02 / Case study
Style in the Age of AI
"Can a model make a physics paper read like art history while keeping every claim intact?"
Explore the research- Role
- ResearcherFull-time summer 2026 · Ongoing research
- When
- June 2026 – Present
- Where
Stanford University, Center for Spatial and Textual Analysis (CESTA)
PI: Dr. Mark Algee-Hewitt
Research schematic
How a rewrite is made, and how it is checked
Writing and evaluation are separate steps: local code and independent model calls check each rewrite.
Stage 0: before the pipeline
10
frontier models
4
local models
Model selection
Blind-tested on the transfer task
Selected: Claude, via the API
Model benchmarking informs the choice of writer; it is separate from the checks run on each rewrite.
Discipline profiles
Measured, not prompted
- NER
- UDPipe
- Sentiment
~1,000
articles per field
19
style features
Instructions come from the measured gap between two profiles, not from prompts like "write like an art historian."
A
Source paper
The text to be rewritten
B
Target-profile gap
Measured difference between source and target field
C
Intensity
YAML level, L1 light touch to L7 full rewrite
L1L2L3L4L5L6L7
Style transfer
Claude rewrites the source toward the target field
Rewritten text
Sent to every check independently
Four independent checks
Check 1
Claim survival
Are the original facts and claims present, changed, or missing?
Check 2
Gist recognizability
Is it still recognizably the same paper?
Check 3
Verbatim overlap
Did the model copy language from the target corpus?
Check 4
Fabrication
Did the rewrite introduce claims unsupported by the original?
Evaluated by local code and separate model calls.
Eval dashboard
Run experiments and compare outputs and evaluations side by side, replacing the command line.
The central difficulty
Voice and meaning are not cleanly separable
Adding a disciplinary voice is computationally trivial. Stripping one removes meaning: removing conventions can erase or distort factual claims.
One early physics-to-art-history test (illustration, not an aggregate metric)
About one-third of 95 factual claims were altered or omitted in that test.
01
The question
Every academic field has a voice. Physicists and art historians don't just study different things; they write differently. The project, "Style in the Age of AI," asked whether that voice can be measured precisely enough to transfer: rewriting prose in the conventions of another discipline without changing its claims.
02
Choosing a model
Before building, I blind-tested 10 frontier and 4 local large language models on the task, then selected Claude, accessed through the API, as the engine for the pipeline.
03
Fingerprinting a discipline
To give the model something concrete to aim at, I scored 19 style features across roughly 1,000 articles per discipline, using named-entity recognition, UDPipe parses, and sentiment analysis. Together these formed a per-field fingerprint.
- Transfer instructions were derived from the measured gap between two fields' profiles
- Those instructions were generated across seven intensity levels, from a light touch to a full rewrite
04
Keeping the claims honest
Style transfer is only useful if the meaning survives. I wrote four automated verifiers to check each rewrite:
- Content: are the original claims still present?
- Gist: does the rewrite say the same thing overall?
- Verbatim overlap: did the rewrite copy language from the target corpus?
- Fabrication: has anything been added that wasn't there?
05
Scaling up
The pipeline grew from 2 to 8 discipline profiles. I tested physics to art history in both directions, and the harder economics to law pairing, where the vocabularies overlap but the conventions don't.
06
From terminal to dashboard
I replaced the command-line workflow with a dashboard so the pipeline could be run and inspected without touching the code. I presented the work at the mid-summer review and spoke solo at the August symposium.
Next case study
NoteCoach.ai