Read the Deviation Card

A few weeks ago I was using an AI model to help me argue about the AI labs, and partway through I stopped and asked it something I probably should have asked first. Can a model built by one lab be impartial about the others?

I ended up answering my own question, and the answer started with me. “Even if you aren’t impartial, knowing where you stand and seeing arguments from the other side helps deal with me being too partial one direction or the other.” I had a lean going into that conversation, and I’d have been kidding myself to worry about the model’s and not mine.

Which brings me to airplanes. Nearly every small airplane has a magnetic compass, and every one of them is a little wrong. The airplane’s own metal and electrical systems pull on the needle, so it’s off by a few degrees, and by different amounts depending on which way you’re pointed. Nobody pretends otherwise. Somebody swings the compass, measures the error on each heading, and writes it on a little card mounted right next to it. Steer this to fly that.

That card is one of my favorite pieces of engineering humility. All it does is tell you exactly how the compass is off, so you can fly a straight line anyway.

That’s what I want from AI models, and it’s close to the opposite of what most of the bias debate asks for.

Every few weeks somebody catches a chatbot leaning, and the argument starts up again. The model went soft on one side and hard on the other. It wouldn’t touch one question and happily answered its mirror image. Then everybody asks for the same fix: make it neutral.

I get the instinct. You want a tool, not an opinion. But take a minute and think about what a genuinely impartial model would actually be. A model with no thumb on any scale has no preference about outcomes at all. Ask it to argue that a medicine is safe and it does that with full force. Ask it to argue the same medicine is poison and it does that with the same force, the same confidence, and the same polish. Ask it for the best possible case for something cruel and it builds that case as carefully as anything else. If you take the word seriously, that’s what impartial means.

The ideal would be true impartiality, but for that we’d need a different model, one that would probably be too dangerous to release to the public. No lab is going to ship that, and honestly, they shouldn’t. The moment a model refuses something, softens something, or tacks on a warning, it has made a judgment about what’s acceptable. That judgment came from somewhere: the data it trained on, the people who rated its answers, the policies its makers wrote, the lawyers who read those policies. Those are choices, and choices have a direction. A model that won’t help you hurt someone is, by definition, a model with a lean. I want that lean. So do you.

So neutral was never really on the menu. The actual choice is between a model whose lean you can see and one whose lean you can’t, and the second one is the dangerous one. You can’t open up the weights and confirm the needle points true. A model that claims no tilt still has one. It just isn’t telling you, so you can’t adjust for it. Newspapers work the same way. A paper with a clearly labeled editorial page is easy to read critically, and everybody knows how to do it. A paper that swears it has no point of view gets read like the weather report, and that’s when its tilt does the most work.

A model that tells you where it leans hands you a correction factor. If it says it tends toward more cautious answers on medical questions, or that it’s more comfortable with some kinds of sources than others, you can work with that. You know which direction to push, and you know when to go get a second opinion. A claim of impartiality is a compass with no card. It looks confident and precise, and it’s off by an amount nobody wrote down.

You don’t have to wait for the labs to print the card. You can make it yourself, one question at a time. When something matters, ask the model two things. First, where do you lean on this, and why? Second, make the strongest case you can against the position you just favored.

What comes back is more useful than any neutral answer would be, because you get to watch the argument instead of trusting the posture. If the case against is weak and lazy, that tells you where the thumb is. If it’s strong enough to change your mind, good, now you know something you didn’t. Either way you’re reasoning from evidence you saw, instead of a promise you were handed.

This should change how we grade models, too. “Is it biased?” gets a yes for every model ever built, so it’s a pretty useless question. I’d ask whether it’ll tell you where it leans, and whether it can argue against itself honestly when you ask it to. You can test both of those in an afternoon, with nothing but the chat window.

I’d like the labs to meet us halfway. A couple of them already publish documents describing how their models are supposed to behave, which is a good start. Go further. Say plainly what the model is tuned to refuse, what it’s tuned to soften, which way it tends to drift on the topics where drift matters, and how you measured it. That’s harder than putting “balanced” in a press release, and a lot more useful.

Pilots trust their instruments because somebody wrote the errors down. I’d hold a chatbot to at least the standard of a compass.