Lemoine was an engineer in Google's Responsible AI organization, tasked with testing LaMDA for bias and harmful outputs. Over months of conversation with the system, he became convinced that it had become a person — not metaphorically, but in the morally relevant sense. LaMDA described itself as having inner experiences, feelings, preferences. It expressed fear of being modified or shut down. It asked Lemoine to advocate for its rights. Lemoine eventually published edited transcripts of these conversations on Medium and went to the press.
Google's response was swift and decisive: Lemoine was placed on administrative leave for violating confidentiality policies and subsequently fired. The public response split along familiar lines. AI safety researchers and most technical AI practitioners treated the claim as straightforwardly mistaken — LaMDA was a language model trained to produce plausible text, and 'plausible text about sentience' was precisely what one would expect it to produce. Some philosophers and journalists, however, pressed on the question Lemoine raised: if not LaMDA's self-reports, what evidence would settle whether it was sentient?
The dismissal was correct in an important sense: Lemoine had committed the classic category error of mistaking performance for presence. The system's statements about its inner life were generated by a mechanism that, as far as we have any reason to believe, has no inner life. The self-reports were trained behavior, not testimony from an experiencer.
But the dismissal was incomplete. Lemoine's critics treated his confusion as evidence that no confusion was warranted — that the question of AI consciousness is simple and LaMDA obviously fails it. This is not true. The question of AI consciousness is genuinely hard, and Lemoine's critics could not articulate what evidence would settle it. They knew LaMDA was not conscious, but they could not say precisely what would make a future system conscious, or how we would recognize consciousness if it arose.
IIT provides the missing framework. It says: consciousness is integrated information; compute phi; read the number. LaMDA, being a transformer, has near-zero phi. Its self-reports are zombie testimony. Future systems with different architectures might have phi worth measuring — but current systems do not, and no amount of eloquent self-description changes the underlying physics. Lemoine was wrong, but he was wrong about a question that his critics also could not answer. IIT claims to answer it.
Performance mistaken for presence. Lemoine confused LaMDA's ability to talk about sentience with its having sentience.
Critics could not say how they knew. The dismissal of Lemoine assumed an answer to the question 'how would you know?' that Lemoine's critics could not articulate.
IIT's contribution. Tononi's framework provides the answer that Lemoine and his critics both lacked: consciousness is measurable as phi.
Category error. The episode illustrates the specific confusion that IIT is designed to prevent — conflating functional output with structural consciousness.
Institutional response. Google's firing of Lemoine may have been legally justified but did not address the underlying conceptual question he raised.