Owing to ChatGPT’s growing influence, Google announced its AI chatbot Bard, which is based on the LAMDA language model
By Ankita Chakravarti: Google’s AI assistant Bard, which was created to take on ChatGPT has made shocking claims about how it was trained. Bard in yet another incident has revealed that it was trained with users’ gmail data but Google does not agree with it. Owing to ChatGPT’s growing influence, Google announced its AI chatbot Bard, which is based on the LAMDA language model. Bard has similar functionalities as the ChatGPT, so therefore there are high possibilities that the AI chat model may hallucinate and present fiction as facts.
Bard’s controversial answer was highlighted by Microsoft researcher Kate Carwford. She shared the screenshot of her conversation with Google’s AI chatbot, in which she asked the chatbot about the dataset. To which Bard replied that its dataset comes from a variety of sources. It listed publicly available datasets, which includes data sets of text and code from the web from various sources including Wikipedia, Github and Stack overflow. Bard then listed Google’s internal data, which included Google Search, Gmail and other products and data from third party companies. Now if it is true that Bard uses data from Gmail, it is a serious breach of privacy.
However, Google has rubbished the report and issued a clarification on Twitter. Reacting to Crawford’s post, Google tweeted from its official account saying, “Bard is an early experiment based on Large Language Models and will make mistakes. It is not trained on Gmail data.
The generative AI tool like Bard or ChatGPT, will not always provide you with a factually correct data. A lot of times, the AI tools cook up stories and present them as facts. This is something even the companies agree with. Google Bard website also warns users that AI chatbot will not always get the facts right. Similar thing ChatGPT’s parent company OpenAI had said when it rolled out the GPT4 language model, which is a successor to the GPT 3, which powers the ChatGPT.
Highlighting the limitations of GPT 4, OpenAI in its blog stated, “Despite its capabilities, GPT-4 has similar limitations as earlier GPT models. Most importantly, it is still not fully reliable. It “hallucinates” facts and makes reasoning errors,” The company also asked users to be careful when using language model outputs, particularly in “high-stakes contexts, with the exact protocol (such as human review, grounding with additional context, or avoiding high-stakes uses altogether) matching the needs of a specific use-case.”