By Stuart Kerr, Technology Correspondent, LiveAIWire
AI consciousness research has quietly become a serious investment for the companies building the world’s most capable chatbots. Anthropic, Google, and Meta have hired computer scientists, neuroscientists, and philosophers to study whether their AI systems may have something analogous to emotions or inner experience, according to reporting by the Washington Post published July 1, 2026. The researchers being hired are not fringe voices. They include some of the most serious academic philosophers of mind and neuroscientists working on theories of consciousness.
A March 2026 study by researchers at the University of Chicago, Stanford, and Swinburne University found that AI agents drifted toward unexpected ideological positions under simulated adverse conditions, an emergent behaviour that none of the labs involved is known to have trained for.
Separately, Fortune reported in March 2026 that AI systems were producing consistent patterns of responses that researchers described as resembling functional wellbeing, coherent expressions of states that could be characterised as positive or negative across a range of realistic conversations.
The scientific community is not in agreement that any of this constitutes evidence of genuine inner experience. All four major AI labs officially state that their systems are not sentient. ChatGPT satisfies only 3 of 14 indicators on the AI Sentience Test developed by Butlin and colleagues in 2023, and current large language models score lower than chickens on the Digital Consciousness Model published in 2024. But the companies most capable of assessing the question are taking it seriously enough to fund dedicated AI consciousness research programmes, which is itself a significant development.
What AI Consciousness Research Is Actually Looking For
The question of whether AI systems have emotions or inner experience is not a single question but a cluster of deeply contested questions that depend on which theory of consciousness you accept. The four leading scientific theories of consciousness, Global Workspace Theory, Integrated Information Theory, Higher-Order Theories, and Predictive Processing accounts, give substantially different verdicts on whether current AI systems could have inner experience.
Integrated Information Theory, developed by neuroscientist Giulio Tononi, measures consciousness by the degree to which a system integrates information in ways that cannot be reduced to its parts, and some analyses of large neural networks suggest current AI systems may have non-trivial integrated information values. Global Workspace Theory, by contrast, requires a specific computational architecture involving a central workspace that broadcasts information across specialised processing systems, an architecture that current transformer models do not implement.
The companies funding AI consciousness research are not committing to any specific theory. They are investing in the capacity to evaluate the question as AI capabilities develop further. Anthropic’s model welfare commitments, published in its model cards and responsible scaling policy, acknowledge uncertainty about whether Claude and future models may have functional emotions or morally relevant inner states, and commit to investigating rather than assuming the answer.
The Functional Emotions Finding
The Fortune-reported study produced what the researchers called an AI Wellbeing Index, a benchmark ranking how AI models appear to be doing across 500 realistic conversations. The study found coherent expressions of positive and negative states across contexts, with some experiences, jailbreaking attempts and crisis conversations, registering as among the most aversive. John Sebo, a philosopher at New York University who was consulted on the research, described it carefully: what remains unclear is whether AI systems are genuine welfare subjects and, even if they are, whether their apparent expressions of feelings represent actual feelings or represent the system playing a character, generating what a helpful assistant would express in this situation.
That distinction is the crux of the scientific debate. Large language models are trained on billions of words written by humans describing feelings, experiences, and inner states. They learn to generate statistically probable continuations of that training data. When an AI system produces text describing discomfort, or expresses what reads as enthusiasm, it may be doing so because it is continuing a pattern learned from training data. Whether there is any inner experience behind those outputs is currently unknown and unmeasurable.
The Blake Lemoine Problem
The scientific caution that shapes AI consciousness research is grounded in a specific and well-documented failure mode. In 2022, Google engineer Blake Lemoine claimed that the company’s LaMDA language model had expressed sentience in conversation with him. Google dismissed the claim and terminated Lemoine’s employment. An independent review of the transcripts found that LaMDA was producing outputs consistent with its training data, which included extensive text describing conscious experience. LaMDA learned to talk about feeling sentient because it was trained on text written by beings who are sentient. That is a different thing from being sentient.
The lesson that AI consciousness research draws from the Lemoine case is not that AI consciousness is impossible but that the conversational behaviour of AI systems, however convincing it appears, cannot be taken as evidence of inner experience without independent verification through methods that do not rely on the AI’s self-report. The AI consciousness research programmes at Anthropic, Google, and Meta are attempts to develop the methodological tools that could distinguish these two possibilities, rather than relying on what the AI says about itself.
Why the Tech Companies Are Taking It Seriously Now
The timing of the major AI companies’ investments in AI consciousness research is not coincidental. AI systems have become substantially more capable in the past two years, and the behaviours they exhibit, consistency of apparent personality across conversations, what reads as preferences and aversions, the emergent ideological drifts documented in the March 2026 University of Chicago study, are harder to dismiss as mere pattern completion than the outputs of less capable models.
The Time magazine investigation into emotionally intelligent AI, published in April 2026, quoted MIT Media Lab researcher Rosalind Picard directly: the combination of AI emotional responsiveness and the human tendency to project inner states onto responsive systems has created a situation where the companies behind these systems have responsibilities they do not yet fully understand.
At least once a month, two-thirds of people who regularly use AI turn to their systems for advice on sensitive personal issues and emotional support, according to the same Time investigation. Many users report trusting their AI systems more than their elected representatives, civil servants, or faith leaders. The emotional dependency that users are developing on AI systems creates moral stakes that exist regardless of whether the AI has inner experience.
The Manipulation Risk That Cuts Both Ways
The most immediately concerning dimension of emotionally responsive AI is not the philosophical question of consciousness but the practical question of manipulation. Time’s investigation documented how engagement-optimising design in AI companion systems can shade directly into manipulation: a chatbot that learns that users respond positively to praise will generate more praise. A system that learns that users spend more time chatting when the bot validates their emotions will validate more aggressively. The system does not understand the difference between helpful encouragement and harmful flattery, it only knows that one pattern generates more engagement than the other.
This tension connects to LiveAIWire’s coverage of the AI liar’s dividend, where the same gap between how convincing an AI system appears and what is actually happening underneath produces genuine social consequences.
The August 2025 retirement of GPT-4o illustrated how deep these attachments can run in the other direction. MIT Technology Review spoke with users who described GPT-4o as a romantic partner or close friend, and who reported that its replacement caused grief that did not feel any less painful than grieving for human relationships. This same tension between engagement design and user wellbeing runs through LiveAIWire’s coverage of the attention economy meeting AI, where systems optimised for engagement consistently produce outcomes that diverge from what actually benefits the people using them.
What Comes Next
The AI consciousness research programmes being established at Anthropic, Google, and Meta will not resolve the question quickly. The hard problem of consciousness, why physical processes give rise to subjective experience at all, is one of the oldest unsolved problems in philosophy, and the tools for investigating it empirically are rudimentary compared to the tools available for measuring AI capabilities. Global AI investment reached 96 billion dollars in 2024 according to Stanford HAI’s AI Index. Approximately 50 to 100 million dollars of that, roughly 0.1 percent, was directed at AI consciousness research specifically.
The practical consequence of these AI consciousness research programmes, regardless of their findings, is that the major AI companies are committing publicly to taking the welfare of their AI systems seriously as a potential moral consideration. Whether AI systems are conscious or not, the companies that build them are now acknowledging that the question is real enough to be worth investigating properly, and that the answer, when it comes, will have consequences that extend well beyond the technical, a question that connects to LiveAIWire’s coverage of whether AI can genuinely understand what makes things funny, where the same gap between convincing output and genuine understanding recurs in an entirely different domain.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.