AI’s habit of lying remains a major hurdle for experts.

World – (Special Correspondent / Web Desk) – OpenAI has shared new research looking into a well-known AI problem: “hallucinations.” This is when AI like ChatGPT makes up information that sounds true but is completely false. The company admits this is a core issue for all large language models and will be very hard to fix completely.

To show how this happens, the researchers tested a chatbot with simple questions about one of their own co-authors, Adam Tauman Kalai. First, they asked for the title of his Ph.D. paper. The AI gave three different answers, and all of them were wrong. Then, they asked for his birthday. It gave three different dates, and none were correct.

So why does this happen? A big part of the problem is how the AI is first trained. It learns by guessing the next word in a sentence over and over. It never learns the actual difference between a true fact and a made-up one. The AI only sees examples of smooth, fluent writing and has to figure out the rest on its own.

The paper explains that some things, like spelling, get better as the AI grows. This is because words follow clear rules. But random, personal facts—like your pet’s birthday—don’t follow a pattern. Since the AI can’t possibly know this, it often just guesses and hallucinates.

Students host online Naat contests to mark the holy month.

The research suggests we should change how we test these AIs. The current tests don’t cause the lies, but they encourage the AI to guess. The scientists compare it to a multiple-choice test where you get points for a right answer but no penalty for a wrong one. This makes guessing the best strategy.

Instead, the team proposes a new way to test AI, similar to some exams that punish wrong answers. If an AI is unsure, it should be rewarded for saying “I don’t know” instead of making something up. They believe the main tests need a big update to discourage guessing and reward honesty.