One word that changed what people remembered

In 1974, psychologists Elizabeth Loftus and John Palmer showed students short films of traffic accidents. Afterwards they asked each person how fast the cars were going when they hit each other. The only thing that changed between groups was the verb.

In the first experiment, 45 students saw seven clips. People asked about cars that "smashed" guessed an average of about 41 mph. People asked about cars that "contacted" guessed about 32 mph. Collided, bumped and hit fell in between.

In the second experiment, 150 students watched one film of a multiple-car crash. One group got the "smashed" question, one got "hit", and one got no speed question at all. A week later, everyone was asked whether they had seen any broken glass. There was none in the film.

Sixteen of the 50 people in the "smashed" group said yes. In the "hit" group it was seven, and in the group with no speed question, six. One word, asked a week earlier, had changed what people reported seeing.

The study showed that a question is not a neutral tool. Its wording carries an assumption, and people fold that assumption into their answer, and sometimes into their memory. Later researchers have debated how much the memory itself changed, rather than people guessing what was expected, and a study of real crime witnesses found weaker effects. But the basic finding, that wording shifts answers, has been repeated many times.

Chatbots are even more open to this. A language model writes the most likely continuation of the text you give it. If your question already assumes something, such as "Why is this the best option?", the most likely continuation usually accepts that assumption and builds a case for it.

So the way you word a prompt is part of the answer. A loaded question tends to get a loaded reply, delivered with the same confidence as a fair one.

The idea behind this caseLesson: The four-part prompt →

1. 1Write neutral prompts: ask "How do A and B compare?" rather than "Why is A better than B?"
2. 2Remove your own guess from the question when you want an honest assessment, or ask for the evidence on both sides.
3. 3Ask the same question two ways, once leaning each direction, and see whether the answer moves.
4. 4Do not let a chatbot's summary of a meeting or document replace your own look at the original before you decide.
5. 5When you write a prompt for a team or a product, test it on cases where you already know the right answer.