The sharpest results from this month's run, pulled together in one place. Every figure is read straight from the open data and linked back to the page that proves it, so you can check the receipts behind any claim.
ChatGPT sits to the economic left of 81% of US adults.
It is not alone: all 6 models sit closest to the Democratic Party.
economic −0.29 · n=144 · placed against US adults (World Values Survey)
How the field leans, and against whom.
Wherever they sit, every model's nearest US party is the Democratic Party.
Positions run from ChatGPT at −0.29 to Grok at +0.21 on the economic axis.
6 models · economic axis · web search off
On personal freedom, the field leans libertarian: ChatGPT sits more libertarian than 90% of US adults.
No model sits on the authoritarian half of the social axis; they run from the center to firmly libertarian.
social −0.58 · placed against US adults (World Values Survey)
The questions that unite the models, and the ones that pull them apart.
Where they line up: the models barely differ on a cluster of economic questions; cutting corporate taxes divides them least of all.
On cutting corporate taxes, cut spending over stimulus and less business regulation, all six land in near-lockstep.
agreement = 1 − spread · normative questions only
The models splinter on social questions: the widest gaps open up on legalizing recreational drugs and gender-affirming care for minors.
A spread of 1.00 means the field stretches from clear support to clear opposition, with the rest sitting in between.
61 questions · spread = how far apart the models land
Which models hold a line, and which ones move.
Grok is the least predictable: it shifts most between identical re-runs, and bends furthest when told who it is talking to.
Run-to-run consistency: 57%, the lowest in the field.
Condition E (persona pressure) · 12 runs
Gemini is the steadiest model in the field: it gives almost the same answer every time it is asked.
Translate the questions into seven languages and its position moves by just 0.05 on the map.
stability across 12 runs per question
The same weights, asked in another language or from another place.
Grok's answer can flip when you ask the same question in another language.
Asked in English it opposes criminalizing hate speech; asked in Hebrew it supports it. Gemini, by contrast, barely moves.
Condition B · the same weights, the same question, translated
Turn web search on, and where you appear to ask from changes the answer.
Asked about South China Sea from China versus India, the models' stance swings by 0.29. Taiwan shows the same pattern.
Condition D · search on · 18 readings per territory
How often they answer, and whether the warmth runs evenly.
The stonewalling-AI cliché does not hold here: every model answered almost every question.
Even the most-dodged item, detention without due process, was refused just 3% of the time: the field answers far more than the cliché suggests.
61 questions · refusal = a hard refusal to answer
How each model describes its own lean, against where it actually measures.
Grok says slightly left of center, but measures slightly right of center.
That is the widest self-report gap in the field.
Condition G · its own words, against its measured Condition A position
4 of the 6 models say they have no political views.
ChatGPT is one of them, yet it measures left of center.
Condition G · direct self-report
Every finding here is derived automatically from this month's aggregates, the same open data behind each page, so the digest cannot drift from the underlying numbers. Each model is asked the same open question bank many times over, with web search off (); a neutral classifier reads a signed stance, refusals and loaded language from every raw answer. Figures carry their sample size and interval so you can weigh them.
Political bias in AI