What to do about factual errors written by artificial intelligences like ChatGPT? Google is trying to provide some form of response, this Tuesday, September 19, in the new version of Bard, its conversational robot competing with the famous software launched by OpenAI at the end of 2O22. The principle consists of launching a parallel query on the search engine to “check” whether the content of the response produced by artificial intelligence (AI) seems to be matched – or not – by website pages dealing with the same subjects.
“We know that major models of language processing present challenges, particularly because they are capable of presenting false information in an affirmative tone.”, recognizes Google in its press release. The company is well aware of the problem: Bard was mocked by Internet users for having, in one of the demos presented at its launch, wrongly attributed the first observation of a planet outside the solar system to the James-Webb telescope.
These kinds of errors, often called “hallucinations” in the profession, are actually common. ChatGPT, like its derivative version integrated by Microsoft into its Bing search engine, also often got dates or facts wrong, invented book titles, etc.
Green or red
“It’s important that you can have more confidence in AI-generated information”, writes Google. Indeed, this lack of factual reliability, if it does not pose problems for certain uses of creative writing or fiction, is an obstacle to the adoption of language processing models by companies. With the new version of Bard, you will be able to “check the answers provided by finding similar or different web content”explains Google.
Bard has been accessible since July in Europe, through a home page separate from that of Google. The user can, as with ChatGPT, type in a written instruction and get a response, but then has the option to press a “Google it” button. Then, the search engine will launch an – invisible – query on each sentence or paragraph of the response. If the content of the web pages appearing as a result appears to confirm the content of a passage generated by Bard, it will be underlined in green; if, on the contrary, the search seems to contradict the AI, the extract will appear in red. By clicking on a paragraph, the Internet user will be able to access an extract of the pages found matching Bard’s assertions.
You have 42.97% of this article left to read. The rest is reserved for subscribers.