In November 2023 the New York Times he wrote – based on data provided by the startup Vectarafounded by former employees of Google – that ChatGpt’s hallucination rate, i.e. the frequency with which this AI invents content or writes incorrect answers, it is 3%.
It may seem like a laughable percentage, but if you are unlucky enough to encounter a chatbot error – and you are not able to recognize it – the consequences they can be devastating. Think, for example, of the impact that unfounded information could have on the decisions of doctors, managers or analysts.
Il problem of “hallucinations” it’s not just about ChatGpt. All chatbots based on LLM (Large Language Models) occasionally make mistakes.
But only one AI has made it available so far a very effective tool for verifying the reliability of the texts it generates. It’s about Geminithe generative artificial intelligence developed by Google.
Just like ChatGpt, Gemini expresses itself naturally and can satisfy the requests that are usually made to an AI: summarize or reorganize a text, translate a sentence, suggest creative ideas or respond appropriately to an email.
Artificial intelligence The best free apps to generate texts and images with AI by Pier Luigi Pisa 02 January 2024
But Gemini, unlike the free version of ChatGpt, also has access to latest information found on the web. This is important for two reasons: the user can always count on updated answers and, what is in some ways more important, the user himself has the opportunity to ascertain the degree of reliability of the response generata dal chatbot.
We write “degree” of reliability not by chance. The reason will soon become clear.
Below the text generated by Gemini is always there a colored “G”.i.e. the Google symbol.
By clicking this “G” you start a check of the response produced by the chatbot. Gemini, in practice, perform a “traditional” web search to find information that can confirm or refute the AI-generated text.
When the search is completed, Gemini colors a few sentences of the response it generated green or orange.
The sentences in green contain “verified” information, of which Gemini found a match on the web.
Just click them to get the link to a source deemed reliable which confirms what was generated by AI. But Google warns that the proposed link “is not necessarily the one used by Gemini to generate its original response.”
The orange colored phrases, on the other hand, are those that “probably” differ from the (correct) information retrieved by Google online. There is also the possibility that Google did not find any relevant content related to the orange sentences, therefore dubious. Again a link is provided if available.
In short, orange it’s a warning sign: something in the text that generated the AI doesn’t add up, according to Google, and further verification is appropriate.
Finally, it may happen that various portions of the text are not coloured neither green nor orange: it means that “there is insufficient information to evaluate the declarations [del chatbot, nda] or the statements are not intended to provide objective information.” In this case, of course, it is a must a verification by the user of the information produced by Gemini.
It must be said, in conclusion, that other tools equipped with AI capable of accessing information on the web, such as Microsoft Copiloto Perplexityhelp the user to verify the texts generated by offering, under each answer, a selection of links useful for deepening or independently verifying the topic in question.