The scandal is potentially huge. Kate Crawford posted on Twitter (see screenshot below) a response by Bard, Google’s conversational agent using generative AI, to a seemingly innocuous question, “Where does Bard’s data come from?” Although Bard’s answer may be a hallucination, it is important to remember that Google has demonstrated its ability to disregard the most basic rules of privacy and personal data protection in the past. I would not be surprised if there were some truth in this story, even if it was denied by Google relatively quickly (see tweet below). Remember that until 2017 Google was scanning its users’ emails to propose contextual ads. Google stopped this highly intrusive practice not to scare its paying customers because of legal actions, including the Matera V Google case. Nevertheless, Google has not stopped analyzing your emails. Features are still deployed in gmail today that use unstructured data contained in emails received via Gmail: adding events to your calendar predictive typing proposing reminders What is clear is that this hallucination of Bard should alert us to the possible dangers of such technological advances. What scares me is not the progress itself but the human beings behind that progress and their choices. In the case of generative AI and the LLMs that form its foundations, the training data worries me the most. Especially since the explanations given by OpenAI at the release of GPT-4 are anything but transparent about the data used to train this new model. Risks of using conversational agents The risks are mainly related to violations of the laws governing the secrecy of correspondence. This was already at the heart of the Matera vs. Google case. Quite extraordinarily, the amicable agreement between the parties was rejected by the judges because the measures proposed by Google were insufficient to guarantee that the causes of the problem would be corrected. This intrusion into the secrecy of correspondence poses an even greater risk for companies: the respect of business secrecy. Can you imagine for a second that your confidential conversations (potentially protected by an NDA) could be used as a training corpus for artificial intelligence? A little further on in this article, I talk about a well-known example. Imagine for a moment this scenario. An investment bank is working on a project to buy company A from company B. The exchanges serve as a training ground for Bard. 10,000 km away, an analyst asks Bard...