Priming

Priming roughly means “preparing” or “pre-conditioning”. The term comes from psychology. There it describes a simple finding: someone who has just read the word “doctor” recognizes the word “nurse” faster afterward. The first stimulus has thus pushed thinking in a certain direction, without the person noticing it. In the tech world, the word is nowadays mostly used for something else: you first give a text program a few sentences or examples, and this input determines how it solves the actual task afterward. In both cases, the same principle is meant – the beginning colors the outcome.

Why a single sentence beforehand changes so much

Programs like ChatGPT don’t work with fixed rules. They continue texts by estimating the most likely next word each time. What is likely depends entirely on what came before in the text. That’s exactly why priming works so strongly here. If you write “You are a tax advisor” before the question, the answer sounds different than without that sentence – same question, different result.

For usage, this is good news. You don’t need to retrain a program to make it suitable for a task. A few lines placed beforehand are often enough. Companies save enormous computing costs this way, since retraining a model costs millions. Priming costs a few seconds of typing.

But there is also a downside. Anyone who cleverly formulates the input can push a model toward statements it is actually supposed to avoid. Security researchers call such tricks jailbreaks, meaning breakouts from the built-in rules. And in psychology, priming is a well-known tool in advertising: someone who sees vacation images before a purchase decision decides differently.

Instruction, examples, role: the three usual approaches

The simplest approach is the role instruction. You tell the program who it should be and for whom it is writing. For example: “Explain it to a tenth-grader, without technical terms.” The rest of the conversation then usually stays in this style. Many chat services have such a hidden instruction permanently built in, the so-called system prompt.

The second approach is examples. You show two or three already-solved cases and only afterward pose the actual task. The model recognizes the pattern and imitates it – for instance, with the question of whether a customer review is positive or negative. Experts call this few-shot prompting, meaning “instruction with few examples”.

An important distinction: priming does not change the model itself. Nothing is stored, nothing is learned. The effect only reaches as far as the text remains within the model’s field of view. If this field of view is full, the beginning falls out of memory, and the instruction fades away. It’s different with finetuning, the retraining process – there the model changes permanently.

Priming in chatbots, search engines and advertising

Priming is most commonly encountered in one’s own interaction with chatbots. Someone who just types “Write something about stocks” gets something arbitrary. Someone who first defines target audience, length and tone gets something useful. Practically all guides on the topic of prompting describe exactly this technique, even if they don’t use the word priming.

In products, priming is built in invisibly. The assistant in a banking app was given an instruction that keeps it polite, cautious and on-topic. AI answers in search engines also have found webpages pushed into the text first, before the question comes. The answer then relies on this material.

In business news, the term also appears in the old psychological sense. Market researchers study how price anchors shape purchasing decisions. A common misconception, by the way, is to consider priming a magic trick. It shifts probabilities, it forces nothing – a model can answer incorrectly despite a perfect instruction.