Korva is a scroll of the internet’s best ideas.
Practice English by planning study abroad
The class task
Agne Kask, the school's head languages teacher, picked a task that would interest the students, and she told them to use their school accounts. The class talked about prompting for a while. Then she shared her own prompt and asked how she had made it. A student answered that she had asked ChatGPT to write it. Of course she had: she gave it a little information about the class and the task the students would do at the end.
The chatbot's questions
One student pasted the prompt in, and the chatbot began asking her questions about her future. Her answers were pilot, flight attendant, and sports coach.
Checking the facts
Students who already knew what they wanted could check whether the advice lined up with their own choice. At the end, Agne told them to do the research themselves and not rely on ChatGPT, since it might have outdated information or be wrong about some things. They were to check the school website to make sure.
Why do chatbots give outdated information?
A language model learns from a snapshot of text collected up to a training cutoff date, so it knows nothing after that unless it is connected to search. It also generates text by predicting likely next words, not by looking facts up in a record. That can produce fluent statements that are false, which researchers call hallucinations. Program deadlines, tuition, and entry requirements change often, so they are the kind of detail most likely to have drifted since training.
Why does chatting help language learners?
Linguist Merrill Swain proposed in 1985 that learners improve by producing the language, not only hearing it, because speaking or writing makes them notice what they cannot yet say. Michael Long's interaction hypothesis from the early 1980s added that negotiating meaning in conversation, such as asking for clarification, drives acquisition. Swain's idea came from studies of Canadian French immersion students who understood French well but spoke it with many errors.
Why does adding context improve chatbot answers?
A model generates each word based on the full prompt, so details about the audience, the goal, and the format narrow the range of likely responses. A vague request matches many kinds of text from training, while a specific one matches fewer. The practice of drafting such instructions is called prompt engineering. Having a model draft the prompt itself is a common technique.
More from The Journal.
