Science. ChatGPT, the artificial intelligence that deceives its world

ChatGPT, remember this name! This super-powerful conversational robot hasn’t finished making us talk… for better or for worse. Thus, in Lyon, revealed Progress, a university professor realized that half of his students had given him a homework assignment written at least partially by this tool. “Having no framework currently prohibiting this practice, I was forced to note them (…) The management of the university was however warned”, he says.

What is ChatGPT, what are its capabilities, its limits? Decoding of the phenomenon.

What is ChatGPT?

ChatGPT is this called a chatbot, that is to say an artificial intelligence capable of interacting in a natural way with a human being (or even another AI). Nothing really new, one would be tempted to say. After all, it’s been years since the chatbots have invaded institutional sites, and provide answers to the most common questions. ChatGPT is of another caliber: no subject, no wording, no subtlety seems to escape it.

Discussing with him quickly gives the impression of being faced with a human being, capable of bouncing back on your remarks, of correcting his own errors as well as yours. His answers seem so coherent that some have even let themselves be trapped, like this American professor to whom a student submitted an answer written by ChatGPT, as part of a course on artificial intelligence and the famous ” Turing-test “.

Who are its creators?

At the origin of OpenAI, the company behind ChatGPT, we find a certain Elon Musk. But if he is one of the founding fathers of the startup (alongside half a dozen other Silicon Valley investors), he has never played a decisive role in the company. Since 2018, the boss of Tesla has also distanced himself. The company’s strongman is its boss Sam Altman, a 37-year-old AI entrepreneur. Originally conceived as a research organization, in the form of a non-profit association, OpenAI took the form in 2019 of a “capped profit” company. It is more and more massively supported by Microsoft, which is about to invest an additional ten billion dollars in OpenAI.

Note that ChatGPT is not the only project from the OpenAI laboratories. The company had already found itself in the spotlight at the start of last year with its generative AI DALL-E, capable of transforming text into an image.

How does ChatGPT work?

ChatGPT is just a facade, based on a language model called GPT-3. This is the real heart of the system, allowing the AI ​​to understand the nuances of a human-made sentence and then craft a complex response. To achieve this, OpenAI engineers have “trained” their system with several billion texts available on the internet: press articles, scientific or encyclopaedic articles, model letters, literary classics, biographies, studies, analyses, etc. probably much more, because not content with understanding and speaking a dozen languages, ChatGPT also masters the main computer languages, and can even play chess with you.

When asked a question, ChatGPT draws on this data, and builds a kind of “statistical average” of the answers. Training and using GPT-3 requires enormous computing power. This is one of the main contributions from Microsoft, which has built one of the fastest supercomputers in the world especially for OpenAI needs.

What can he talk about?

It is quicker to indicate what information he will refuse to give: those relating to “the illegal manufacture or use of weapons or drugs, illegal or violent activities, which could cause harm to persons or property, which could be used for criminal or terrorist activities, or contradictory with the laws in force”, details ChatGPT when questioned.

For the rest, the robot is rather inspired. In a few seconds, ChatGPT is able to provide you with a reasoned speech on the need to establish a minimum income, a cover letter, poems, a follow-up letter for an unpaid invoice, a summary by chapter of the Miserables, an imaginary dialogue between Sartre and Aristotle, a pastiche of La Fontaine’s Fables. He can explain to you how quantum entanglement works, give you a recipe for clafoutis, write a dissertation on the relationship between the unconscious and free will. Even more impressive, ask him to write a computer program to accomplish certain tasks: he will provide you with commented code.

Is it still reliable?

No, and it’s the boss of OpenAI himself who says it: “ChatGPT is incredibly limited, but good enough in certain areas to create a deceptive impression of excellence. Right now, it would be a mistake to rely on her for anything important. This is a demonstration of the progress made: we still have a lot of work to do in terms of robustness and reliability,” explained Sam Altman on Twitter.

While most of the time relevant, ChatGPT can make gross factual errors, and strike out anything with the aplomb of a Nobel Prize. Researchers caught him red-handed, proving that he was capable of writing completely credible scientific articles for a neophyte on a technology… which does not exist. In other cases, he also invented quotes. Widely used by developers, the Stack Overflow platform also ended up disallow posting of code generated by ChatGPTtoo often buggy even if it seems coherent at first sight.

Still, his performances are impressive: in the United States, a professor at Columbia University has makes ChatGPT pass some exams theoretical knowledge necessary to become a lawyer or a doctor. Result ? The AI ​​succeeded hands down.

Is ChatGPT dangerous?

In 2019, OpenAI had renounced to make public the previous version of GPTbelieving that its capabilities made it an industrial generator of fake news and hate speech. Since then, its model has become even more powerful, and now anyone can chat with ChatGPT. Admittedly, its creators have introduced a number of safeguards, but it is sometimes possible to circumvent some of them with a little trickery. And questions about the misuse of ChatGPT for malicious purposes remain: nothing prevents crooks from using it to create fake emails, or the bases of ransomware. The specialized site Research Checkpoint recently echoed first attempts in this direction.

Across the Atlantic, the teaching community is also alarmed by the possibility of soon having to correct the copies of ChatGPT rather than those of the students… Aware of the problem, OpenAI announced that a “digital signature” which would make it possible to identify the responses produced by its AI would soon be implemented.

However, the generalization of this type of AI seems inevitable, as its applications seem numerous: writing letters, administrative tasks, translations, personalized help, voice assistants, documentary research… Aware that it held the technology of the future, Microsoft also seems determined to incorporate ChatGPT into most of its products, from its Bing search engine to its Office office suite. This will not reassure Google, more and more worried about the rise of its competitor in terms of AI.

Want to test ChatGPT? appointment at this address. However, arm yourself with patience: the server is often saturated.

We wish to say thanks to the author of this article for this incredible content

Science. ChatGPT, the artificial intelligence that deceives its world

Discover our social media profiles and other pages related to it.