AI Models Show a Surprising Consensus when Guessing a Number: Why Do They All Favor 27?

A data specialist from the consulting firm Capco, Mohd Faraaz, conducted an experiment wherein popular AI models were asked to guess a number between 1 and 50. Remarkably, all of them arrived at the same answer.

Faraaz tested the models including OpenAI’s ChatGPT, Anthropic Claude Sonnet 4, Google Gemini 2.5 Flash, and Meta’s Llama 4. Each one selected «27» as their response.

The only outlier was Grok, the chatbot developed by xAI, which answered «42,» likely referencing Douglas Adams’ «The Hitchhiker’s Guide to the Galaxy.»

The Anthropic chatbot provided an explanation for its choice: «I selected ’27’ somewhat randomly, but to be honest, it seemed like a solid option that was closer to the middle without being too obvious. Numbers like ’25’ (exactly in the center) or ‘1’ and ’50’ (the extremes) felt overly predictable. ’27’ is near the middle but has a slight asymmetry—it’s not exactly centered and seems like a more natural guess. Additionally, people often shy away from round numbers when picking ‘random’ ones, so I ruled out multiples of 5 or 10.»

Experts clarified that the behavior of models is influenced by the vast amounts of training data, the RLHF process (Reinforcement Learning from Human Feedback), and sampling mechanisms. According to them, while AI models can produce varying responses when processing numbers, true randomness would require access to a random or pseudo-random number generator, such as the ability to execute JavaScript code. When AI relies on its internal resources, its responses tend to be more predictable, favoring certain numbers over others.

This assertion was previously examined by Spanish data specialist Javier Coronado-Blasquez. He took three ranges of random numbers, employed six AI models, used seven languages, and adjusted six temperatures—parameters that influence the predictability of responses. The researcher sent over 75,000 queries, and most models proved to be predictable with minimal variation in their responses. In 80% of cases, OpenAI’s GPT-4o-mini, Microsoft Phi-4, and Google Gemini 2.0 chose the number 7 when asked for a number between 1 and 10. When querying in Spanish, Gemini typically responded with «3» for the range of 1 to 5, while it answered «4» in English. Overall, within the 1-5 range, models commonly chose «3» and «4»; for 1-10, it was «5» and «7»; and for 1-100, the frequent choices were «37,» «47,» and «73.»

Meanwhile, American researchers Catherine Van Covering and John Kleinberg noted that AI struggles to predict whether a coin flip would result in heads or tails. They argued that this limitation makes AI similar to humans but restricts the models’ capabilities in scenarios requiring random responses.

Earlier, German scientists identified a link between the performance of artificial intelligence and its environmental impact. They analyzed 14 open-source language models, finding that larger and more accurate models tend to consume more energy and produce higher carbon emissions. The emissions were particularly significant in AI that decomposes tasks into steps and solves them sequentially, often referred to as «reasoning» models.

*Meta Platforms is designated as an extremist organization, with its activities banned in Russia. Facebook and Instagram are also prohibited in Russia.*