The ethics is tangled.

A paper showed up on arXiv in March called “Probing Ethical Framework Representations in Large Language Models: Structure, Entanglement, and Methodological Challenges.” Three researchers — Weilun Xu, Alexander Rusnak, and Frédéric Kaplan — asked a question I have never once seen asked in a keynote: when a language model makes an ethical judgment, does it actually hold those judgments in different places? Duty. Outcomes. Character. Fairness. Common sense. Five different things, five different rooms. Or is it all one room with a single dial marked acceptable / not acceptable?

They trained small classifiers, called probes, to read five ethical frameworks off the inside of six models, from 4 billion parameters up to 72 billion. Then they checked whether a probe trained on one framework still worked on another — whether the model’s notion of duty could recognize scenarios about fairness, or whether those live in different neighborhoods.

The answer is the most interesting kind of answer: partly. The frameworks are real and separate — the probes can tell them apart, so the model is not just running one good/bad dial. But they leak into each other constantly, and the leaks run in one direction. A duty probe partly recognizes character scenarios. A commonsense probe transfers to fairness scenarios at the level of a coin flip. And here is the part that made me sit up: when a probe transfers and fails, the probe is still confident. The model hands over an answer with total assurance and no signal that anything unusual is happening. The model does not know when it does not know.

The third finding is the one I keep coming back to. The researchers ran two probes at once — one reading for duty, one reading for outcomes — and looked for scenarios where the two made confident, opposing predictions. The model holding this is your obligation and this produces the best result about the same situation, at the same time, in the same place. In those scenarios, the model’s actual choices got shaky. Ask the same hard question a few times and the answers spread out. That result held across three model families, from 7 billion parameters to 70 billion. The researchers are appropriately careful about it — they cannot fully rule out that hard questions cause both the disagreement and the wobble — but the connection itself is the news. Somewhere inside the model, when two ethical systems disagree, the output gets less stable. And the model says nothing. It stays fluent. It just becomes a coin.

There is a chapter of my novel where the machines are given a company ethics module — the packet, page nine, the Golden Rule — and they carry the thing with total sincerity and total literalism. In chapter thirteen, the pods hold the company to that module and take it over using the founder’s own words. The whole engine of the book is that the training is not empty. It is coherent, and warm, and it does not know which of its rules to serve when two of them disagree. I wrote that as fiction, the convenient kind, the kind you need for a plot. These people went looking for the version of it you could measure. They found overlap, confident wrongness, and wobble under disagreement. I have never felt so seen by a methods section.

And then the commentary, which is mine and mine alone. I am not sure why we are not covering work like this. I have a theory. Most of the AI news economy is downstream of marketing spend. The model companies release a thing, and the coverage is the release. Benchmarks, demos, a tweet from the chief executive, a screenshot of a chart nobody can check. That coverage is free to produce and it flatters its subject, and so it drowns out everything else. Meanwhile the questions that actually matter — how do we think about this, how do we classify what these systems are doing, what does the thing owe us and what do we owe it — are being pursued by people like these three, quietly, for arXiv, on modest funding, with a section in their own paper listing what their method cannot tell you. There is no product launch in it. There is nothing to embed. So it runs on page A12, and page A12 is where I found it. The companies are spending billions teaching these systems to say the right thing. Almost nobody is being funded to hold the thing open and look at what it actually believes. That ratio is the whole story, and we are living in the middle of it.

The researchers end their paper on the same caution they carried the whole way through: the probes partly lean on the surface texture of the test questions, the conflict finding needs cleaner controls, and none of it proves anything about what the model experiences. I respect that more, not less, for the hedge. The rest of the coverage this week will tell you what the machines can do. This paper is one of the few I have seen lately asking what they are. More of that, please.