OpenAI says new model GPT-4 is more creative and less likely to invent facts

The artificial intelligence research lab OpenAI has released GPT-4, the latest version of the groundbreaking AI system that powers ChatGPT, which it says is more creative, less likely to make up facts, and less biased than its predecessor.

Calling it “our most capable and aligned model yet”, OpenAI cofounder Sam Altman said the new system is a “multimodal” model, which means it can take images as well as text as inputs, letting users ask questions about pictures. The new version can handle massive text inputs and can remember and act on more than 20,000 words at once, letting it take an entire novella as a prompt.

The new model is available today for users of ChatGPT Plus, the paid-for version of the ChatGPT chatbot, which provided some of the training data for the latest release.

OpenAI has also worked with commercial partners to offer GPT-4-powered services. A new subscription tier of language learning app Duolingo, Duolingo Max, will now offer English-speaking users AI-powered conversations in French or Spanish, and can use GPT-4 to explain the mistakes language learners have made. At the other end of the spectrum, payment processing company Stripe is using GPT-4 to answer support questions from corporate users and to help flag potential scammers in the company’s support forums.

“Artificial intelligence has always been a huge part of our strategy,” says Duolingo’s principal product manager Edwin Bodge. “We had been using it for personalizing lessons and running Duolingo English Tests. But there were gaps in a learner’s journey that we wanted to fill: conversation practice, and contextual feedback on mistakes.” The company’s experiments with GPT-4 convinced it that the technology was capable of providing those features, with “ninety-five percent” of the prototype created within a day.

OpenAI claims that GPT-4 fixes or improves upon many of the criticisms that users had with the previous version of its system. As a “large language model”, GPT is trained on vast amounts of data scraped from the internet and attempts to provide responses to sentences and questions that are statistically similar to those that already exist in the real world. But that can mean that it makes up information when it doesn’t know the exact answer – an issue known as “hallucination” – or that it provides upsetting or abusive responses when given the wrong prompts.

By building on conversations users had with ChatGPT, OpenAI says it managed to improve – but not eliminate – those weaknesses in GPT-4, responding sensitively to requests for content such as medical or self-harm advice “29% more often” and wrongly responding to requests for disallowed content 82% less.

GPT-4 will still “hallucinate” facts, however, and OpenAI warns users: “Great care should be taken when using language model outputs, particularly in high-stakes contexts, with the exact protocol (such as human review, grounding with additional context, or avoiding high-stakes uses altogether) matching the needs of a specific use-case.” But it scores “40% higher” on tests intended to measure hallucination, OpenAI says.

The system is particularly good at not lapsing into cliche: older versions of GPT will merrily insist that “you can’t teach an old dog new tricks” is factually accurate, but the newer GPT-4 will correctly tell a user who asks if you can teach an old dog new tricks that “yes, you can”.