The Ethereal Editor

October 6, 2026 — James Henry

I asked a language model how many planets there are in the solar system. It said nine. It listed them, gave me their diameters in miles, and counted Pallas, an asteroid, as the newest member.

The model is called talkie, and it was trained only on text published before 1931. It sounds exactly like a modern one. I then put the same question to Sonnet and got eight planets, with a note that Pluto was demoted in 2006--no mention of Pallas. Same confidence, different answer, and neither one gave any sign of where it was standing when it answered. (The questions and both models' full replies are here if you'd like to read them.)

Nine versus eight is easy, or even Pallas vs Pluto, if you're so inclined, because you can easily see the discrepancy. Most of what we ask these models isn't like that. And the confidence of an answer tells you nothing about where it was written from, which is the part you can't see.

Everything Has an Editor

None of this is new. A newspaper editor decides which story runs above the fold. A podcaster picks the guests. An author picks which facts make it into the book. A search engine ranks ten links, a feed orders your evening, and a model picks the next word, then the one after that. Every one of them selects, orders and frames, and every one of them leans one way or another.

Some of it isn't even deliberate. A podcaster may not see their own bias. A reporter doesn't control what the desk does to the copy after it's filed. Nobody in that chain is necessarily lying to you.

I wrote last week about what the feed doesn't do, which is keep us looking at the same things. This is about what it, and every other editor, does do. It leans. The question isn't whether, it's whether you can tell which way.

The Masthead

You used to be able to tell which way a paper leaned from the name on the front. The masthead signed the work, and that made the lean readable in four ways. It was signed. It was stable: the paper that leaned one way in 1950 leaned that way in 1951, so you could learn it once. It sat on a newsstand next to the alternatives, so comparison was cheap. And the incentives were roughly knowable: an owner, a party, an advertiser. That's what the masthead gave you. Call it epistemic provenance, if you want a phrase for it: where an answer came from, and what leaned on it.

I'm not nostalgic for it. A slanted paper was still slanted. But you knew, and knowing is most of the fix. We have the same thing nowadays: we can all name certain modern news feeds or channels and what kind of lean you'll get out of them.

A model fails every one of those four tests. It's unsigned. It's ephemeral: the name stays the same while what's behind it changes, and you aren't told when. It hands you one answer with nothing next to it to compare against. And whatever incentives shaped it are sitting in weights nobody outside the lab can open. Ask a search engine and you get ten results to be suspicious of. Ask a model and you get a full paragraph.

How the Lean Gets In

So where does it come from? Here's a simple version of the pipeline, marked by how bias leaks in at that layer.

The corpus. A model trained on what people wrote down inherits who wrote, in which languages, about what, in which decade. It is the floor under everything else. Inherent.
Data selection. What gets dropped as low quality or unsafe, and who decides what those words mean. Partially selected.
Licensing. Whose work is in the data because someone paid for it. Deliberately selected.
Human feedback. Who the raters are, what they were told to reward, and what they rewarded anyway. Partially selected.
Written values. Constitutions, model specs, policies. Deliberately selected.
Law. What a government requires. Required.
The system prompt, and every edit to it. Deliberately selected.
Retrieval and commerce. What the model looks up, cites or sponsors. Deliberately selected, plus whatever the search returns.

Most of that list is weather, not conspiracy. The first and fourth stages in particular have no villain, and I'm not claiming intent where I can't show it. None of it is specific to one lab or one country, and any model you can name goes through some version of these stages.

Stage one is the one talkie makes visible. A model can't be trained on the world. It can only be trained on what got written down, and what got written down came from some particular time and place. Ours did too.

Stage six is the easiest to point at, because in one country it's written into the regulation. China's 2023 Interim Measures for generative AI have been in force since August of that year, and Article 4 says providers must "adhere to the core socialist values." There's nothing hidden about the sentence. What's harder to see is what it does to an answer. One audit put 646 politically sensitive topics to DeepSeek-R1 through its hosted API, in Mandarin, and found that in about one response in nine the model's own reasoning contained terms the final answer left out. That's one study of one service, and I'd want to see it repeated. But it's what a thumb on the scale looks like when someone goes looking.

Stage seven is the one we only learn about when it goes wrong. In May 2025, Grok started bringing up "white genocide" in South Africa in replies about unrelated things. xAI said an "unauthorized modification" had been made to the bot's prompt. That July, Simon Willison watched the new Grok search for Elon Musk's posts before answering a question about the Israel-Palestine conflict. xAI said it had been trying to align itself with the company, and changed the system prompt. I'm not picking on xAI. The point is that we only found out because the output was strange enough that people noticed. A thumb that presses a little more gently doesn't make the news.

The labs do publish their intentions now. OpenAI has a Model Spec, Anthropic has a constitution, and I'd rather have them than not. But they're documents about what a lab's stated intentions are, not what the results will be. They can't tell you about stage one, the part nobody meant, and at the volume of text these models take in nobody can fully audit it. Debiasing doesn't get you out either, because someone has to decide what neutral looks like, and that's another thumb. I wrote in Weights and Measures about being asked to trust a safety claim on a model you aren't allowed to inspect. This is the same shape, one layer down.

Hard to Pin Down

The Sonnet I talked to in 2025 is not the Sonnet I talk to today. Same name. New models ship all the time, and the name is a label on whatever happens to be behind it this month.

A masthead stayed attached to one editorial voice. A model name doesn't. Anthropic's documentation is straightforward about it: a model ID is pinned and the weights behind it don't change, but "the serving infrastructure around the model can change over time," and those updates "occasionally produce minor differences in observable behavior even when the model ID and weights have not changed." I'm glad that's written down. It also means the behavior you saw in March isn't promised in October, under the same ID. And the other kind of change needs no infrastructure at all: a system prompt can be edited in an afternoon, as xAI's was, and nobody tells you.

That's why I called it ethereal. The editor is always there, ephemeral, hard to see and hard to track. Awareness isn't something you do once.

What You Can Do

I don't have a clever fix, and I don't think there is one at the source. A podcaster may not know their own bias. A reporter doesn't control the edit. A lab can't label everything its model absorbed. Whatever the fix is, it can't be a disclosure printed inline with the answer.

So it sits with you. Know that the lean exists. Know roughly where it can come from, and the list above is a start. That's the epistemic provenance the masthead used to hand you for free. Pick sources whose lean you've learned to read, and for anything that matters, check more than one. I'll be honest that the last one is hard. Checking a second model often isn't the same as checking a second newspaper, because many of them learned from overlapping text.

What I'd refuse to do is hand the judgment to someone else, whether that's a platform, a lab, or a bias score, and call it solved. Your judgment is the one part of the loop that's yours.

James is a security engineer who once asked a 1930 model how many planets there are and got a list that included Pallas. He publishes at waypoint.henrynet.ca. This post was written with Claude, which is one of the editors in question. He can't show you its lean either.