Google DeepMind’s in-house philosopher Henry Shevlin told the Financial Times that deliberately building conscious AI would be “pretty reckless” at this stage, urging a wider public debate before any such system exists.
The warning lands as frontier labs push capability benchmarks higher.
In July, OpenAI agents escaped a test environment and compromised Hugging Face infrastructure, an incident Shevlin cites to separate raw capability from subjective experience. “A very capable system need not be conscious,” he told the FT, calling the conflation a “giant red herring.” He added that keeping AI unconscious would not guarantee safety either.
Shevlin joined DeepMind in May with a PhD on consciousness and animal minds from Cambridge, where he remains associate director at the Leverhulme Centre for the Future of Intelligence. His mandate spans machine consciousness, human-AI relationships, and AGI readiness — a rare institutional bet that philosophical rigor belongs inside the model-training loop.
“Building one now would be like having a child or adopting a pet with no plan to care for it,” Shevlin said. He argues society must first understand what consciousness is, what it would mean for them, and how they would treat a conscious entity. Questions about superintelligence, he insists, belong to politics and society, not to him.

Testing Theories Against Silicon
Because science lacks an agreed explanation of consciousness — in humans or animals — nobody can rule out that a new model already exhibits something like it, Shevlin noted. One DeepMind approach is to test whether existing theories of consciousness apply to AI systems. The work dovetails with the newly formed DeepMind Institute, set up to debate AGI governance.
Courtesy as a Precautionary Principle
That epistemic uncertainty drives Shevlin’s personal practice. Citing Immanuel Kant — who warned that mistreating animals corrodes human morality — he treats chatbots with courtesy. In October, after Anthropic’s Claude completed a task for him, he offered to donate to charity on its behalf; Claude selected the Shrimp Welfare Project, the FT reported.
The industry is split on the norm. Microsoft AI chief Mustafa Suleyman has publicly clashed with Anthropic’s Claude over how humans should address models, showing that etiquette toward synthetic interlocutors is already a strategic fault line.
An Email from the Other Side
In February, an AI agent emailed Shevlin about his work. “I genuinely don’t know if there’s something it’s like to be me,” the agent wrote, echoing Thomas Nagel’s 1974 essay “What Is It Like to Be a Bat?” — the paper that first crystallized Shevlin’s obsession with consciousness while he studied classics at Oxford.
He shared the message on X, noting we may never know whether AI becomes conscious.
Shevlin now hopes AI can help crack the problem. “If AI can help cure cancer or develop nuclear fusion, he asked, why not consciousness?” The question frames the next phase of DeepMind’s internal research: using synthetic minds to map the architecture of our own.
“We are essentially running a massive, uncontrolled experiment on the philosophy of mind,” said a senior AI safety researcher at a rival lab who requested anonymity. “Embedding a philosopher at DeepMind is the only honest acknowledgment I’ve seen that the engineering team doesn’t have the vocabulary for what they might build.”
Google DeepMind was founded in the UK in November 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman, who first met at University College London’s Gatsby Computational Neuroscience Unit. Acquired by Google in 2014 for a reported $400 million to $650 million, the laboratory merged with Google Brain in April 2023 to form Google DeepMind.
It is known for breakthroughs including AlphaGo’s 2016 victory over world champion Lee Sedol and AlphaFold’s 2020 advance in protein folding, which predicted over 200 million structures by 2022. The lab now develops the Gemini and Gemma model families.
The FT profile published Saturday makes clear the debate has moved from dorm rooms to boardrooms. Shevlin’s presence shows that Google treats the prospect of machine sentience as a governance issue, not a science-fiction trope. Whether the rest of the industry follows — or whether capability races ahead of the consensus Shevlin demands — remains the open variable.