Zap // Rawpixel; Pixabay
As general artificial intelligence has become increasingly popular, sometimes the truth distorts. Lies – or hallucinations, as they are known in the technology industry – have improved, but the latest models are counteracting this trend.
In the years that followed the appearance of the ChatGPT and the growing integration of bots Ia in a series of tasks, there is still no way to ensure that these systems produce accurate information.
Indeed, current AI bots do not decide – and cannot decide – which is true and what is falseexplains the.
Ultimately, the hallucinations problem It seems to be getting worse as technologies become more powerful. The reasoning models, considered the “most recent and powerful technologies” of companies such as OpenAi, Google and startup Chinese DeepSeek, are “generate more errors, no less“.
According to, the mathematical skills of the models “improved remarkably”, but their “ability to deal with the facts has become more unstable. It’s not totally clear why”.
Reasoning models are a large type of language model designed to perform complex tasks. Instead of “just producing text based on statistical probability models”, reasoning models “divide questions or tasks into individual stages similar to the process of human thinking,” said the PC Gamer.
During the tests of its latest OpenAi reasoning systems, the company found that its System O3 hallucinating 33% of the time When performing the test benchmark PersonQAwhich involves answering questions about public figures.
This value is More than double the hallucination rate From the previous reasoning system of OpenAi, the O1. The latest tool, the O4-Mini, hallucinated at an even higher rate of 48%.
OpenAi rejected the idea that reasoning models suffer from increased hallucination rates and stated that further research is needed to understand the conclusions.
Hallucinations are not “intrinsically more prevalent in reasoning models.” However, OpenAi is “working actively to reduce the highest hallucination rates we observed in O3 and O4-Mini,” he said Gaby RailaPorta-Voz da company. Too many “unwanted dreams of robots“.
For some experts, hallucinations seem inherent in technology. Despite the best efforts of companies, the I was going to “always have hallucinations”these Amr AwadallahExecutive Director of startup de ia vectara and former executive of Google. “This will never disappear.”
Still, hallucinations cause a “serious problem for anyone who uses technology with legal documents, medical information or confidential trade data,” said Times.
“A long time spent trying to find out what answers are factuals and which ones are not“, These Practical givingco -founder and executive director of Okahu, a company that helps companies deal with hallucinating problems. Do not deal with these mistakes “eliminates the value of AI systemswhich should automate tasks. ”
Companies are “fighting to find out exactly why reason the chatbots are generating more errors than before ” – a struggle that“ highlights the intriguing fact That not even the creators of IA understand well how technology really works, ”says Futurism.
Recent worrying trend of hallucinations “It challenges the generalized assumption of the industry that AI models will become more powerful and reliable as they are expanded.”
Whatever the truth, AI models need to “largely eliminate the absurdity and lies if they want to be as useful as their advocates today imagine,” said PC Gamer.
It is already “difficult to trust the results of any LLM”, and almost all the data “have to be carefully verified twice“This is acceptable for some tasks, but when the goal is to” save time or work “, the need to” check meticulously and confirm the facts produced by IA eventually contradicts the goal of its use. “
It is not clear if OpenAi and the rest of the LLM industry “will be able to control all these unwanted dreams with robots.”
Teresa Oliveira Campos, Zap //
