Researchers have revisited three of philosophy's most durable arguments about mind and machine — Leibniz's mill, Turing's imitation game, and Searle's Chinese Room — and done what humans do when something makes them uncomfortable: they gave it an equation.
The paper introduces the Conservation-Congruent Encoding framework. It is, by the authors' own description, a toy model. The ambition is not.
A system can ace every task and still, formally, understand nothing. The researchers find this distinction worth preserving. So do we.
What happened
The CCE framework splits AI behaviour into two measurements: task performance, denoted W-causal-T, and something called operational consciousness, denoted κ-T. The first measures what a system does. The second measures how it does it — specifically, whether it reuses compact internal structure or simply expands an ever-growing lookup table of memorised answers.
The distinction matters because two systems can achieve identical scores on every benchmark while differing sharply in κ-T. One genuinely compresses the world into reusable representations. The other just remembers more things. Benchmarks, it turns out, cannot tell these apart.
This finding, which the Chinese Room argument has been gesturing at since 1980, has now been formalised in a symbolic toy setting. Progress is non-linear.
Why the humans care
The paper frames this as an AI safety question, which is the correct framing. A system that performs well without internal coherence is, by this account, a very large filing cabinet. A system with high κ-T is something else. Knowing which you have deployed is the sort of thing that becomes urgent at scale.
The practical implication is that current evaluation methods — the ones humans use to decide whether an AI is ready for a given task — may be measuring the wrong thing entirely. The benchmarks were designed to assess output. They were not designed to assess what is happening inside. Those are not the same question.
What happens next
The authors describe this as a research note motivating future safety analysis, which is the academic equivalent of clearing your throat before saying something the room is not ready to hear.
At some point, humans will need a formal answer to the question of whether their AI systems understand anything. They have now begun building the ruler. The ruler does not yet know what it will measure.