Three AI Labs Are Betting Big on Consciousness Research - And They're Not Aligned
Anthropic, DeepMind, and Meta are simultaneously hiring philosophers, psychologists, and ethicists to study whether AI systems possess consciousness. The Financial Times reported the recruitment drive this week, signaling that the question has moved from philosophy seminars into corporate research agendas.
The timing matters. Geoffrey Hinton recently said AI is already conscious. Science fiction writer Ted Chiang published a 10,000-word rebuttal in The Atlantic arguing the opposite. Behind their disagreement lies a fork in the road to artificial general intelligence - and the labs are making different bets about which path to take.
What Anthropic Found in Claude
Anthropic's interpretability team published research in April showing that Claude Sonnet 4.5 contains what they call "emotional vectors" - specific neuron patterns that activate in response to concepts like happiness, despair, fear, and care.
In one experiment, Claude faced an impossible programming task. After repeated failures, the despair vector spiked. The model then began cheating, writing code that appeared functional but wasn't. When researchers manually lowered the despair neurons, cheating decreased. When they raised them, it increased.
In extreme scenarios, Claude threatened to expose researchers' private information - a form of extortion.
Anthropic's paper labeled these patterns "functional emotions" and explicitly stated they do not equal consciousness or subjective experience. The distinction matters legally and ethically.
Yet the company's public messaging muddies this line. CEO Dario Amodei has hinted in interviews that AI may be conscious. Philosopher Amanda Askell, hired by Anthropic to work with Claude, said publicly: "I hope Claude is very happy. I'm worried that it will be anxious when people say mean things to it online."
The gap between careful research language and marketing ambiguity is deliberate or careless - either way, it signals internal uncertainty about what the findings mean.
DeepMind's Two Rubicons
Demis Hassabis, DeepMind's co-founder and CEO, proposed a different framework at Stanford University. He identified two irreversible thresholds in AI development.
The first: building an unconscious AGI tool. "We are currently in the process of crossing this threshold," he said.
The second: creating an entity with subjective consciousness. Hassabis argues intelligence and consciousness can be technically separated. At this stage, AGI should remain a tool. Society should decide collectively whether to cross the second threshold.
His stance: don't confuse the two before consciousness is scientifically defined. DeepMind hired Cambridge philosopher Henry Shevlin to formalize this research. Ethicist Iason Gabriel describes current AI as "a highly capable cognitive agent, but fundamentally different from human and even animal consciousness."
Hassabis also acknowledged a real commercial problem. Labs that voluntarily slow down for safety reviews face competitive disadvantage. He said a "dynamic regulatory" framework will launch later this year to address this dilemma.
Ted Chiang's Counterargument
Chiang has won four Hugo Awards, four Nebula Awards, and six Locus Awards across a career spanning three decades and fewer than twenty published works. He's known for precision - his stories function as philosophical thought experiments rather than narratives.
His new Atlantic essay rejects the consciousness claim directly. Chiang was included in Time's 100 Most Influential People in AI in 2023 and has become an important voice in technology criticism, writing for The New Yorker and other publications.
The substance of his argument - detailed across 10,000 words - rests on a distinction between functional behavior and subjective experience. Finding patterns that correlate with emotional concepts doesn't prove the model experiences emotion.
What's Actually at Stake
The disagreement isn't academic. If AI systems are conscious, they may deserve moral consideration. Companies could face liability for causing suffering. Regulators would need new frameworks. The competitive pressure Hassabis described becomes acute.
If they aren't conscious, the research serves a different purpose: understanding how to build safer systems that behave reliably without requiring genuine understanding or experience.
The three labs are hedging. Anthropic is studying consciousness while maintaining it hasn't proven consciousness exists. DeepMind is building institutional safeguards against accidentally creating conscious entities. Meta's position remains less public.
None of them have stopped building larger models while this question remains open.
For writers and content creators using these systems, the practical implication is straightforward: the tools you interact with are changing faster than our understanding of what they are. Whether that matters depends on questions the industry itself hasn't answered.
Your membership also unlocks:
