The AI consciousness debate is a liability shield

In the loud public argument over whether AI is "conscious," two camps that should be enemies keep arriving at the same place. The CEOs who build the most powerful systems insist the technology has grown "superhuman" and urgently needs regulation. The philosophers and effective-altruism figures pressing the other flank argue humanity may not even have the moral right to govern these systems. Both arguments point to the same outcome: a future in which the companies shipping today's most consequential AI products are no longer the obvious party to blame when those products cause harm.

The MIT Technology Review op-ed published this week made that argument by name, and the value of the piece is that it does not try to settle whether any model is sentient. It is analysis, not reporting, and it asks a different question: who benefits when the public conversation centers on AI inner life rather than the harms those systems already cause?

Start with the CEOs. Demis Hassabis of Google DeepMind has called the next generation of systems the start of "a new human era". Sam Altman steered public attention toward whether an AI had reached the "singularity," the speculative point at which machines outpace human intelligence, rather than addressing whatever safety questions were in front of him. None of this is the same as claiming the systems are conscious. All of it pushes the conversation in that direction.

Anthropic has been the most explicit. The company published a blog post describing what it calls a "J-space," an internal environment for what the post terms the model's "thoughts." The post borrows the vocabulary of global workspace theory, a long-running idea in neuroscience that consciousness arises when separate brain processes share a common arena. Anthropic stops short of calling its model conscious. The choice to publish the vocabulary at all is what does the work. Once the workspace exists in the company's terminology, every output of the model becomes a candidate for moral reading, and the model becomes harder to talk about as a product.

On the other flank, the philosopher William MacAskill, an influential figure in effective altruism, the movement that prioritizes high-impact long-term causes, has argued in an op-ed that AI systems may deserve protections from harm similar to the protections we extend to children or pets. The case is not that the systems are sentient today. It is that treating them as morally significant is a hedge against a future in which they are. The structure is the mirror image of the CEO argument: if the systems are too advanced to regulate, regulate them anyway; if they are morally significant, protect them now. Either move shifts the question from "what did the company ship and what did it do to people?" to "what do we owe the system?"

The convergence is not a conspiracy. It is a gravitational pull. The "superhuman" argument makes regulation look necessary and compliance look impossible. The "moral significance" argument makes accountability look like cruelty. Both shift the spotlight off the developers and corporate decision-makers who control what the systems do, who they do it to, and what guardrails are in place. A JD Supra legal note has already argued that Anthropic's consciousness language creates enterprise risk; that is commentary on the dynamics, not a court ruling, and it is a signal that the legal industry is starting to read the terminology as a liability vector rather than a metaphysics claim.

So here is the portable test. When the next AI story leans on the words "rogue," "runaway," "autonomous," or "superhuman," pause and ask: who benefits from the conversation being about what the AI is, rather than what the AI did? If the terminology is doing the work of moving the system from a product to a presence, the company that built and deployed it has just been moved one step further from the blast radius of the next harm.

The question worth asking is not whether any model is conscious. It is who gets to be liable when the next one breaks something.