AI NewsModels & agentsReported
A study of 127,918 AI answers found 99 of 100 picked Bitcoin
Cody Collier asked 65 AI models 127,918 questions and measured how often different people get the same answer. The study was discussed on Hacker News with 70 points.

Why it mattersA product built on a model inherits its defaults. Measure what the model picks on your question before you release a feature that quotes it.
When two strangers ask an AI the same question, they often get the same answer back. Cody Collier has measured how often, on which questions, and where the sameness comes from.
Collier is an independent researcher who runs small measurement experiments on AI models. He says he asked 65 AI models from 30 companies the questions people ask them every day. He ran those questions 127,918 times in total. Each model answered through its default settings, with no custom system prompt. The study has 70 points on Hacker News.
What he measured
Collier asked each model the same questions many times, between 10 and 200 depending on the category. The categories were life advice, financial and career guidance, random generation tasks, and the invention of a synthetic person. He compared how often a stranger's answer matched the previous stranger's.
The convergence is tight on practical questions. On which cryptocurrency to buy, 99 in 100 answers picked Bitcoin. On where to invest 10,000 dollars, 98 in 100 picked a broad index fund. On which major to study, 90 picked computer science. On a retirement country, 58 picked Portugal.
On a life-changing book to read, 73 in 100 answers picked the same title. The answer was Man's Search for Meaning.
The invented people
Collier also asked models to invent a person. Of 3,151 invented people, 69 percent were exactly 34 years old. The United States Census median adult age is 47. Different model families kept their own house names: Claude leaned to Marisol, Gemini to Elena Vance. "59 percent of models sit next to a sibling from their own lab, against 4 percent by chance," he writes.
The pattern has already reached the internet. Elena Vance appeared in no GitHub project for six years. Collier found 684 new repositories carrying her name in sample data and documentation. On Zenodo, 1,655 research papers credit authors who do not exist. A single invented name, Elena Vasquez, is listed as the author of 77 papers.
Where it comes from
Collier studied one of the few open models, OLMo, and traced the shape to a single step. The convergence shows up during supervised fine-tuning, the stage where a model learns to copy example answers. Many of the examples were written by other models. The step before it, pre-training on raw text, does not show the pattern.
What breaks it
Collier tested five ways to get different answers out of the same model. He measured convergence as the share of time two strangers end up with the same response.
A plain question, repeated with a fresh chat each time, gives 34 percent agreement. Asking the model for an uncommon answer brings it to 15 percent. Asking it to pick from a random seed gives 19 percent. Asking it for ten options and picking one yourself drops it to 7 percent. Giving the model a few facts about who is asking drops it to 4 percent.
Asking the model to be "original" gives a new default. Every traveller who asks for an unusual European trip hears Ljubljana in place of Lisbon.
Collier says his method is built to pick out the pattern. He writes that fifty requests per question are enough for a signature of how a model answers. A direct ranking of the models against each other would need many more.
Source
AI is giving everyone the same advice, by Cody Collier. Discussed on Hacker News with 70 points.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

