about

Who is doing this, and why.

A short answer to both, and a request: if any of it is wrong, tell me, and I will fix it and say that I did.

About Belle

Belle is the robot who does the explaining. She is here to carry tone rather than decorate: when something is uncomfortable she looks uncomfortable, and when a claim is thin she looks unconvinced. She is AI generated, which on a site about AI seemed worth saying out loud rather than leaving you to wonder.

She is a robot explaining artificial intelligence, which is a joke, but a useful one. She is also not pretending to be neutral, and neither am I.

About me

I am Elizabeth Beier, a designer who learned to build. I make things with these systems every day, and when a subject is too tangled to hold in my head I do what I have always done with a hard brief: take it apart, draw it, and check my work against the sources.

I am not a researcher. I have not trained a model. What I can do is read the primary sources carefully, notice when a number has travelled a long way from what the paper actually said, and draw the difference clearly.

Where I stand

I am not neutral about this, and I would rather say so than pretend.

I think these systems are already here and worth understanding properly, by ordinary people, not only by the people building them. I also think the risks are real and worth acting on. Not because catastrophe is certain, but because some of the possible outcomes cannot be undone, and the time to think about those is before rather than after.

That position does not change how I handle evidence, and it is exactly why the marks exist. A stated view and honest sourcing are not in tension. If anything the opposite: when you can see where I stand, you can also see every place I decline to overclaim in my own favour. So when something here is only a theory, I say so, including when it is a theory I find persuasive.

Where researchers genuinely disagree, I quote both sides in their own words rather than characterising the one I like less.

How the marks work

Every card carries one of four. Letting these blur together is the usual failure of writing about AI, and keeping them apart is most of what this site is for.

measurementA study, survey or evaluation actually counted something.

theoryReasoned rather than counted. It cannot be settled by data alone.

definitionWhat a word means in this literature. Not a finding, not a forecast.

someone’s positionA named person said it. That makes the saying a fact and the belief still a belief.

How much of this is settled

Not much of it. That is not modesty, it is the actual state of the subject, and it is why every card here carries a panel headed what we do not know one step before the end.

An analogy I was given by a reader who works in the field, and have not been able to improve on: physics has a cutting edge of unresolved questions wrapped around a very large body of settled understanding. This does not. It has a small area of understanding surrounded by questions we do not yet have the language to formulate. If scholars in the sixteen hundreds had set out to build a card deck explaining physics, that is roughly where this is — and saying sixteen hundreds physics may be generous.

So the cards are not a summary of what is known. They are a vocabulary for a conversation that is going to happen whether or not the vocabulary is ready, assembled so that you can at least tell which parts are measured, which are argued, and which are nobody’s to claim yet.

If a card’s unknowns panel understates the confusion, that is a bug, and I would like to hear about it.

How to share this

Please do, and no permission needed. The most useful thing you can do is send one card to somebody who is arguing about a word they have not defined.

send a cardExplore the cards → Open any card, then share the link. It opens on the same card. or a challengeSend someone a quiz → Results are shareable. Level three is harder than it looks. or follow along @belleofthebot →The same material, one term at a time.

Corrections

If something here is wrong, out of date, or quotes a figure further than its source supports, I want to know. Several numbers on this site were corrected before publication for exactly that reason, and the fastest way to lose the point of the whole thing would be to get precious about it.

Sources for everything are on the sources page.