ChatGPT probably beat you at the philosophy baccalaureate… and he wasn’t even all the way

This Wednesday, June 14, hundreds of thousands of high school students took the baccalaureate philosophy test. But they weren’t the only ones: a competition between the philosopher Raphaël Enthoven and ChatGPT took place. No suspense: the AI ​​lost, but still got the average.

Is artificial intelligence capable of philosophizing? This is the question we ask ourselves, although, from a technical point of view, this is not at all what it does. Be that as it may, the Paris School of Technology and Business (PST&B) computer and business school organized this Wednesday, June 14 a confrontation between Raphaël Enthoven and ChatGPT on the test of the general baccalaureate in philosophy which took place at the same time. The subject : ” Is happiness a matter of reason?», chosen by the philosopher in the flesh.

How did ChatGPT pass the philosophy baccalaureate

If one simply gave the philosophy baccalaureate question to ChatGPT, it would formulate an answer of around 200 words, free of details, definitions, references and quotations, by formulating a classic outline thesis-antithesis-synthesis , sacrosanct of high school philosophy. Several successive questions were asked to ChatGPT in order to guide its answers and, above all, to deliver a complete copy. Interviewing Mahdi Zargayouna, a computer scientist who formed ChatGPT, The Parisianwrites that ” ChatGPT had weeks of preparation“.

Behind the scenes (not trivial) of the match between the philosopher Raphaël #Enthoven and… GPT chat around the philosophy test of the #Bac2023

👉 #EverythingIsPol with @PaulLarrouturou pic.twitter.com/pGtTHE66Fh

— LCI (@LCI) June 15, 2023

We learn that about twenty lines were given to the OpenAI chatbot for the generation of the response, with the help of an engineer, a technician and philosophers. This is called the prompt engineering, command engineering in French, which refers to all the techniques for creating requests for generative AIs. In a nutshell, it is the “how do you ask him the question“. In a first, the engineers asked ChatGPT for the essential authors around the subject. In the second request, ChatGPT was asked to write the essay, asking him to mention the authors, which he gave in response to the first request.

ChatGPT had to use “an academic and persuasive tone“, while writing”a clear, coherent, fluid, orderly dissertation, successively examining different possible answers», ReportsLe Figaro. To have the most advanced copy, the GPT-4 language model was used, available to those who subscribe to the ChatGPT Plus subscription. It is more efficient than GPT-3.5, used for the free version of the chatbot. The latest version of OpenAI’s language model allows for deeper reasoning and is said to be better for tasks that require creativity.

ChatGPT and Raphaël Enthoven do not talk in the same way at all

After an hour and a half of writing, the twoopponentsreturned their copy. In fact, ChatGPT only took one minute and thirty seconds to “to write» a nine-page dissertation, proofread by the team of experts, who cut a few passages that were too long.

They were then copied by two different people, so that the jury could not determine who had written this or that copy. A jury also composed by Eliette Abécassis, philosopher, and Lev Fraenckel, philosophy teacher in high school and university, better known as Serial Thinker on TikTok.

YouTube linkSubscribe to CssTricks

When rendering scores and conclusions, the jury was very direct: they quickly recognized who had written which copy. For Lev Fraenckel, “ChatGPT does not make problematic, write empty sentences, with approximate quotes“. The chatbot would not have really discussed, argued, reasoned, but rather written beautiful sentences. He even adds thatin the reference to the authors, it is very weak, because there are errors“. For Eliette Abecassis, “ChatGPT returned a copy of History of Philosophy, a catalog with what different philosophers think»

Why Raphaël Enthoven got 20/20 and ChatGPT “that» 11/20

For the philosopher Eliette Abécassis, her colleaguehas developed his thinking superbly» and developed «a philosophical reflection so well thought out, well written, surprising“. Which earned him a logical 20/20. But it should be remembered that Raphaël Enthoven is an associate professor of philosophy and a temporary teacher: he also taught in high school. In addition, several terminale students obtain the maximum mark each year in the discipline.

As for ChatGPT, it only had “that» 11/20, which may not seem like much. However, it is above average, which means that the copy was sufficient with regard to the expectations of the National Education for the baccalaureate test. If ChatGPT made inaccuracies and mistakes that cost it valuable points, it’s because of how it works.

GPT-4 designed by Midjourney

Since he is meant to predict the next word in a sentence or paragraph, he writes down the one that seems most likely to him. This creates a bias, which generates very consensual sentences: it also means that ChatGPT can write false things. This is why he is devoid of opinion on the question, even less than a student doing a philosophy dissertation, scrupulously respecting the methodology.

Strange element, Serial Thinker assures that “in the copy of ChatGPT there is not even a problem“, according to his remarks reported by AFP (viaOBS). However, by asking ChatGPT to write an essay with the subject of the baccalaureate, the chatbot is formulating a problem for us.


Similar Posts