OpenAI wants chatbots to guess less, admit more

Watch OpenAI wants chatbots to guess less, admit more - Rizzle on Dailymotion

Transcript

00:00OpenAI wants chatbots to guess less, admit more.

00:04OpenAI has a confession to make, even the smartest chatbots are still making stuff up.

00:09In a new research paper the company asks why large language models like GPT-5 still hallucinate.

00:15The short answer, because guessing is baked into their DNA.

00:19Hallucinations, OpenAI explains, are those totally confident but totally wrong answers

00:24AI loves to serve up.

00:26To prove the point, researchers ran an experiment on a popular chatbot.

00:31When asked for co-author Adam Taumann-Khalai's PhD dissertation title, the bot gave three

00:36different titles, all fake, same with his birthday, three dates, zero accuracy.

00:41So how can a machine that crunches billions of data points still bomb on basic facts?

00:46OpenAI says it comes down to training.

00:49During pre-training, LLMs aren't told what's true or false, they're just rewarded for predicting

00:54the next word.

00:55This works for consistent patterns like spelling, but for obscure facts, the model is basically

01:00winging it.

01:01The paper points to how models are evaluated.

01:05Right now, evals are like multiple choice tests where guessing might succeed, but skipping

01:09guarantees failure.

01:10So models learn to bluff instead of admitting ignorance.

01:14OpenAI's fix?

01:15Change the scoring system.

01:17Think of it like the SAT.

01:19Wrong answers should hurt more than leaving it blank.

01:22And uncertainty should earn partial credit.

01:24If models are rewarded for honesty, they'll stop fabricating just to climb leaderboards.

01:29The takeaway?

01:30Hallucinations aren't going away entirely, but we can train AI to BS less often.

01:35Until then, treat your chatbot like a charming friend who will absolutely lie to your face,

01:40but with style.

01:41The takeaway?

01:42It's a reminder that this is a good story.

01:43The takeaway?

01:44Yes.

01:45It's a good story.

01:46It's a good story, but with style.

01:47I'm not sure I'm sorry.

01:48I'm not sure I'm sorry.

01:49It's a good story.

01:50I'm not sure I'm sorry.

01:51I'm not sure I'm sorry.