘1 In the process, Lemoine - and others - increasingly felt they were talking to something intelligent, and part of his offence was to publish a transcript of some of the conversations.2
They certainly make fascinating reading: a chatbot capable of giving plausible responses to quite probing questions about itself comes across almost as a piece of surrealist science fiction. But Lemoine claimed that LaMDA was more than intelligent, it had feelings, so it must be a person. This was too much for Google. They held an internal review and announced that there was no evidence to support the claim that LaMDA was sentient (and lots of evidence against it). No matter that the question of sentience is not reducible to its expression in speech, as if non-human animals didn't count. Their spokesperson explained that such systems ‘imitate the types of exchanges found in millions of sentences,