Artificial intelligence is making things up - can this problem be fixed?

Spend a lot of time using ChatGPT, or “chatbot” of artificial intelligence and it won't take you a long time to realize that they're telling lies.

Described as hallucinations, false information, or invention of things, these have now become a problem for every business, organization and high school student using artificial intelligence generation systems to develop documents. Some are using artificial intelligence systems for tasks that potentially have major consequences, ranging from psychotherapy to research on the drafting of legal documents.

I don't think you can find any pattern today that doesn't suffer from any hallucination”, said Daniela Amodei, cofounder and leader of the Anthropic, creative “chatbot” Claude 2.

“[ The Chatbots are designed to predict the next word”, Amodei said. “Therefore, there is a degree in which the model does this incorrectly”.

Anthropic, the founding company of ChatGPT, OpenAI, and other major developers of artificial intelligence systems, also known as major linguistic models, have said they are working to make these systems more reliable.

How long will this take and if they are good enough to be considered safe for medical advice, it remains to be seen.

This cannot be fixed”, said Emily Bender, professor of linguistics and leader of the Computer Linguistics Laboratory at Washington University. “is a difference between technology and the cases in which technology” is used.

Many rely on the reliability of artificial intelligence generation technology. The McKinsey World Institute predicts that this technology will add the world economy to an equivalent of 2.6 trillion to $4.4 trillion. <x) Almost all these tools include several language components.

Google has only started to present an artificial intelligence product that writes news to media organizations, and that's why accuracy is paramount. The Associated Press, too, examines access to technology as part of its partnership with OpenAI, which is paying to use some of the Associated Press texts archive to improve its artificial intelligence system.

In partnership with hotels management institutes in India, computer scientist Ganesh Bagler has been working for years to obtain artificial intelligence systems, including a predecessor of the ChatGPT, in order to create new recipes for South Asian cuisine. A single “halinant” component can make the difference between a tasty and intangible food.

When OpenAI chief Sam Altman visited India in June, Bagler, who is a professor at the Indrapasthra Institute for Information Technology in Dehli, he raised several questions.

“I think hallucinations in ChatGPT are still acceptable, but when a recipe is affected by this, then it turns into a big problem”, Bagler said, as he asked Almat at an event in New Delhi.

“

Altman expressed optimism that this problem could be fixed.

I think we will greatly improve the hallucination problem”, Altman said. “I think it'll take us a year and a half, two years. Something like that. But then we won't talk about these problems. There's going to be a balance between creativity and perfect precision and the model needs to be learned when we want one, or the other”.

But for some experts studying technology, such as languager Bender from Washington University, these improvements will not be enough.

Bender described the language model as a system for “modeling lines of words”, offering some written data on which the model is trained.

This is how spell checking programs are able to find out if a word is incorrectly written. This model also helps strengthen automatic translation and transliteration services, “making the result appear more like a typical text in a particular language”, Bender said. Many people rely on a version of this technology every time they use the “autocomplete” when sending messages or emails.

But, the “chatbots -- the last”, like ChatGPT, Claude 2 or that Bard of Google Company -- tend to go one step further, generating new textile paragraphs. However, Bender said these “chatbot” constantly select only more reliable words to unify the sentence.

When they are used to generate texts, language models “are designed to invent things. That's all it does to”, Bender said. This technology of artificial intelligence is good for imitating forms of writing, legal contracts, and television software scripts.

But since these models just invent things, when the text that creates turns out to be something that we consider to be accurate, it's coincidence”, Bender said. “Even if programmed to be more accurate, there are still ways in which they can fail, and failures will likely occur when it is harder for a person who reads the text to note them”.

These mistakes are not a big problem for marketing areas that have required Jasper AI services to write ideas, said company president Shane Orlick.

“Hakcinations are actually an additional” bond, Orlick said. “We have consumers who all the time show us how they came to ideas, how Jasper handles them from different angles, the way they would never think themselves”.

This Texas headquarters start co-operates with companies like Open It, Anthropic, Google or Facebook, to offer consumers a set of linguistic models of artificial intelligence for their needs. For someone concerned about accuracy, this start-up offers the Anthhopic model, meanwhile, for someone concerned about the safety of their source data is offered another model, Orlick said.

Orlick said he knows hallucinations can easily improve. He expects from companies like Google, which according to him, “have high standards for accurate 48x1> for his research machine to work on finding the solution to this problem.

I think they can fix this problem”, Orlick said. “They must address this. I don't know if it's ever going to be perfect, but I believe it will constantly improve”.

Technology optimists, like Microsoft company cofounder Bill Gates, have predicted an optimistic future.

“I am optimistic that over time artificial intelligence models can be taught to distinguish truth from falsehood”, Gates said in a July post concerning his thoughts regarding the dangers in society that present artificial intelligence.

He cited a 2022 text of the Open Hei as an example “that there will be good work in this direction”. Finally, researchers at the Federal Institute for Technology in Cyril, Switzerland, said they have developed a method to detect some, but not all, the hallucinating content of ChatGPT, and, according to them, the method will automatically remove this content.

But even Altman, who creates products for various uses, does not expect models to be true if he asks for information.

Perhaps I believe the answers I receive from ChatGPT as much as anyone on Earth believes”, Altman said during the incident held at a university in New Delhi. / REL