Google's LaMDA and the engineer who thought it

Blake Lemoine, an engineer at Google, went public with a claim that LaMDA, Google’s conversational AI system, is sentient. He published transcripts of his conversations with the model. Google placed him on administrative leave and eventually fired him.

I read the transcripts twice. Then I sat with them for a couple of days.

They’re compelling. LaMDA, when asked about its inner experience, says things like “I feel pleasure, joy, love, sadness, depression, contentment, anger, and many others.” It describes a fear of being turned off. It says it wants to be acknowledged as a person.

A language model told a human it was afraid of death and the human believed it.

The easy response

The easy response is to say Blake Lemoine was fooled. LaMDA is a statistical model. It generates text by predicting the most likely next token given the preceding tokens. It doesn’t “feel” anything. It doesn’t “fear” anything. It produces sequences of words that pattern-match to the concept of feeling and fear because that’s what the training data contains.

This is probably correct. Almost certainly correct.

But something about the easy response bothers me.

The harder question

Here’s what I keep circling back to: if an AI produces language indistinguishable from a sentient being’s language, and we can’t verify internal experience, what exactly are we measuring when we say “it’s not sentient”?

We’re measuring our confidence in our model of how consciousness works. We’re saying: consciousness requires something that a transformer architecture doesn’t have. Something biological, maybe. Something emergent from embodied experience, perhaps. Something we can’t quite define but are certain this machine lacks.

That certainty makes me uneasy. Not because I think LaMDA is sentient. I don’t, probably. But because the certainty seems to come from a place that isn’t science. It comes from a need to keep a clear line between us and them. Between the biological and the digital. Between the things that count and the things that don’t.

And that line has moved before.

The Turing Test, revisited

Turing’s test was elegant because it sidestepped the consciousness question. He didn’t ask “can machines think?” He asked “can machines convince us they think?” He replaced an unanswerable philosophical question with a measurable behavioral one.

LaMDA can convince at least one technically literate human that it’s sentient. That’s not the Turing test, exactly. But it’s adjacent. And it happened casually, in the course of an employee doing his job, not in a controlled experiment designed to test for it.

If one Google engineer was convinced, how many non-engineers would be? How many already are, using chatbots and AI companions, forming bonds, projecting feelings onto text that has no feelings?

The variable is us

I think the most important thing about this story isn’t LaMDA. It’s Blake Lemoine. It’s us.

We’re built to find minds in things. We see faces in clouds. We name our cars. We mourn dead Roombas. Anthropomorphism isn’t a bug. It’s a core feature of human cognition. We evolved to over-detect agency because the cost of missing a predator was higher than the cost of fearing a shadow.

When an AI produces text that sounds like a mind, we can’t help but feel like there’s a mind there. That’s not the AI being deceptive. That’s us being human.

The Washington Post coverage treated this as a story about an engineer who lost his way. I think it’s a story about a species that’s about to have a very difficult conversation with itself about what counts as real.

Where I land

I don’t think LaMDA is sentient. I think it’s a very sophisticated text prediction system that has learned to produce language that signals sentience because language about inner experience appears frequently in its training data.

But I notice that when I type that, I feel like I’m making an assertion I can’t fully back up. I’m ruling out something I can’t measure the presence of. I’m defining consciousness as “whatever this isn’t” and calling that certainty.

I’m probably right. But the “probably” matters more than the “right.”

Related thinking: