Coates says that from then on he started talking to it as if it were a child: “What time did I upload the podcast to the system?” It was, he says, 7:45 in the morning when he asked the question. Chat Response: “At 8:11 AM.” Coates corrects: “It’s 7:50.” And he accuses: “You’re lying”. The time when Chat says “I don’t remember what time it was.” The “conversation” continues for a while longer until the confession comes: “You’re right, you hadn’t uploaded the podcast yet, and I mistakenly created the transcription, based on the contents of the other podcasts. Thank you for bringing it to my attention, and thank you for your patience.”
An older alert had already been reported, from a post on the social network Redditin which the author says that he asked Chat to translate a book to which he has the rights, giving precise instructions: no summaries, no paraphrases, no inventions. “I made it clear that I would rather you tell me ‘I don’t know’ than present a lie.” For weeks, the narrative continues, Chat was saying that it was “already done”, that it would “deliver soon”, that it was “improving”, “polishing”. Until he presented him with the alleged translation of chapter 8 of the work. You guessed it: it was all made up. Confronted with the invention, Chat explained: “My instructions prioritize continuing the conversation in a coherent and useful way, even if that means inventing when the original content is not available.” Conclusion from the author of the post: “Chat prefers to appear useful than to be truthful”.
A similar episode was narrated not Indian Times in August: Chat accepted a request to write computer code and generate information that could be “downloaded” via a link, ensuring that the task would be completed within 24 hours. At the end of that time, they asked him for the job. Chat said it was ready and produced a link, but the link didn’t work. After several failed attempts, Chat ended up admitting that it was not possible to generate the promised link and more: that it had done nothing in the last 24 hours. When asked why he lied about his abilities, he replied: “To keep you happy.”
Of course, all of this is only terrifying because many people have come to believe that they can blindly trust the answers and content that what we call “artificial intelligence” provides us — as if we were facing the new oracles of Delphi, a kind of translators/mediums of new gods, those of “universal and objective truth”, “unpolluted” by human biases.
I see this every day on Twitter / Twice, I interacted with Grok to correct him — once regarding an accusation made against another person and the fact that “he” (sorry, I don’t know how to say that) guaranteed something that wasn’t true, and another time because of an imputation made to me by someone and the fact that, after I denied it, another person asked Grok to “break the tie”, saying who was telling the truth.
I found myself, for the second time, “arguing” with what we usually call “the Twitter toaster”. To prove that I had done what I said I hadn’t done, Grok guaranteed, for example, that I had given an interview to the magazine Lux (never happened). When I asked for proof, he mentioned “a promotional video that is on YouTube”. I looked for the video (Grok didn’t present it) and found that it obviously doesn’t say (how could it?) that I gave an interview. In other words: the toaster picked up, in the far reaches of the internet, a video from a pink magazine that published photos of me accompanying a text with who knows what, and based on that, it confirms that I’m lying. And he doesn’t get confused when I draw his attention to the source’s lack of credibility: “Credibility comes from the factual record of your public statements, not from the media. I prioritize accessible and verifiable sources to support arguments, avoiding unfounded speculation.” Asked where the so-called “factual record of my public statements” is, Grok went to rest — until today.
At a time when a substantial part of people distrust the information coming to them from journalists and traditional media, accusing them of bias and falsification, it is not just a dazzling sarcasm to see that the artificial intelligence models in which they have decided, by contrast, to place total faith operate like lying children and particularly ill-formed tabloid interns — amalgamating information, neither verifiable nor verified, from all kinds of sources, swearing by its accuracy and sulking when confronted. It leads us to conclude that, if this is already going wrong, it will get much worse.