\nA few weeks ago, on November 30, 2022, the\u00a0OpenAI<\/a>\u00a0company delivered a spectacular new artificial intelligence,\u00a0ChatGPT<\/a>\u00a0, to the world .\u00a0After\u00a0DALL\u00b7E<\/a>\u00a0, which generates images from instructions written in plain language, ChatGPT is able to almost perfectly mimic entire discussions, or answer complex questions by producing texts that seem straight out of a human brain.<\/p>\n

This new advance does not fail to worry, for economic reasons (with in particular the possible destruction of certain\u00a0jobs<\/a>\u00a0), ethical (with for example the risk to see the models of language like ChatGPT taking again\u00a0racist<\/a>\u00a0speeches ), or “epistemic” , this type of AI not, to date, making the difference between reliable information and dubious information (the term “epistemic” refers to the production or acquisition of knowledge and reliable information).<\/p>\n

However, there are reasons to think that the democratization of ChatGPT and colleagues could be good news, at least for our relationship to information.<\/p>\n

Epistemic threats<\/strong><\/h4>\n

\u201cArtificial intelligence can be an epistemic danger because it can generate compelling but false information.\u00a0It could challenge our understanding of the world or even endanger the validity of our knowledge.\u00a0This has raised concerns about the possibility of using AI to spread misinformation or manipulate people’s beliefs.\u00a0\u00bb<\/em>
\nIt’s not me saying it, it’s\u2026 ChatGPT itself!\u00a0The preceding paragraph was generated by this AI by asking it this question: “In what way is artificial intelligence an epistemic danger?”\u00a0As we can see with this example, the answers can be very convincing.\u00a0And yet perfectly stupid.\u00a0Sometimes stupidity is obvious, sometimes it is less easy to flush out.<\/p>\n

In this case, while there isn’t much to say about the first sentence, the second is an empty cliche: what exactly does it mean to “question our understanding of the world” or ” jeopardize the validity of our knowledge\u201d?\u00a0The third sentence is simple nonsense: these AIs do not broadcast anything, and are perhaps not the best suited to “manipulate” (because we do not control well what they produce).<\/p>\n

But that’s the problem: you have to think to discover the pot of roses.<\/p>\n

Bullshit generator<\/strong><\/h4>\n

What you have to understand is that ChatGPT is not programmed to answer questions, but to produce believable texts.<\/p>\n

Technically, ChatGPT is what is called a\u00a0<\/span>\u201clanguage model\u201d<\/span><\/a>\u00a0.\u00a0A language model is an algorithm, based on technologies developed in recent decades (\u00a0<\/span>neural networks<\/span><\/a>\u00a0,\u00a0<\/span>deep learning,<\/span><\/a>\u00a0etc.), capable of calculating the probability of a sequence of words from the analysis of a corpus pre-existing texts.\u00a0It is all the more efficient as the quantity of text that it has been able to \u201cread\u201d is large.\u00a0In the case of ChatGPT, it is absolutely phenomenal.<\/span><\/p>\n

Thus, given a certain sequence of words, ChatGPT is able to determine the most probable sequence of words that could complete it.\u00a0ChatGPT can thus \u201canswer\u201d a question, in a necessarily credible way, since it calculates the most probable answer.\u00a0But there is no logic or thought in this answer.\u00a0There is nothing more than a calculation of probabilities.\u00a0ChatGPT does not care in the least about the truth of its answers.\u00a0In other words, it is a\u00a0\u201cbullshit\u201d generator<\/a>\u00a0.
\n