AI models never cease to fascinate us with both their abilities and limitations, exhibiting strange behaviors that shine a light on their basic form. One noteworthy observation is that these algorithms appear to have preferences for specific numbers, similar to how humans select numbers. This is not merely a surface-level oddity; it provides information about how these models function.
Humans have a well-documented problem with randomness, frequently overthinking or misinterpreting what it really means. For example, when asked to forecast the result of 100 coin flips, human predictions typically do not include the heads-or-tails streaks that are characteristic of real random sequences. In a similar vein, when choosing a number between 0 and 100, people tend to prefer more "neutral" options, such as those that end in 7, rather than extreme or visually striking numbers.
Unexpectedly, AI systems exhibit the same bias in number selection as humans do. In an informal experiment conducted by Gramener engineers, many major language models (LLM chatbots) were asked to select a number between 0 and 100. The final results were by no means random. Even when the settings were changed to promote unpredictability and increase randomness, each model continually showed a preference for particular numbers. For instance, the GPT-3.5 Turbo model from OpenAI frequently selected 47, Anthropic’s Claude 3 Haiku went with 42 while Gemini opted for 72. Likewise, other models had their own number preferences.
Additionally, these models steered clear of low and high numbers and rarely opted for double digits or round figures unless the temperature setting was maximized, which led to less common selections by Gemini.
This is not due to consciousness or a knowledge of randomness, but rather to a reflection of the training data of the models. Without any real logic, these systems produce answers based only on how frequently they have been fed data. A number is less likely to be chosen by the model if it doesn't frequently occur as a response in the training set.
This mimicking of human-like choice patterns in seemingly simple tasks like number selection illustrates how AI models reflect the biases and behaviors present in their training data, often leading to anthropomorphic interpretations of their actions.
Despite appearing to "think" like humans, these models lack comprehension and consciousness; they function only on the data they process, exposing both the potential and limitations of present AI technologies.