Did Google use some data from ChatGPT to train its own AI, Bard? This is what transpires from an investigation by the American media The Information. Google denies.
The war between the giants of the net in artificial intelligence has perhaps its first major controversy. In an article published on March 29, The Information tells how Google is mobilizing its troops internally to face OpenAI, the company whose products — ChatGPT, DALL-E — are in the news.
A controversy that almost overshadows the heart of the site’s investigation, which focuses on how the two divisions of the Mountain View firm most advanced in AI were brought to collaborate when they were historically rivals. Cooperation illustrates the old adage: the enemy of my enemy is my friend, especially if it is an outside competitor.
Google Brain specializes in AI frameworks, like TensorFlow, and learning methods, like AutoML. Its tools have made it possible to design technologies like BERT, used with the search engine. For its part, DeepMind is a startup acquired by Google, whose great feat was to beat one of the best days of go.
The two companies have been somewhat pressured to work together, according to The Information, through a new initiative called Gemini. In the line of sight, OpenAI, whose two main products have been in the spotlight for several months — and more particularly ChatGPT, whose specialty is the creation of texts on demand.
ChatGPT, Bard’s model in spite of himself?
However, in this effort to deceive, Google is implicated: the group would have gone so far as to train its own conversational agent, Bard, with certain data from ChatGPT – using this information found on ShareGPT, a site in which we can share the “craziest” conversations with ChatGPT. There are more than 116,000 listed.
Contacted by The Verge, Google rejects these allegations outright. “Bard is not trained on data from ShareGPT or ChatGPT,” a spokesperson said. But has this been the case in the past? On this second question, the company evaded. She only commented on the present time, not on what may have happened in the past.
The Conversation says an AI engineer working at Google ended up joining OpenAI. Before switching companies, he reportedly sought to warn his ex-employer not to dig into ChatGPT’s data, as it would risk violating the terms of service — and risk generating responses that were too similar.
Also according to the site, Google would have actually used data, then stopped after these repeated warnings. It’s not clear what happened to the data used previously — whether it was retained or deleted after it was used to train Bard. The chatbot is based on LaMDA, a language model identical to GPT-3.5 or GPT-4.
Bard’s beginnings were difficult. Presentation of the chatbot has been minimal, its accuracy is questionable and its availability is limited. The tool, released much later, has some difficulty existing against ChatGPT, which is omnipresent in people’s minds. Above all, Google does not seem eager to go in this direction, for business model reasons, unlike its rival, Microsoft, which mobilizes colossal means, because it sees a strategic opportunity in it.