A new paper doesn't claim to answer whether AI can be conscious. It claims to measure how confident we should be, and shows that confidence depends almost entirely on which theory you already believe.
The researchers extend Marr's classic three levels of analysis - the framework used to describe systems at different grains, from behavior down to hardware - into five levels: behavioral, computational, intrinsic causal-structural, organismic, and organism-environment. Each major theory of consciousness gets slotted into the level it treats as decisive, and each level gets its own testable indicators. Those indicators feed a Bayesian model that combines a theory's assigned credence with how well a system scores on its indicators, producing a probability rather than a yes-or-no verdict. Run against current large language models, that probability swings from below 0.01 to roughly 0.8 depending entirely on which theoretical assumptions get plugged in.
That spread is the real finding. It isn't that LLMs are secretly conscious or obviously not - it's that the disagreement about AI consciousness has never really been about the AI. It's about which philosophical camp you already trust. The paper also notes that the indicators tracking consciousness overlap heavily with the architecture needed for general intelligence, meaning more capable models may start checking consciousness boxes almost as a side effect of getting smarter.
That's a useful reframe for an argument that usually generates more heat than data. It won't settle anything - the authors call their approach "structured agnosticism," a polite way of saying the scorecard still runs on your priors. But at least now everyone arguing about it has to say which priors those are.