ChatGPT still prone to errors, best used as 'second opinion': OpenAI
text_fields
New Delhi: OpenAI’s latest language model, GPT-5, may be more advanced and accurate than its predecessors, but the company has cautioned users against treating ChatGPT as their primary source of information.
Nick Turley, Head of ChatGPT, said the AI tool should be used as a “second opinion” since it remains prone to errors despite significant improvements. In an interview with The Verge, he acknowledged that GPT-5 continues to grapple with “hallucinations”, instances where the system generates content that appears credible but is factually incorrect.
OpenAI claims to have substantially reduced such errors, yet the model still delivers wrong answers around 10 per cent of the time. Turley emphasised that achieving complete reliability is an immense challenge.
“Until we are provably more reliable than a human expert across all domains, we’ll continue to advise users to double-check the answers,” he said. “I think people are going to continue to leverage ChatGPT as a second opinion, versus necessarily their primary source of fact.”
Large language models like GPT-5 are designed to predict words based on patterns in massive datasets. While this enables them to generate fluent and natural-sounding responses, it also means they can provide incorrect information on unfamiliar topics.
To counter this, OpenAI has integrated search capabilities into ChatGPT, giving users the option to cross-check results with external sources. Turley expressed optimism that the hallucination problem will eventually be resolved, but cautioned that the solution will not be immediate.
“I’m confident we’ll eventually solve hallucinations, and I’m confident we’re not going to do it in the next quarter,” he remarked.
Meanwhile, OpenAI continues to expand its ambitions. Reports indicate that the company is working on developing its own browser, while CEO Sam Altman has even suggested that OpenAI could consider acquiring Google Chrome if it were ever available for sale.
With IANS inputs