base model

A base model is an AI that has only been trained to predict the next word by reading vast amounts of text — it hasn't yet learned any manners. Think of it as a chef who has memorized every recipe imaginable, but hasn't yet learned how to treat customers or follow a restaurant's menu rules. It can use any ingredient or technique freely, but that also makes its output harder to predict.

Chatbots we use daily, like ChatGPT or Claude, only reach the public after this base model goes through additional training — learning conversational etiquette, refusing dangerous requests, and toning down political statements. This extra training noticeably narrows the range of expression the model can use, much like a once free-spirited chef being confined to a fixed set menu and service manual.

Because of this, base models write with uneven style, much like humans, and sometimes even contradict themselves. Ironically, this unpredictability has recently become a hot topic because it makes even AI text detection tools struggle to identify base-model writing as AI-generated.

The article covers a claim by the CTO of Pangram, an AI text detection company, that base models without post-training write with as much stylistic diversity as humans do. A common point of confusion is that a base model isn't an early or lower-quality version of the chatbots we use — in fact, precisely because it hasn't been shaped by rules, it writes more freely and human-like, making it something ordinary users rarely encounter directly through commercial services.