Information on tide times is easily available online – a search engine will throw up multiple websites all giving the same, correct information.
However, when the BBC asked ChatGPT the same question as the unlucky visitors, it generated exactly the same, incorrect, response.
On another attempt it was out by five hours.
Open AI, which owns ChatGPT, said improving factual accuracy was a “significant focus” for the industry, and new models had significantly fewer hallucinations, external but still occurred.
Prof Steven Schockaert, head of the AI and data analytics research section at Cardiff University’s school of computer science, said in this instance the problem may lie with how the information is presented online.
“I think the issue here is that the tide tables are shown in a table, and the model struggles with extracting the correct value from the table,” he explains.
“Traditionally tables have been quite challenging for language models because they’ve been trained on language, so words [written] left to right. Tables don’t have that kind of structure.”