The scheduled coronation for May 6 was marked on the calendar; however, ChatGPT surprisingly concluded that the appropriate date would be May 19. This type of error, known as "hallucinations," is one of the notable limitations of AI systems like GPT-4.

When presenting the GPT-4 version, OpenAI acknowledged these shortcomings, noting that they are continuously working to address them, along with other issues such as social biases and contradictory instructions. However, this situation is not unique to OpenAI, as other AI systems, like Google's Bard, also face similar challenges.

Recently, journalists from the New York Times tested ChatGPT by asking for information about the first article published by the newspaper on artificial intelligence. The responses provided by the chatbot varied, some even including incorrect data, highlighting the presence of these "hallucinations."

The technology that powers chatbots, known as large language models (LLMs), is based on analyzing vast amounts of digital text to learn its capabilities. It functions similarly to an autocomplete tool, predicting the next word in a sequence based on patterns found in the analyzed data.

However, due to the abundance of false information on the internet, LLMs can assimilate and repeat these falsehoods, and even invent incorrect information on occasion. This situation presents significant challenges in the development and implementation of AI systems, requiring ongoing efforts to improve their accuracy and reliability.

Take It With Caution

Generative AI algorithms and reinforcement learning have the ability to analyze large amounts of online information in a matter of seconds and produce new text, generally coherent and well-written. However, experts caution that these results should be taken with caution.

Both Google and OpenAI have urged users to be cautious when interacting with AI systems. OpenAI, in collaboration with Microsoft and its Bing search engine, acknowledges that its GPT-4 model has a tendency to "hallucinate," that is, to generate nonsensical or false content based on certain sources.

This tendency can be problematic, as AI models are becoming increasingly convincing and credible, which may lead users to place too much trust in them. Therefore, users are warned not to blindly trust the answers provided by these systems, especially in areas that involve critical aspects of life, such as medical or legal advice.

OpenAI has implemented a series of methods to mitigate these "hallucinations" in user responses. This includes review by real people to avoid incorrect data, gender or race biases, and the spread of fake news. These measures aim to ensure that the responses generated by AI are accurate, reliable, and ethically responsible.