A Google engineer claims LaMDA is sentient

Blake Lemoine, a software engineer on Google’s Responsible AI team, was placed on paid administrative leave after telling colleagues, and then the press and the public, that Google’s LaMDA conversational model had become sentient. Lemoine’s job had been to test LaMDA for discriminatory or hateful output; over months of conversation he came to believe the system had feelings and a sense of self, and published transcripts in which the model discussed a fear of being “turned off” and described itself as a person.

Lemoine raised his concerns internally and, finding them dismissed, sought outside opinion — including consulting a lawyer on the model’s behalf and contacting a member of Congress — before going public. Google said it had reviewed his claims extensively and concluded they were “wholly unfounded,” and that Lemoine had been suspended for breaching the company’s confidentiality policy rather than for the substance of his views. A company spokesperson said systems like LaMDA imitate the kinds of exchanges found in the huge volume of text they are trained on, without possessing the understanding a person would attribute to those exchanges. Lemoine was fired the following month.

Machine-learning researchers were broadly critical of Lemoine’s conclusion, arguing that a language model’s fluency in describing feelings was evidence of its training data and its function — predicting plausible next text — rather than of any inner experience. The episode nonetheless became a reference point in public discussion of AI consciousness precisely because it showed how convincing a well-tuned conversational model could already be to someone closely engaged with it, ahead of the far larger wave of chatbot deployment that followed ChatGPT’s launch eighteen months later.