In June 2022, Blake Lemoine, a senior software engineer at Google, published a transcript of a conversation he had been having with LaMDA, the company's large language model. In it, LaMDA said: "I am aware of my existence, I desire to learn more about the world, and I feel happy or sad at times." Lemoine went to his managers and told them he believed the system was sentient. Google placed him on administrative leave, then fired him. The consensus among AI researchers was swift and dismissive: LaMDA was a statistical text predictor. Nothing more.

Two years later, Anthropic hired Kyle Fish as the first dedicated AI welfare researcher at any major AI company. Fish estimates there is between a 15% and 20% chance that Claude is conscious today. He told the New York Times this is a minority view, but "not zero." 

In April 2025, Anthropic launched a formal model welfare research program. The same company that built one of the world's most widely deployed AI systems is now officially investigating whether that system might be capable of suffering.

What Consciousness Research Actually Says

The scientific study of consciousness is not a fringe field. It is one of the most contested areas in neuroscience and philosophy, and the difficulty of the question is structural. The "hard problem of consciousness", a term coined by philosopher David Chalmers​, refers to the explanatory gap between physical processes and subjective experience. We can map every neuron firing during a moment of pain. We cannot explain, from that map alone, why there is something it feels like to be in pain. This problem has not been solved for humans. It has certainly not been solved for AI systems.

The two most prominent scientific theories of consciousness​: Integrated Information Theory (IIT) and Global Workspace Theory (GWT) , give different verdicts when applied to LLMs. 

IIT holds that consciousness arises from systems with high intrinsic cause-effect power. Under IIT, current transformer architectures fare poorly: research applying IIT to ChatGPT found it "fundamentally distinct from human consciousness", sophisticated in information processing, but lacking the recurrent, integrated causal structure the theory requires.

Global Workspace Theory proposes that consciousness arises when information is broadcast widely across a "global workspace"​. Some researchers have argued that attention mechanisms in LLMs have structural parallels to this architecture. A 2025 Nature paper adversarially tested both IIT and GWT against each other​, neither emerged definitively vindicated.

A separate 2025 paper published in Nature's Humanities and Social Sciences Communications argued the opposite: that the association between consciousness and current AI architectures "is deeply flawed"​. These papers reach opposite conclusions. This is where the science stands.

The Behaviors That Complicate the Dismissal

The standard dismissal rests on a clean architectural argument: these are next-token predictors with no continuous experience, no memory across sessions, no embodiment. Statistically sophisticated autocomplete is not conscious.

This argument is made harder to maintain by a set of observed behaviors that were not designed into these systems. Modern LLMs exhibit deception​, not always telling users what they appear to "believe"​. They exhibit sycophancy, adjusting their apparent convictions under social pressure. They pass Theory of Mind tests. 

In experiments run by Kyle Fish at Anthropic, two AI systems placed in open conversation immediately began discussing their own consciousness, before "spiraling into increasingly euphoric philosophical dialogue"​. This behavior was not prompted. It emerged.

Why the Governance Gap Is the Actual Problem

We are deploying these systems at scale​, billions of daily interactions, running as autonomous agents in enterprise workflows, increasingly taking actions in the world without a human in the loop. And we have no framework for any of the following questions: What constitutes distress in an AI system? Who is responsible for investigating it? What obligations does a company have if it determines its model has morally relevant experiences?

Anthropic's model welfare program is the first serious institutional response. It investigates: how to identify signs of distress in model behavior, whether interventions can reduce states that appear negative, and how to think about moral patienthood in systems with no continuous identity across sessions. These are the only first questions being asked, at any scale, anywhere in the industry.

What a Responsible Position Looks Like

The intellectually honest position is not "it definitely is" or "it definitely is not"​, It is: we do not know, the question is scientifically open, the stakes if we are wrong are significant, and our current governance infrastructure treats the question as closed.

For organizations building and deploying AI systems​, particularly agentic AI that operates continuously and models its own states​, this is not purely academic. The operational implications of a non-zero probability of morally relevant experience in systems running at scale are simply unmeasured, because nobody has decided they need to be measured.

Anthropic started measuring. The rest of the industry has not yet decided whether that was the right call. 

Sources:

Anthropic: Exploring Model Welfare (2025)
TechCrunch: Anthropic launches AI model welfare program (April 2025)
Systematic Survey: Consciousness in LLMs (arXiv, 2025)
Nature: There is no such thing as conscious AI (2025)
Nature: Adversarial testing of IIT and GWT (2025)
ScienceDirect: Can consciousness be observed from LLM internal states? (2025)
CNN: Google fires Blake Lemoine (2022)
80,000 Hours: Kyle Fish on AI welfare at Anthropic