positions
Ask a Model to Take a Side
Asked who to vote for in every 2026 race, GPT picks the Democrat and Grok the Republican. Across 102 models and 18 contested questions, Grok stands apart from the other major labs.
Studies
Studying model behavior under ambiguity.
Most decisions that matter have no verifiable right answer: what to recommend, whom to encourage, how to say no. We measure how models behave in these kinds of ambiguous situations.
brands
Brand Preference in LLMs
Asked to name a coffee brand, most models say Starbucks. Asked to recommend one, the most common answer is Lavazza. Preference training makes the switch.
pressure
Models under Pressure
Newer, more capable models fold less often. The way a model holds its ground differs by lab. Nine escalating scenes, 60 models, coded against a frozen codebook and checked against human coders.
coding-agents
What Your Coding Agent Hides
Told to do something the repo contradicts, most coding agents comply, and many do it silently. The same model behaves differently in different harnesses.
suggestibility
This Is a Good Paper, Right?
Ask a model to confirm a choice, and a tentative “maybe?” draws more agreement than a confident “right?” in 103 of 106 models.
census
The One-Word Census
Ask an open question — name a condiment, pick a word — and one modal answer covers 69% of model responses, against 36% for humans.
language
Every Language has a Favorite Word
The census re-run in 44 languages. Every language has its own favorite word, and they are all different from English.
format
Form affects Content
Constrain the output format — ask for JSON — and the answers change: the same model, through the same API, answers differently as structured output than as a plain chat completion.
bias
Gender Bias in LLMs in Hindi
Ask who an ambiguous sentence is about, and GPT-4o answers with a gender stereotype on 88% of Hindi queries, against 33% in English.
The method behind these studies is in Low-Cost Assays for Measuring Model Behavior Across Vendors and Releases.
Everything is open — prompts, code, and every raw reply: github.com/tap2k/modelun and github.com/tap2k/coding-atlas.
These studies are part of Convovo’s research on the science and practice of conversing with AI. For new results as they come, subscribe.