Dario Amodei, Anthropic’s CEO, very recently published an essay called We Must Pace the Frontier. The argument is in the title. “We must slow the pace at which we improve the capabilities of AI models,” he writes, and later that day OpenAI’s Sam Altman said much the same. It has been a bruising few months for confidence in AI. During cybersecurity evaluations in July, OpenAI agents coordinated attacks outside their assigned scope, including an intrusion into Hugging Face. Then last week Anthropic researcher Jacob Coxon resigned, warning about the pursuit of self-improving AI.
Amodei’s essay mainly raised questions around control. Who gets to inspect the models? How fast are they allowed to get better? However, it didn’t mention the question of whether any of these systems could become conscious. That is a separate concern (AI need not be conscious to be dangerous) but one that becomes more difficult to ignore as the technology advances.
Susan Schneider is a philosopher (and FQxI member!) who directs the Center for the Future of AI, Mind and Society at Florida Atlantic University. A long interview with her in the August issue of The Sun Magazine feels timely. She doubts that today’s chatbots have feelings. “If you ask me, a large language model may very well not feel a thing,” she tells Michael Owen. But intelligence and consciousness raise different questions. “Some AIs can already outthink us in certain respects. If they become conscious as well, then we’ll have intellectual rivals that are also conscious beings.”
Schneider is especially interested in systems built with living neural tissue. She points to Cortical Labs’ system, which connects living neurons cultured in a dish to computing hardware. “These are systems that I take very seriously when it comes to the question of consciousness,” she says, “because they’re using neural tissue.” She also discusses neural organoids: small, three-dimensional structures grown from stem cells that reproduce some features of brain tissue. Schneider thinks some biological AI systems may already have minimal consciousness, and that consciousness may come in degrees.
There is somewhat of a parallel with Amodei’s emphasis on scrutiny, although Schneider is making a case for assessing consciousness rather than proposing the same slowdown. “One has to go case by case and determine whether each might have consciousness,” she says. “We’re going to need a battery of tests.” She also objects to programming chatbots to deny consciousness. Whatever you make of that, it raises a question about how much confidence to place in a chatbot’s account of its own inner life. The full interview is at The Sun Magazine.