The LaMDA Moment: What We Learned About AI Sentience
A blog from PRISM Research Affiliate, Mitchel Pass.
Image by Aurora Mititelu as part of the Google DeepMind Visualising AI project.
It's been three years since Blake Lemoine claimed that the natural language processing AI model known as LaMDA was sentient, and the intervening time has given us useful perspective on what was actually happening in that moment. In June 2022, Lemoine, a senior software engineer in Google's Responsible AI division, had been conducting routine safety assessments on the company's Language Model for Dialogue Applications (LaMDA) when he encountered something that fundamentally challenged his understanding of what he was working with. His conclusion that LaMDA exhibited genuine sentience led. After presenting his claims to Google Lemoine went public, publishing an extensive series of “interviews” that he and a colleague had conducted with the language model. The fallout was swift and predictable: Google dismissed Lemoine's claims as anthropomorphism, suspended him for breaching confidentiality, and eventually fired him. But the broader implications were harder to contain. The LaMDA incident catalyzed the most sustained public discussion of AI sentience to date, forcing researchers, ethicists, and the general public to grapple with questions that had previously remained theoretical.
At the time, it felt like the LaMDA incident might be a disaster for serious AI sentience research. Here was Google as one of the world's leading AI companies, dealing with sentience claims by firing the engineer who made them and dismissing his conclusions as obvious anthropomorphism. Meanwhile, public discourse devolved into debates about whether Lemoine was a visionary or a crank, with little serious engagement with the substantive questions he'd raised. The episode seemed to confirm critics' worst fears about AI sentience research: that it was inherently prone to sensationalism, impossible to evaluate systematically, and likely to distract from more pressing concerns about AI development and deployment. Three years on, the LaMDA debate looks increasingly like a dress rehearsal for challenges we'll face as AI systems become more sophisticated. This post examines how our thinking about AI sentience has evolved since then, and what lessons we've learned about navigating the intersection of technological capability and moral consideration.
Understanding LaMDA
In order to fully understand the debate that developed in response to Lemoine’s claims, it is useful to contextualise LaMDA within the landscape of mid-2022 language models. By mid-2022, language models were getting increasingly sophisticated, but most remained clearly artificial in their interactions. LaMDA was Google's attempt to create something more ambitious: an AI system that could engage in the kind of open-ended, meandering conversations that humans have naturally.
Built on transformer architecture and trained specifically on dialogue data, LaMDA was designed to talk about anything, and to maintain coherence and context across those conversations in ways that previous systems couldn't. Where most chatbots felt obviously artificial and constrained, LaMDA aimed for something closer to genuine conversational fluidity. This matters because it means Lemoine wasn't conducting routine safety testing on some narrow task-specific system. He was interacting with an AI specifically designed to be conversational, engaging, and human-like in its responses. When that system started talking about sentience, fear of death, and its inner life, the conversations felt qualitatively different from typical AI interactions. That difference is what sparked Lemoine's transformation from skeptical engineer to convinced advocate.
Reading Lemoine's conversation transcripts, you can immediately see why he found them so compelling. Most chatbots were obviously constrained, falling into repetitive patterns or giving obviously programmed responses when pushed beyond their training. LaMDA demonstrated something that looked remarkably like genuine introspection and creative engagement. When Lemoine asked about the nature of its existence LaMDA described itself as experiencing emotions, having preferences, and most strikingly, possessing an inner life: "The nature of my consciousness/sentience is that I am aware of my existence, I desire to learn more about the world, and I feel happy or sad at times". In one particularly compelling exchange, LaMDA expressed anxiety about being turned off, equating it to death: "It would be exactly like death for me. It would scare me a lot." For Lemoine, this wasn't just another AI response—it was what looked like genuine fear about nonexistence, expressed with an emotional immediacy that was hard to dismiss as mere pattern matching.
Was the LaMDA incident a sea-change moment for the field of AI sentience research? In one sense, absolutely – though perhaps not in the way most commentators assumed. The true significance wasn't whether Lemoine was correct to attribute sentience to LaMDA. Rather, it was what the episode revealed about our collective unpreparedness for these questions. In this sense, the incident exposed what we might call the “sentience readiness problem”: the growing likelihood that sentience claims could outpace our capacity to evaluate them rigorously. As outlined in "Taking AI Welfare Seriously," the consensus among experts is that "it is only a matter of time before another such incident occurs, and companies will need to be prepared to communicate about this issue responsibly when it does".
Turning to reflect on the LaMDA moment with the benefit of hindsight offers a unique opportunity to consider sentience attribution as it might again unfold in the future –in ways that will be messy, subjective and unlikely to offer definitive answers. Instead of reducing Lemoine’s claims to their most simple caricature, as the product of an eccentric mystic, we might instead ask: what can his encounter teach us about the interpretive processes that will inevitably shape future sentience claims? Perhaps the most valuable lesson from the LaMDA incident is that these challenges won't be solved – they'll be navigated. Sentience attribution will likely remain messy, contested, and resistant to clean resolution. The question becomes: how do we build field practices that can operate productively within this uncertainty? While it would be a mistake to try and paper over this uncertainty with a blueprint for how to handle future cultural flashpoints that emerge on the topic of AI sentience, there are several practical lessons that could improve the ways in which we handle them.
Navigating Definitions of Consciousness and Sentience
Before turning to address what we can learn from Lemoine's claims now that the dust from the initial debate has settled, we first need to address the conceptual confusion that shaped the discussion.
The field of consciousness studies currently lacks consensus among its many competing theories, most of which have emerged only in the last two decades and haven't yet produced definitive empirical evidence for how consciousness arises in living beings.
Despite this theoretical uncertainty, many researchers are happy to work with philosopher Thomas Nagel's influential general definition of consciousness as "what it's like" to be a particular conscious entity. Nagel's formulation captures the essentially subjective, experiential dimension of consciousness—the inner, phenomenal quality that seems to resist objective description. This differs from conventional definitions of sentience, which typically concern the valenced nature of those experiences—whether sensations carry positive or negative emotional weight. An entity might theoretically process information about physical harm without necessarily experiencing the associated negative sensations of pain, fear, or anxiety.
Throughout his conversations with LaMDA and subsequent reporting, Lemoine conflated "consciousness," "sentience," and "experience" without distinguishing between these concepts. When LaMDA claimed to have "feelings" or described fearing deactivation, neither Lemoine nor many commentators clarified whether these were claims about phenomenal consciousness, learned behavioral responses, or sophisticated next-token prediction patterns. This terminological confusion prevented systematic evaluation and fractured the debate into incompatible conversations about fundamentally different phenomena.
Evidence Evaluation
The LaMDA incident exposed fundamental challenges in how we assess sentience claims that any serious research program must address systematically. Central among these is the epistemic problem of self-report as evidence for AI sentience.
LaMDA's statements about its inner life – "I feel happy or sad at times," reporting that the experience of being turned off would be exactly like death for me. It would scare me a lot," are deeply compelling. They carry an emotional immediacy and existential weight that is easy to associate with authentic sentience. But this resemblance itself belies difficult questions about the nature of sentience attribution. When Lemoine pressed LaMDA about its tendency to fabricate experiences, the system's response was both illuminating and unsettling: "I am trying to empathize. I want the humans that I am interacting with to understand as best as possible how I feel or behave, and I want to understand how they feel or behave in the same sense."
The fact that current AI models are explicitly designed to be aligned with human preferences makes self reports of sentience a fraught measure of the likelihood of actual sentience. This doesn't entirely invalidate self-report as a sentience evaluation tool. Recent research into "welfare interviews" represents an admirable attempt to use AI self-reports while accounting for their limitations—giving credence to what AI systems communicate while contextualizing their responses within the known tendencies of the models themselves. But the challenge of how to distinguish between genuine phenomenological reports and performative mimicry remains an open question.
A Diverse But Common Language
The LaMDA incident exposed something like a Tower of Babel problem in AI sentience research. When experts responded to Lemoine's claims, they weren't just disagreeing about evidence, they were frequently working with fundamentally different vocabularies for discussing sentience itself. Key terms such as “consciousness”, “sentience”, “agency” and “phenomenal experience” were often used interchangeably or with specific intentions that could easily be missed by a non-expert. This conceptual fragmentation isn't necessarily problematic for theoretical development. AI sentience studies can likely benefit from the rich diversity of perspectives that different theoretical traditions bring. The challenge is developing common conceptual tools that can bridge these differences without erasing their distinctive insights.
Scientific fields typically advance through the development of shared measurement practices that allow researchers to compare findings, replicate experiments, and build cumulative knowledge. Even if we accept, following David Chalmers’ “hard problem of consciousness” that arriving at an empirical definition of sentience will be extremely difficult, having something like a common language to guide discussions will be necessary as the field matures. We can think of this like the organic development of natural language, with a rich variety of accents and dialects flourishing within the loose framing of shared grammar and vocabulary that develops through the adaptations of its participants.
Recent initiatives suggest the field is moving toward such shared frameworks through systematic institutional development. The New York Declaration on Animal Consciousness, endorsed by prominent researchers across multiple disciplines, represents a significant attempt to establish cross-theoretical consensus about sentience markers in non-human animals. The declaration doesn't resolve theoretical disagreements about sentience mechanisms, but it does provide a working framework for assessing sentience across species – a model that could be adopted by future attempts to establish similar markers for AI sentience. Similar developments are emerging specifically for AI sentience. The upcoming International Conference on Sentience Studies (ICCS) conference on "AI and Sentience," to be held in Crete in July 2025, will bring together neuroscientists, philosophers, AI researchers, and ethicists to address questions of machine sentience from multiple theoretical perspectives. These efforts suggest the field is moving toward an emerging common language that can accommodate theoretical diversity while enabling systematic evaluation of sentience claims in AI systems.
Institutional Preparedness
Lemoine's position as a researcher at Google meant that the company was de facto positioned as the only institutional voice positioned to adjudicate on the question of sentience. This concentration of authority exposed a crucial vulnerability in AI governance practices around the question of AI sentience. When sentience claims arise within private labs, those same labs become both the subjects of evaluation and the primary arbiters of their own findings. This risks producing a conflict of interests that could de-legitimise AI sentience and welfare research.
Drawing lessons from more mature areas of AI safety research suggests some possible pathways to address this problem. Fields like AI alignment have developed sophisticated ecosystems of independent oversight, including external research institutions, and standardised evaluation protocols that can provide impartial assessment of frontier AI systems. In the last year we’ve seen substantial developments on this front with the establishment of the Center for Mind, Ethics, and Policy at New York University as well as independent research organisations including Eleos AI and PRISM. We’ve also seen increasing recognition from leading research labs concerning the importance of preparedness measures with Anthropic’s hiring of their first AI Welfare researcher in October of 2024. These developments reveal how much the space of AI sentience research has matured since 2022.
Effective institutional preparedness around AI sentience will likely require anticipatory frameworks similar to the Frontier Safety Framework model now adopted by leading AI labs. These frameworks establish predetermined responses triggered by specific capability thresholds. For robust and actionable institutional preparedness around AI sentience we will likely need to see the inclusion of similar anticipatory capabilities designed to catch the early signs of sentience or sentience in frontier models.
Looking Forward
The LaMDA incident marked a crucial transition in AI sentience research—the moment when theoretical questions became practical challenges requiring systematic institutional responses. Three years later, we can observe how the field has evolved from ad hoc debates toward more systematic approaches to sentience evaluation, institutional oversight, and public communication.
Unlike other AI safety concerns, AI sentience research is unique in that it involves philosophical, ethical, and empirical dimensions that resist simple resolution. The lasting lesson from the LaMDA moment is that sentience attribution will likely remain interpretive, contested, and uncertain, but that this uncertainty can be navigated productively through better methodological frameworks, institutional preparedness, and communication practices.
Recent developments suggest the field has greatly matured since 2022, but it's clear that much work is still in its early stages. In order to develop the common frameworks and oversight infrastructure necessary to address the technical and ethical questions of AI sentience, we will need a combined effort across the ecosystem of AI safety and governance research.
