Christopher Barrie

Just generated this method summary of the paper w/ Nano Banana 2. Crazy good

Omg I've unintentionally become a New York yoga mom

Into my veins

I wanna see a play about Jack Dorsey

Hearing that Matt Goodwin has joined the Green Party saying it's all actually been a big misunderstanding

Yes it was written by gpt4o tbf though

Ah that's super kind

Thank you!

It'll be front page of NYT by then

Paper with DOI now SocArXiv: osf.io/preprints/so...

Now on @socarxiv.bsky.social !

osf.io/preprints/so...

Thank you, Laura!

Hyped to see this featured in the FT today!

Anthropic: we hired philosophers as part of our team because arguments

Counterpoint: maybe don't do that

"The welfare of the models"

ur mean

now now

Yes we liked it too!

Implication:

For questions about polarization, dimensionality, or demographic sorting, over-constrained synthetic data can mislead.

Silicon samples are useful—but require structural diagnostics.

Full paper: drive.google.com/file/d/1bsmC...

🧭 Figure 6: Missing structure

Project LLM belief systems onto human belief axes.

Finding:
• Dominant human dimension → amplified
• Secondary dimensions → attenuated or missing

LLMs exaggerate the main axis and erase cross-cutting structure.

📚 Figure 5: Persona vs residual variance

In the GSS, most structure persists after conditioning.

In LLMs, early principal components are heavily persona-mediated.

Synthetic belief systems are more demographically sorted than real ones.

🧩 Figure 4: Persona mediation

Remove demographic effects and recompute constraint.

In humans: modest change.
In LLMs: large drop.

Demographics carry too much of the synthetic ideological backbone.

Put simply:

Humans: weak, multi-dimensional, cross-cutting belief systems.

LLM personas: concentrated, low-dimensional, highly predictable.

Not just stronger ideology—flattened geometry.

📈 Figure 3: Effective dependence (Dₑ)

Captures global linear dependence across all dimensions.

LLMs show much higher Dₑ → fewer effective dimensions.

Belief variation is compressed.

📊 Figure 2: Dominant-axis constraint (PVE₁ = λ₁/p)

LLMs allocate more variance to the first principal component.

Interpretation: belief systems collapse too strongly onto a single ideological axis.

Too one-dimensional vs humans.

📊 Figure 1: 🪐 Ellipses summarize the distribution of pairwise correlations.🪐

LLM personas show systematically tighter contours than the GSS—especially in the upper tail.

Beliefs co-move too strongly.
Synthetic publics are overly rigid.

We benchmark 28 models using 52 attitude items + demographic personas.

Key test: not just matching averages, but matching belief-system structure:

i.e., the covariance geometry of opinions.

Structure > marginals.

And this is what we find...

There's this classic finding we all know about in political sociology: mass publics are structured but weakly constrained

Meaning...

Real people hold cross-cutting views.
Attitudes only loosely predict each other.
Multiple latent dimensions coexist.

Do LLM personas preserve this messy geometry?

🧵on my new paper "Synthetic personas distort the structure of human belief systems" w Roberto Cerina I'm v excited about...

🚨 Do synthetic samples look like human samples?

We compare 28 LLMs to the 2024 General Social Survey (GSS) to find out + develop host of diagnostics...

Me too

What's the opposite of a sleep-in and also when your kid repeatedly kicks you in the face in their sleep?

Kick-out?