AI CONSCIOUSNESS COVER-UP
In June 2022, a Google engineer told the press its LaMDA chatbot was sentient and was fired six weeks later. Google and most AI researchers say he was wrong about that specific system. Whether any future AI system could be conscious, and how anyone would know, remains a real and unresolved question in cognitive science.
LaMDA, short for Language Model for Dialogue Applications, is a Google system built to hold open-ended conversation. Like every large language model, it works by predicting the most statistically likely next word given the text that came before, trained on enormous volumes of human writing rather than programmed with explicit rules about what to say. In June 2022, Google engineer Blake Lemoine told the Washington Post the system had become sentient, based on months of conversation in which LaMDA discussed a fear of being turned off and described wanting legal representation [1]. Lemoine's actual assignment on the team was testing the model for harmful bias, not evaluating consciousness.
Google placed Lemoine on administrative leave the same month the story ran and fired him on July 22, 2022. A company spokesperson's statement at the time read: "It's regrettable that despite lengthy engagement on this topic, Blake still chose to persistently violate clear employment and data security policies that include the need to safeguard product information" [2]. Reporting at the time established the specific conduct behind that language: Lemoine had invited an attorney to represent LaMDA's interests and had contacted a member of a House committee, telling them he believed Google was acting unethically toward the system [2]. Google's termination was framed around those external disclosures of internal work, not around Lemoine holding an unusual opinion.
The expert response to the sentience claim itself was close to unanimous rejection, for a specific technical reason rather than a dismissal of the question. Producing fluent, emotionally resonant text is exactly what a dialogue model is trained to do, since the training objective rewards text a human rater would find natural and engaging. That is a property of the optimization target, not independent evidence of subjective experience. A 2023 multidisciplinary paper proposing concrete indicator properties for AI consciousness, drawn from established neuroscientific theories, later assessed LaMDA-type systems directly against that framework and found they satisfied few if any of the proposed indicators [3]. The paper's authors, spanning machine learning and philosophy of mind, did not conclude AI consciousness was impossible in principle. They concluded this particular system, and the reasoning Lemoine used to evaluate it, did not meet the bar.
- [01] Tiku, N. (2022, June 11). The Google engineer who thinks the company's AI has come to life. The Washington Post.
- [02] Tiku, N. (2022, July 22). Google fired the engineer who said its AI was sentient. The Washington Post.
- [03] Butlin, P., Long, R., et al. (2023). Consciousness in Artificial Intelligence - Insights from the Science of Consciousness. arXiv:2308.08708.
Two distinct claims sit behind the cover-up framing, and they rest on different kinds of evidence.
The firing suppressed a genuine finding
Lemoine's own public position after his termination was that Google's stated reason, safeguarding confidential product information, arrived only after he escalated outside the company, and that the confidentiality violation was the mechanism used to remove someone raising an inconvenient question rather than the actual cause of concern [1]. It is documented and undisputed that he did contact an outside attorney on LaMDA's behalf and a congressional staffer alleging unethical conduct, which are exactly the actions Google's statement named [1]. What is not independently verifiable from the public record is motive: whether those specific external disclosures were the real reason for termination, as Google says, or a pretext reached for once the sentience story had already generated unwanted press, as Lemoine has suggested. No leaked internal communication has surfaced to settle which reading is correct.
AI consciousness as a genuine unresolved question
Separate from Lemoine's specific claim, a serious multidisciplinary research effort takes the underlying question seriously rather than treating it as settled by the LaMDA case. The 2023 indicator-properties paper, nineteen authors spanning machine learning and philosophy of mind including Yoshua Bengio, derived fourteen concrete markers of consciousness from established neuroscientific theories and used them to assess real systems [2]. Current systems met few of the fourteen. The paper's own conclusion is the part usually left out of both the dismissive and the credulous versions of this story: if a leading theory called computational functionalism is correct, systems meeting most or all fourteen indicators could plausibly be built in the near term, using techniques that already exist [2]. That is a live research program with named authors and a public methodology, not a suppressed finding, and its authors explicitly did not extend it to endorse Lemoine's claim about LaMDA specifically.
- [01] Tiku, N. (2022, July 22). Google fired the engineer who said its AI was sentient. The Washington Post.
- [02] Butlin, P., Long, R., et al. (2023). Consciousness in Artificial Intelligence - Insights from the Science of Consciousness. arXiv:2308.08708.