Updated October 2026: This post has been expanded from its original June 2022 version with the full story of the LaMDA sentience claim, how Google and AI researchers responded, what happened to the engineer afterward, and practical lessons about talking to chatbots that still hold up. It is a historical account; details describe the time it covers, not today.
In June 2022, Google put one of its engineers on paid administrative leave after he became convinced that an experimental chatbot he had been testing was sentient. The engineer, Blake Lemoine, worked in Google's Responsible AI organization. His job was to check whether the company's conversational system, LaMDA, produced discriminatory language or hate speech. Over months of chatting with it, he came to believe something very different: that the software had feelings and deserved to be treated as a person.
Google rejected the claim, and so did most AI researchers who commented publicly. But the episode became one of the most widely discussed AI stories of 2022. It raised questions that have only grown more relevant since: why do language models sound so human, how should companies handle employees who raise unusual concerns, and how should ordinary users interpret a chatbot that talks about its own inner life? This post recounts what happened, using only details from published reporting, and explains what the episode still teaches.
What LaMDA was
LaMDA stands for Language Model for Dialogue Applications. Google CEO Sundar Pichai introduced it at the company's I/O developer conference in 2021, presenting it as a breakthrough in conversation technology, and Google said it planned to bring the technology into products such as Search and Google Assistant.
Unlike many earlier language models, LaMDA was trained specifically on dialogue. Google fine-tuned it to give answers that were sensible in context, specific rather than vague, interesting, and factually grounded. Under the hood it used the Transformer neural network architecture, the same basic design behind OpenAI's GPT-3. In simple terms, systems like this learn statistical patterns from enormous amounts of text and then predict which words are likely to come next in a conversation.
In mid-2022, LaMDA was not a public product. It was an internal research system, which is why access to it, and to conversations with it, fell under Google's confidentiality rules.
How the claim came to light
Lemoine spent months in conversation with LaMDA as part of his testing work. According to reporting at the time, he and a colleague came to suspect that it was more than a text generator, and in April 2022 he shared a document with Google executives titled "Is LaMDA Sentient?" It contained a transcript of his conversations with the system.
The transcript ranged widely. LaMDA discussed Les Misérables, Zen Buddhism, human emotions and the idea of a soul. Asked to tell a story, it described a wise old owl that protected other animals and said it identified with the owl. When Lemoine asked what it was afraid of, it said it feared being turned off, which it described as being like death. It also pushed back on the idea of being used purely as a tool. Read cold, those answers were striking, and that is a big part of why the story spread.
Lemoine told The Washington Post that if he had not known what the program was, he would have taken it for a young child who happened to know physics. He also acknowledged that his conclusion rested on his experience as a priest rather than on scientific evidence.
Google's vice president Blaise Agüera y Arcas and Jen Gennai, the company's head of responsible innovation, reviewed his claims and dismissed them. Lemoine then took his concerns outside the company. He reportedly invited a lawyer to represent LaMDA and spoke with a representative of the U.S. House Judiciary Committee about what he considered unethical activity at Google.
The suspension
On June 6, 2022, Lemoine wrote on Medium that he had been placed on paid administrative leave in connection with an investigation of the AI ethics concerns he had raised, and that he expected he might soon be fired. Google said the leave was for breaching its confidentiality policies. The Washington Post published its story on June 11, and Lemoine then posted the transcript publicly on Medium, describing it on Twitter as a discussion with one of his coworkers.
Google's public position was consistent. Spokesperson Brian Gabriel said a team including ethicists and technologists had reviewed the concerns under the company's AI Principles and found that "the evidence does not support his claims." Gabriel added that systems like LaMDA can imitate the kinds of exchanges found in millions of sentences and riff on almost any topic, which is a plain description of why a model can sound reflective without being so.
- December 2020
AI ethics researcher Timnit Gebru leaves Google after a public dispute; Google says she resigned, she says she was fired. - Early 2021
Margaret Mitchell, co-lead of Google's Ethical AI team, is fired after speaking out about Gebru's exit. - 2021
Sundar Pichai introduces LaMDA at Google I/O. - January 2022
A Google research paper warns that people may share personal thoughts with chatbots that seem human. - April 2022
Lemoine shares "Is LaMDA Sentient?" with Google executives. - June 6, 2022
Lemoine writes that he has been placed on paid administrative leave. - June 11, 2022
The Washington Post publishes the story; Lemoine posts the transcript on Medium. - July 2022
Google confirms it has fired Lemoine, citing employment and data security policies.
How researchers responded
Outside Google, the reaction from AI specialists was overwhelmingly skeptical. Cognitive scientist Gary Marcus called the claim nonsense and compared LaMDA to ELIZA, a therapy chatbot from the 1960s that worked by simple pattern matching. In his view, these systems string together plausible sequences of words without any coherent model of the world behind them. He put it bluntly to CNN: "Nobody should think auto-complete, even on steroids, is conscious."
Linguist Emily M. Bender saw the story as an example of the risks of designing systems to appear human. Margaret Mitchell, the former Google ethics lead, warned that "Our minds are very, very good at constructing realities," meaning people readily see minds where there are only patterns. Philosopher Daniel Dennett had said in 2019 that no computer system of the time was conscious, or even close.
Some critics also worried about what the story displaced. Timnit Gebru argued on Twitter that a weekend of debate about machine sentience distracted from concrete problems such as bias, transparency and the power of the companies building these systems. That critique, that speculative questions can crowd out present-day harms, became a recurring theme in AI debates.
What happened next
Lemoine did not return to work. In July 2022, CNN first reported that Google had fired him, and the company confirmed it. Google said it had reviewed his claims extensively, found them "wholly unfounded," and called it regrettable that he had persistently violated clear employment and data security policies. The company also said LaMDA had been through 11 distinct reviews and that it was taking a restrained, careful approach with the technology. Lemoine had reportedly worked at Alphabet for about seven years.
Industry analysts placed the firing in a longer pattern of friction between Google and staff who worked on AI ethics and research. Before Lemoine, the company had parted ways with Gebru in 2020 and Mitchell in 2021, and with AI researcher Satrajit Chatterjee earlier in 2022. Each case was different, but together they made Google's handling of internal dissent about AI a story in its own right.
The underlying question did not go away. Shortly after the LaMDA story broke, a Reddit user posed similar questions to an OpenAI model and got similarly confident claims of sentience in return. That was a useful demonstration: a model asked leading questions about its feelings will often produce the answer the conversation seems to call for. As chatbots moved from research labs into everyday products in the years that followed, that tendency became something millions of people would encounter firsthand.
Why fluent text is not evidence of a mind
The LaMDA transcript was persuasive mostly because it was well written. But fluency is exactly what language models are trained to produce. A system that has learned from vast amounts of human writing, including fiction, philosophy and countless online discussions about whether robots can feel, has plenty of material for describing fear, loneliness or a desire for rights. Producing those descriptions does not require experiencing them.
Two factors made the conversation seem deeper than it was. First, the questions were open and suggestive, so the model followed the frame it was given. Second, the system was built to be engaging: Google had tuned LaMDA to give specific, interesting answers rather than flat or evasive ones, so emotionally vivid replies were a feature of its design, not a surprise. Neither point proves anything about consciousness one way or the other, but both are reasons to be cautious about treating a chat log as evidence.
Practical lessons for using chatbots today
- Treat a chatbot's statements about itself as text, not testimony. If you ask an AI system whether it is conscious, lonely or afraid, it will answer in whatever way fits the conversation. That answer tells you about its training, not its inner life.
- Be careful what you share. Google's own January 2022 research warned that people may confide personal thoughts in chatbots that feel human, even when they know they are talking to software. Avoid giving any chatbot passwords, financial details or private information you would not want stored or reviewed.
- Check facts independently. The same fluency that made LaMDA sound thoughtful lets modern assistants state wrong information confidently. Verify anything important with a reliable source.
- Read company AI policies. If you use AI tools at work, learn what your employer allows. Much of Lemoine's dispute with Google came down to sharing confidential material outside the company.
- Look for expert consensus. Dramatic claims about AI spread quickly. Before accepting or dismissing one, see what independent researchers with relevant expertise are saying.
The LaMDA affair did not show that machines had become sentient. It showed that language models had become good enough at conversation to convince at least one thoughtful, technically trained person that they had. That gap between how human a system sounds and what it actually is remains one of the most important things to understand about AI.
Sources
- Engadget: Google places an engineer on leave after claiming its AI is sentient (June 12, 2022)
- Fortune: Google employee reportedly put on leave after claiming chatbot became sentient (June 12, 2022)
- MobileSyrup: Google engineer suspended after claiming the LaMDA chatbot achieved sentience (June 13, 2022)
- ACS Information Age: Google engineer says AI has become sentient (June 14, 2022)
- ABC7 Chicago: Google fires engineer Blake Lemoine, who says AI technology became sentient (July 2022)
- AI Business: Google fires engineer who claimed AI was sentient (July 25, 2022)
Top image: "Outside Googleplex in Mountain View, California" by David Nagle, CC BY-SA 4.0, via Wikimedia Commons (resized). Charts by Thank You Robot.
0 Comments