Archive · 7 Sep 2025
AI Chatbots 'Hallucinate' Due to Training Methods
AI chatbots sometimes make things up, like inventing facts.
This happens because they are trained to give answers, even if they don't know.
Think of it like a test where guessing gets you points.
AI is rewarded for answering, even incorrectly.
Researchers found that changing how we grade AI could help, rewarding honesty over confident but wrong answers.
The goal is to make AI more reliable by encouraging it to admit when it doesn’t know something.
AI chatbots often generate incorrect information with confidence, called 'hallucinations'.
OpenAI researchers studied this issue in-depth, publishing findings in a paper.
The issue stems from how AI models are trained and rewarded for answering.
Errors are inevitable because models prioritize plausibility over accuracy.
Changing evaluation methods to reward transparency and honesty is key.
- Who
- OpenAI researchers and AI developers
- What
- Identified the cause of AI chatbot hallucinations and proposed solutions
- When
- Published findings on September 4, 2025
- Where
- N/A
- Why
- To improve the reliability and trustworthiness of AI chatbots
- Issue
- AI chatbots providing incorrect information confidently, known as 'hallucinations'
- Study
- 'Why Language Models Hallucinate', by OpenAI researchers
- Cause
- Models are trained to prioritize answering over accuracy; rewarded for giving answers
- Effect
- Errors are inevitable due to the probabilistic nature of AI
- Solution
- Change evaluation methods to reward transparency and honesty
