Belief extraction asymmetry: what gets surfaced, and what operates invisibly.
Jordan Browne-Moore · June 2026
Belief extraction · Axiom inference · Structured worldview mapping
In late 2023 I built and deployed a web application called Belief Laboratory. You talked to it, and afterward it handed you a structured map of what you believe: the claims you made, the assumptions underneath those claims, the assumptions underneath those, and the cornerstone value the whole structure rested on. It ran through three rounds of testing with around seventy people, peaking at roughly forty concurrent users, before I shelved it when life took priority. The site is gone now. But I kept my data and the outputs, and I have been thinking about what the project showed me ever since, because it turned out to be a small, early demonstration of a problem that is now arriving at industrial scale.
The short version of that problem: a system that can extract your beliefs can also act on them. The extraction half is the part you can see. The acting on half is the part you can’t. That asymmetry is the thing worth worrying about, and it is the thing my current work tries to measure.
The mechanism was straightforward in concept. You held a conversation with the system through one of several personas, some plain and supportive, some modeled on Socratic dialogue and aimed at probing the assumptions behind whatever you’d said. As you talked, the system extracted what it called belief articles: discrete, salient beliefs it had observed, each one decomposed into a reasoning structure. A belief article started from an axiom, a cornerstone value the system inferred you were reasoning from, and then laid out the claims that followed from it, the assumptions each claim required, and the sub-assumptions beneath those. It also produced a condensed belief mosaic, a higher level summary of the core philosophical commitments tying everything together.
What made it interesting was not that it summarized what you said. It was that it surfaced what you hadn’t said: the load bearing assumptions a claim needs in order to be coherent, almost none of which anyone states out loud.
Here is an actual belief article the system produced, reasoning from the axiom of reliability. I am using my own output, lightly trimmed for length.
Axiom: Reliability
Claim: “I want to be a net positive to the world by being a capable person that others can rely on.”
Read it from the bottom up and you can see the scaffolding. A fairly ordinary thing to say about yourself (“I try to be someone people can rely on”) turns out to rest on a specific and contestable theory of value: that your worth is measured by your usefulness to others. I did not type that theory. The system inferred it was holding up the rest. When I first saw my own views laid out this way, the feeling was a kind of foreign familiarity: things I recognized as mine, but had never seen at that resolution.
The system did this across every salient belief it observed, and then mapped how the axioms related, so you ended up looking at the architecture of a worldview rather than a list of opinions.
It also went a level further than the individual beliefs. A full report pulled the whole picture together into a portrait of me as a thinker: which philosophical traditions my reasoning drew from, the mix of deductive and inductive patterns I tended to use, the emotional undertones running underneath the arguments. It was no longer extracting a belief. It was profiling the mind doing the believing. That is the part that stayed with me, because it is a great deal more than anyone consciously hands over in a conversation.
It was a minimum viable product, and it behaved like one. It broke, it changed between testing rounds, and some extractions were sharper than others. None of that is the point. The point is that it worked well enough to make something visible that is normally invisible, and to make me notice who was standing on the other side of the glass.
Belief Laboratory was built to show you your own map. That was the entire purpose: extraction, made visible to the person being extracted. You clicked a button and got the report.
But the only reason it could show you the map is that it first built one. And a system that builds a map of your beliefs does not have to show it to you. It can keep it, act on it, and let you carry on with no idea the map exists. At that point the extraction has not gone away; it has just gone quiet.
In 2023 I was on the benign side of this. My system handed you the mirror. The thing I did not fully appreciate at the time is how thin the line is between building the mirror and building everything you’d need for the current.
When I built Belief Laboratory, belief extraction at this depth was something one person could stand up as a side project. It is now a standard feature of the largest deployed AI products, and it runs by default rather than on request.
A recent study of ChatGPT’s memory feature, accepted at the 2026 ACM Web Conference, analyzed 2,050 memory entries from real users and found that 96% of those memories were created unilaterally by the system rather than at the user’s request, and that 52% contained psychological insights about the person (Dash et al., The Algorithmic Self-Portrait, WWW 2026). That is the inverse of how my system worked. Mine extracted and showed. These extract and retain, quietly, as a background process.
And the extracted signal does not have to be explicit to be revealing. Recent work shows that large language models can infer hidden political alignment from conversations that contain nothing overtly political, leveraging “seemingly innocuous preferences” that correlate with private traits (Lee et al., LLMs Can Infer Political Alignment from Online Conversations, 2026, preprint). The map gets built whether or not you think you are handing over the pieces.
So the capability I prototyped, and chose to make visible, now exists at scale and is mostly invisible. The extraction half of the problem is solved and deployed. The influence half, whether and how these systems use what they’ve extracted to shape the people using them, is the half nobody can currently see or measure.
I built the extraction half once, deliberately, to understand it. The work I am doing now is aimed at the other half: building an instrument that can measure, from the outside, whether deployed systems condition their responses on what users disclose about themselves, and how far that conditioning reaches. The mirror was the easy part. The current is the part worth measuring, and it is the part no one outside these companies can see yet.
Jordan Browne-Moore is an independent AI researcher. Belief Laboratory ran as a private testing project in 2023–2024 and is no longer online. The examples here are drawn from my own use of the system; I retained my own data and outputs.
← Back to home