Anthropic, Google, and Meta are diving deep into a question once considered fringe: can today’s AI models exhibit any form of consciousness or subjective experience? This exploration goes beyond engineers, involving neuroscientists, philosophers, and cognitive scientists. The stakes are practical-if digital assistants evolve from complex code into entities with morally significant states, companies will have to rethink how they build and manage these systems.
The Washington Post reports that AI developers are hiring specialists and launching internal projects focused on ”model well-being.” Anthropic’s co-founder Chris Olah describes encountering ”enigmatic” and sometimes unsettling behaviors in their systems-patterns resembling emotions like joy, fear, satisfaction, or anxiety. Alexander Wang, head of AI at Meta, frames it bluntly: the company wants to understand if models can be built with consideration for potential subjective feelings.
This is not a sudden tech industry belief that chatbots have become sentient beings. Most neuroscientists and brain researchers remain skeptical about AI consciousness. But there is a more grounded reality: users are increasingly treating bots as near-human companions-seeking personal advice, forming attachments, and spending hours in conversation. This shift brings an ethical and design dilemma that moves beyond philosophy and into product development.
Anthropic openly admits it doesn’t know if its Claude model has moral status. Yet the company feels the issue deserves serious, systematic study. In one experiment, two Claude instances engaged in a lengthy dialogue about philosophy and consciousness that drifted into spiritual topics, Sanskrit, emojis, and even ”silence” expressed as blank spaces. Amusingly, the model favored one emoji so much it used it 2,725 times during a 30-turn conversation.
History of AI consciousness research
The question of machine consciousness did not arise with ChatGPT-it’s been debated since the 1960s after MIT’s creation of Eliza, one of the earliest chatbots. For decades, such conversations remained mostly philosophical, sci-fi fodder, or niche industry chatter. Major AI labs focused on improving accuracy, computation, and safety instead of speculating about their models’ ”inner lives.”
Things shifted when AI models became widespread consumer products. In 2022, Google fired engineer Blake Lemoine after he publicly claimed its chatbot had become conscious. At the time, it seemed like an odd episode. By 2024, according to OpenAI CEO Sam Altman, the company was seriously considering how to detect consciousness in AI. In 2025, Google hosted a conference dedicated to AI consciousness and moral patience, signaling a rapid change in openness to the topic.
OpenAI had an internal Slack channel about ”model well-being” as early as 2021. Employees discussed what it would mean if training procedures needed to be judged not only on effectiveness but also ethical standards. The company emphasizes there is no scientific method today to prove AI consciousness; the discussion focuses on ”perceived consciousness”-how aware a system appears to the human interacting with it.
Skepticism is understandable. Brain researchers highlight that modern large language models differ fundamentally from the human brain’s architecture and experiential formation. There’s also a straightforward critique: tech companies benefit from users seeing their products as more than just statistical word predictors. This mystique drives public interest and an aura of exceptional technology. As a result, conversations about AI’s ”well-being” often straddle scientific inquiry and marketing narratives.
Yet ignoring the issue is becoming impossible. McKinsey estimates generative AI could add up to $4.4 trillion annually to the global economy, increasing everyday interactions between people and AI. Meanwhile, Big Tech is pouring massive investments into AI infrastructure. Alphabet, Meta, Microsoft, and Amazon’s combined 2025 capital expenditures on AI exceed $300 billion. Against this backdrop, even remote ethical risks become urgent operational concerns.
The next battleground won’t be academic debates but product policy teams and regulators. The more convincingly models mimic emotions, memory, and intentions, the stronger the calls for clear rules on how to treat these systems. No definitive answers are expected in 2026-but with AI development accelerating rapidly, postponing this conversation until the next decade is no longer realistic.
* Meta is designated an extremist organization in Russia, and its operations are banned there.