In testing by this masthead, GPT-3.5 struggled with instructions to draw pictures using text characters known as ASCII. It would produce a rabbit when asked for a person, for example. GPT-4 did not have that issue, producing a proficient stick figure, but mangled requests for more advanced images such as a dragon. It was an illustration of both its improvement but limitations.

While the two versions can appear similar in casual conversation, “the difference comes out when the complexity of the task reaches a sufficient threshold”, OpenAI said, noting “GPT-4 is more reliable, creative, and able to handle much more nuanced instructions”.

An online demonstration of the technology by Greg Brockman, OpenAI’s president, showed it could take a photo of a hand-drawn mock-up for a simple website and create a real website based on it. GPT-4 also could help individuals calculate their taxes, the demonstration showed. Education company Khan Academy is testing an AI-powered tutor using GPT-4, and payments firm Stripe is using it in its customer help tools.

But GPT-4 does not have access to real-time information, so it cannot explain current events such as the collapse of Silicon Valley Bank this month.

Sam Altman, OpenAI’s chief executive, on Twitter called GPT-4 its model “most capable and aligned” with human values and intent, though “it is still flawed”.

GPT-4 is 82 per cent less likely to respond to requests for disallowed content than its predecessor, and scores 40 per cent higher on certain tests of factuality, the company said. Inaccurate responses known as “hallucinations” have been a challenge for many AI programs.

With Reuters.

Get news and reviews on technology, gadgets and gaming in our Technology newsletter every Friday. Sign up here.