If the algorithm ChatGPT was a student at the Aristotle University of Thessaloniki (AUTH), at which point he could boast of having passed with a very good grade (8 out of 10) the demanding digital imaging course in the winter semester exams at the Department of Computer Science.

ChatGPT, which in just two months after its launch “won” 100 million active users worldwide and combined the title of the fastest growing application in history (Google and Facebook took five years, and TikTok nine months), took out the creators of the “white man” in the context of a related experiment conducted by Professor of Computer Science Ioannis Pitas, President of the International Academy for Doctoral Studies in Artificial Intelligence (AIDA).

“We sent requests to ChatGPT and received responses. Except for some hiccups, the algorithm worked very well. I would rate it an 8 out of 10. Some rather surprising findings were that ChatGPT can mimic mathematical reasoning quite well, for example. in 2D systems z and Fourier transform and complex numbers. He can also represent the properties of physical phenomena quite well with some “reasoning” – with or without quotation marks – which is beyond the scope of a successful classical text composition. ChatGPT can still provide scheduling solutions, eg. to calculate the arithmetic mean and the output of a moving average filter in Python,” says Mr. Pitas of APE-MPE.

At the end of January last year, the same algorithm, the “child” of the American OpenAI, in which Microsoft announced that it would invest $ 10 billion, passed (albeit with a low score, C +) final exams at Minnesota Law University. The School, after answering topics – a total of 95 multiple-choice questions and 12 essay questions – from constitutional law to taxation and torts. Shortly thereafter, he “scored” over 50% on the United States Medical Licensure Exam (USMLE).

The risk of awarding degrees to non-existent students

Until recently, notes Mr. Pitas, “we thought that applications like ChatGPT could be used mainly in the humanities and medicine. After this experiment, we saw that the algorithm can also work very well in the sciences related to mathematics and engineering, and this highlights an important problem that arises in education as well. We should not be technophobes, as this is a useful technology, if used correctly, the rewards will be much greater than the risks, and in the medium to long term will have a very positive impact on work and markets, but action is certainly needed. should be taken, even the extinguishing of a fire, in the sense that it must be immediate.

According to Mr. Pita, the first issue that everyone involved in education needs to pay attention to is exams and assignments: “In Greek universities, where exams are taken in person and there is personal control, the problem is not acute. However, if ChatGPT had fallen into the Corona virus conditions, things would have been very different. The same applies to universities working with distance learning, since now the question will arise not only about the personality of the examiner, which already existed, but also about the content. If such universities have flexible admissions policies and flexible assignment/examination controls, they may end up awarding degrees to non-existent students. Who answered the exam questions or wrote the paper?

Student or algorithm? Today’s digital tools do not seem easy to check copy, so it requires the personal work of a professor or other person specifically to check copy, which is almost impossible, given that there are several hundred professors in each university, and many thousands of students. I think the real issue is how can we use generative DNNs to our advantage in education. This deserves careful study and could lead to revolutionary innovations in education,” he emphasizes.

He adds that systems such as ChatGPT must remain open to the international scientific community as to how they work so that precious gray matter is not lost when trying to decipher their secrets, sometimes by reverse engineering to make a product look like it. that already exists. , but we don’t have enough information on how to do it). But isn’t it in the legitimate interest of a company developing such a product to keep its secrets secret?

“Because of their enormous social impact, and for maximum socio-economic progress, the advanced core technologies of artificial intelligence systems must be opened up. It is possible that the management of such powerful tools as ChatGPT is not a matter of legislation, but a matter of policy applied by each company. AI-related data needs to be at least partially democratized. Therefore, appropriate robust financial compensation systems should be in place for champions in AI technology to compensate for any loss of profit due to the open nature of code and data and ensure significant future investment in research and development in artificial intelligence (AI),” notes .

The biggest threat

To achieve positive social impact, especially creative AI and big language models like ChatGPT, the president of AIDA puts forward additional proposals based on key assumptions: “In my opinion, the biggest threat today comes from the fact that such AI systems can remotely deceive an extremely large number of ordinary citizens with low (or secondary) education and/or little investigative skills. This can be extremely dangerous for democracy and any form of social and economic progress. Another big threat lies in their use in illegal activities: cheating in university exams is a rather “soft” use of them compared to other opportunities they provide for criminal activity. To do this, in accordance with international law, AI systems must be registered in the “AI systems registry” and inform their users that they are talking to the AI ​​system or using its results,” he concludes.

MONKEY BEE