Impressed by the current version of internet chatbot ChatGPT? Well, prepare yourself for a new artificial intelligence milestone.


That's according to ChatGPT creator OpenAI, who has just unveiled GPT-4, a major upgrade to the current version of the online service GPT-3. The company says the latest model is not only more creative, but more intelligent – and less likely to invent facts.

The upgraded chatbot comes with several new features. While its predecessor only responds to text prompts, GPT-4 can also interpret images, allowing users to ask questions about pictures they post. This means a user, for instance, can post a picture of ingredients and GPT-4 can suggest a recipe that will use them all.

The new ChatGPT model – initially only available to subscribers of ChatGPT Plus (a service with a $20 monthly fee) – can process up to 25,000 words at a time, eight times more than the previous version.

Additionally, OpenAI has announced they are partnering GPT-4 with several commercial services, including the language learning app Duolingo, which will use the technology to offer personalised language lessons.

Launched in November 2022, ChatGPT became a viral sensation after it offered human-like responses to user text prompts. The technology is capable of writing jokes, essays, news articles and full movie scripts.

The service is currently incorporated into Bing, the search engine owned by Microsoft. The tech giant has already invested $10bn into OpenAI.

With GPT-3's user base growing rapidly, concerns have been raised about its reliability: the technology does not comment on its own accuracy and could be used to create misinformation.

Seemingly in response to these worries, OpenAI claims that it has spent six months improving safety features for GPT-4, saying it is 40 per more likely to produce factual responses than GPT-3.

However, the technology can still 'hallucinate', effectively inventing facts or providing offensive answers when responding to some prompts.

As OpenAI says: “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.”

GPT-4 does not learn from experience and has a limited context window', meaning it may not remember information from previous user prompts.

OpenAI is also responsible for programmes such as Dall-E 2, Point-E, which generate images based on user text prompts. Other text-to-image AI services, such as Midjourney, NightCafe and Craiyon are available.

Read more:



Thomas Ling
Thomas LingDigital editor, BBC Science Focus

Thomas is Digital editor at BBC Science Focus. Writing about everything from cosmology to anthropology, he specialises in the latest psychology, health and neuroscience discoveries. Thomas has a Masters degree (distinction) in Magazine Journalism from the University of Sheffield and has written for Men’s Health, Vice and Radio Times. He has been shortlisted as the New Digital Talent of the Year at the national magazine Professional Publishers Association (PPA) awards. Also working in academia, Thomas has lectured on the topic of journalism to undergraduate and postgraduate students at The University of Sheffield.