Jack Loughran Wed 4 Mar 2026
Collected at: https://eandt.theiet.org/2026/03/04/openai-promises-25-drop-hallucinations-gpt-53-launch
OpenAI has released GPT-5.3, with a focus on delivering more accurate answers and better contextualised results when searching the web.
ChatGPT-5 was first released in August last year and was touted as a major upgrade and a “significant leap in technology” with fewer hallucinations than previous versions.
One of the major flaws with AI chatbots, which include other models such as Google’s Gemini and Anthropic’s Claude, is their tendency to prioritise providing users with a satisfying response, even if the information given is heavily flawed or incorrect – hence hallucinations.
While GPT-5.1 and 5.2 were both positioned as releases that would improve accuracy, users continued to complain that it all too frequently responded with incorrect answers. With GPT-5.3, OpenAI claims to have reduced this error rate by over 25% as it shifts the model’s focus from speed to accuracy. It will also take a more natural and less “cringey” tone than previous models that had irritated some users.
The GPT-5.3 update has been designed to improve how the model integrates web results into its answers, balancing external information with built-in knowledge to give more relevant responses. This is particularly relevant given a growing movement to optimise content for AI in a bid to drive users towards particular content, or even force it to provide incorrect information.
A BBC investigation recently saw a journalist publish a completely fabricated blog post about himself and, within 24 hours, both ChatGPT and Gemini were repeating that false information as fact.
So far, GPT-5.3 will only be available on the Instant models, with OpenAI saying it plans to update the Thinking and Pro to 5.3 “soon.”
While OpenAI claims the new model has made meaningful progress on everyday usability, the firm admitted that there was still more optimisation work needed, especially the responses given to non-English users, and concerns that the tone could still need some tweaking in the future.
GPT-6 is not expected until mid-to-late 2026 at the earliest, but analysts say it will place a greater focus on remembering users’ preferences and providing more autonomous assistance outside of the chat interactions.
Leave a Reply