How do you write an article for the PÚBLICO newspaper about ChatGPT? Here’s what he would recommend (pictured below).
Fantastic! ChatGPT is an artificial intelligence generation tool, that is, a tool from a field of artificial intelligence dedicated to the study of algorithms capable of generating original content such as text, images, music, etc.
Interestingly, there are a number of statistical insights related to how ChatGPT works. This AI platform is based on an “auto-regression” text generation model, in which each word is generated taking into account the previous ones, there is also a statistical distribution based on the randomization of the word generation process and is further adjusted. a statistical regression model known as a neural network.
For additional details on how ChatGPT performs statistically, I recommend reading Wolfram’s CEO’s blog. But we don’t know exactly everything about its current operation, we know more about the previous versions and the reason for this is as follows. In 2019, OpenAI released the code and data on which its statistical model, then called GPT-2, would have been trained. But since secrecy is the lifeblood of the business, when OpenAI released GPT-3 in 2020, it became less transparent, did not reveal the codes – as it had done in the past – and only gave Microsoft an exclusive license to use GPT-3.
While this is clearly an extremely useful and powerful work tool, I understand that many of our readers may also be concerned at this time about its potential misuse – for example, in terms of plagiarism, in the context of teaching or press. The good news is that there are already many tools being developed to mitigate these concerns. For example, the GPTZero tool was created for this purpose, which allows to classify whether or not a text was created using artificial intelligence techniques.
According to the creators of GPTZero, this platform correctly classifies 99% of human-written articles and 85% of AI-generated articles. But there are several questions that remain unanswered: will it implementation Is the classification of tools like GPTZero the same in English and Portuguese? And how implementation Does it vary from subject to subject? For example, it will implementation Is GPTZero ranked better in philosophy than in history?
Speaking of history, I would now like to go back in time to give another perspective on these matters. First, let’s go back to the 19th century to observe that the search for automations to create text predates the history of the modern computer. For example, inventor John Clark (1785-1853) built the Eureka, which is a machine that has been producing Latin poems since at least 1845! The machine was repaired in 2015 by the University of Exeter, is still in operation and is still being studied by academics looking into the workings of this mechanical way of writing poetry that offers us a bridge between science and art.
Second, let us now go back to the middle of the twentieth century. More specifically, let’s travel to 1954, the year that the famous book How to lie with statistics, in which Darrell Huff recalled how critical it is to develop a critical sense in order to be alert to falsified data or misleadingly presented evidence. But if chatGPT is based on statistics, do Darrell Huff’s principles apply to this AI platform as well? Let’s see how ChatGPT reacts when you ask something about the author of this article (pictured below).
This text shows well how it is necessary to be alert, as we begin to enter the realm of so-called hallucinations – which is the technical term of genetic artificial intelligence used to describe the reveries of these algorithms. For example, the third paragraph does not correspond to the truth, since I do not have a program in RTP 2, nor do I have a blog in the newspaper. Express. In other words, this is a modern way of lying with statistics, a way that I’m sure even Darrell Huff wouldn’t imagine!
These tools are very welcome, opening up a world of opportunities for students and faculty, as well as employers and employees. But we must continue to invest heavily in training and literacy on these issues – taking into account the educational, ethical and legal aspects involved.
The views are those of the author and do not represent necessarily those of the University of Edinburgh and the Portuguese Statistical Society
The author writes according to the new spelling convention