Latest News | Teachers, students, and researchers are forced to update their practices due to the availability of artificial intelligence tools, but lack of knowledge about how new technologies operate is still significant
*By Giordano Benites Tronco
*First published April 13, 2023
*Image: Flávio Dutra – Dall-E/ JU- In the first line, representations about the topic generated by Dall-E, an application that creates images using Artificial intelligence. In the second line, the same images were put through image processing.
In November 2022, the OpenAI organization provided free access to a computer program that speaks like people – and it is open to anyone. You type a question, the program answers – no groundbreaking performance apart from the fact that the answers are good, very good.
The program – called ChatGPT – has the accumulated knowledge of nearly the entire internet and answers complex questions, takes extremely specific steps, creates unpublished stories in the style of established writers and even improves a previously written text – all of that in human natural language.
In universities, ChatGPT created a widely felt earthquake that fostered discussions: are ChatGPT answers a form of plagiarism? How to evaluate students’ knowledge if they can answer assessments using the program? Can researchers write scientific articles in collaboration with ChatGPT? To what extent using artificial intelligence in a field hitherto exclusively human can be considered ethical?
As it could not be otherwise, artificial intelligence tools have also transformed journalistic practices. This, for example, is the first time that the reporter writing to you uses ChatGPT to verify the credibility of information. And this primary contact with this type of technology yielded some interesting insights, including about the very own relationship between humans and machines. For this reason, I have included comments throughout the article regarding the behind-the-scenes of the news report.
How does ChatGPT not work?
First and foremost, let’s be clear: ChatGPT is not Google. Yes, you can insert in the prompt (space intended for the insertion of the text or user question) the same question that you would type in Google. And, yes, the program will generate a response. But while Google offers a list of links among which the user might find their answer, ChatGPT does not clarify where its information comes from.
If pressed to answer, artificial intelligence will say that it has been fed a large amount of data (news articles, books, academic articles, conversations) and that it uses this prior knowledge to generate its answers. This knowledge, however, is based on information available up to 2021.
This is the first common mistake about ChatGPT: it is not constantly updating its database.
Ask about the Covid-19 pandemic and the tool will tell you about the lockdown measures implemented by countries and the first vaccines developed. Ask about who won Big Brother Brazil 2022 and it will not know how to answer. Searching for recent news on ChatGPT is therefore useless.
“Sorry, but as a language model trained in 2021, I have no information about future events”.
ChatGPT
OpenAI’s goal with ChatGPT is to produce an artificial intelligence capable of understanding and communicating with human natural language. It is not a search engine: it is a machine that knows how to speak very well.
The truth is that this technology works more like a text editor than an internet search engine. To be more precise, it is closer to a text corrector.
As it reports Joel Luís Carbonera, professor of the Department of Informatics at Universidade Federal do Rio Grande do Sul (UFRGS), ChatGPT’s artificial intelligence was trained to perform a predictive task – given a set of words, it predicts the next most likely element. It is no different from what happens when you write a text message on your mobile and the app automatically completes the word based on the first letters you type. The difference is that ChatGPT prediction is capable of generating not only words or phrases, but entire texts.
Understand how ChatGPT was trained in the infographic below:
Carbonera explains that ChatGPT generates the most plausible answer from a linguistic point of view, not factual. Therefore, we cannot assume that it is factually correct.
The program’s core commitment is not with the “truth” but with the production of a semantically correct text.
“Here’s the biggest trap: if we use ChatGPT as an oracle, we may not be getting the accurate and factual answers we need.”
Joel Luís Carbonera
Since I was concerned about the program not mentioning sources in its answer to a specific question, I asked ChatGPT which texts from its database had been used to generate that answer.
Very politely, the program apologized and said it was impossible to determine specific texts. “The answers,” it said, “are not based on a single source of information, but rather on a wide range of textual data that the AI model used to learn natural language.”
The program explained to me that ChatGPT is a model of language and not a source of knowledge or information. Its goal is to generate coherent and meaningful responses. It is up to the user to critically verify the information provided.
We then had a heated discussion about the difference between coherence and truth, until I realized that several of my questions began with “please.” As in: “Please, could you give me the answer about x?”. I was subconsciously worried about being kind to the machine. I ended the conversation thinking about how vulnerable human beings are to any simulation of empathy. This observation I preferred not to share with ChatGPT.
Digital hallucinations
Ask ChatGPT something it does not know, and the answers may vary. It can apologize and say that the question refers to something that, for the program, is in the future – that is, after 2021. Regarding controversial matters, it can dodge the question or present multiple existing points of view about the topic. And sometimes it might invent answers.
These texts about things that do not exist or have not happened are called “hallucinations” and are quite common. In an interview with the website Metrópoles, artificial intelligence researcher Anderson Rocha told an anecdotal situation: he asked the program the names of five artificial intelligence professors who had died in the past five years. The program presented the names and described the circumstances of the death of each one of them. However, out of the five allegedly deceased, three are still alive. One of them shares a room with Rocha.
Curiously, hallucinations are very common when academics and scientific papers are involved. There are several reports on the Internet of ChatGPT responses that mention researchers who do not exist or with links to non-existent articles.
Professor Joel Carbonera can’t explain why hallucinations happen, but he has a theory: during the machine’s training, programmers failed to fine-tune the responses, which he calls “putting safety handrails on the experiment.” This adjustment process is meant to prevent artificial intelligence from generating inappropriate replies, such as sexist or prejudicial responses. It also serves to prevent unreasonable responses, such as the so-called hallucinations. According to the teacher, hallucinations happen when the “bugs” exceed the filter acquired in the training.
One of the first uses of ChatGPT for an agenda was in source search. I asked the program to generate a list of researchers that had already published papers about the use or the impact of ChatGPT in the academic environment. Only later, I realized that, obviously, the program would not know the answer, since ChatGPT itself and any papers written about it were released after 2021.
This did not seem to be an obstacle to the program, which immediately delivered the name of seven researchers, the name of the institutions they work for and the subject of their research. I refined the question and asked it to add links to these researchers’ studies.
Several were broken. One of them was directed to research about cell biology.
It wasn’t even from one of the listed authors. It is important to mention that the program crashed several times when this question was asked.
Interestingly, once I have altered minimally the content of the initial question, asking for “Brazilian researchers”, the AI program provided the appropriate answer: there is no information, because the database only goes until the year 2021.
Academic prejudice
The constant occurrence of hallucinations and the lack of source of information practically makes it impossible to use only ChatGPT to write scientific articles. Still, artificial intelligence can be used as a tool by the researcher. It can be used in repetitive tasks that do not require creativity, such as formatting texts, for example.
As soon as the program was launched, the journalist and doctor in communication Marcio Telles started publishing on his Twitter profile a series of tips on how to use ChatGPT to write academic articles, to prepare classes and to other useful aids for researchers. In one of the posts, for example, he demonstrates how to ask the program to organize the references of an academic paper in the APA format, a style of reference citation.
In his posts, Telles points out that ChatGPT is a working tool, and not a replacement for authors. Even after such warning advice, he could not avoid negative reactions on the part of some members of the academic community. Telles declined to be interviewed for this report, claiming that his comments in favor of the use of ChatGPT have not only generated “unease” among colleagues, but also raised suspicion about his own work, what made him decide not to talk about it anymore. Since then, he stopped posting tips on Twitter.
According to Luiza Santos, researcher of artificial intelligence and professor of Communication at the Universidade Estadual de Ponta Grossa (UEPG), this cautious stance of the academic community is historical and comes from Ancient Greece: Plato warned us of the evil effects that books could cause to human memory. The same fear took over with the emergence of computers and even Google. Time has passed, and today these technologies are considered essential to human thinking.
According to the researcher, having a stance of complete “fear” or “fascination” with the academic use of ChatGPT does not help at all. The technology is still new, and a lot of opinions are based upon pure speculation. However, she believes that the community should establish consensus and guidelines. She also mentioned the scientific journal Nature, which published rules for the use of ChatGPT in scientific studies. One of them says that artificial intelligence systems cannot be credited as an author or co-author of a study.
“What ethical rules should the Academic Community follow? We need to establish consensus and guidelines. I believe that consensus will vary according to each field.”
Luiza Santos
Could ChatGPT write a scientific paper for me? I have tested it: I have asked the program to write a ten-page article already formatted in ABNT, containing a literature review on online deliberation in internet forums. The article should have, among other sections, an abstract in Portuguese and English, a conclusion section and another section of references.
A paper titled “findings and trends of online deliberation in Internet forums” was promptly generated. It was much shorter than I expected: it had approximately five pages, counting the bilingual summary. The system already generated the keywords, which was an interesting touch.
Some studies were mentioned in the body of the text but were identified only by the year of publication (“a 2014 study showed that…”). The article ended with a list of references already formatted in ABNT. The years of publication did not match those of the cited studies.
To learn so you do not “get swallowed”
Engineering professor at the Federal Institute of Santa Catarina (IFSC) Jesué Graciliano da Silva remembers the date he first heard of ChatGPT: January 25th. Overnight, that was all anyone was talking about. Most of the news and messages that reached him were catastrophic. All student assessment tests would become obsolete. According to some professors, education would be destroyed.
Jesué had no doubts about how to proceed: he went to test the tool. And he acknowledged: “It was a little bit above my expectations”.
While testing the ChatGPT prompt, he asked the machine questions about its use in education. He liked what he saw. Since he works with distance education and digital learning tools, the teacher has no prejudices against new technologies.
The result of his interaction with artificial intelligence was the digital workbook “Elevando a aprendizagem com ChatGPT (Elevating learning with ChatGPT)”, written by ChatGPT and arranged by Josué. The publication provides a presentation about how ChatGPT operates, its potential and limitations, as well as suggestions for its use in the classroom.
Jesué is not naive: he knows that the program has biases and that it requires attention when handled in the educational context. Therefore, his suggestion is to work with the question, formulating it together with the class and then discussing the answer. If the answer turns out to be wrong, this is also an interesting point for discussion in the classroom.
For those who disapprove of ChatGPT, Jesué has a warning: even if the teacher does not know how to use it, the student does – and the program will be used for plagiarism purposes. It is with this purpose that the guide was organized: to facilitate the transition of teachers to this new scenario of artificial intelligence.
“The more teachers try using it, the better. Teachers need to seek self-improvement. It is their obligation to expand their capacities. And it is up to the educational institution to open the discussion”
Jesué Graciliano da Silva
More than interesting knowledge, more than being good or bad, it is about survival. Jesué understands that ChatGPT will change the educational environment forever. If teachers do not adapt, there may be no place for them in the future.
“Some of my colleagues tell me I am too enthusiastic,” he says. “No, I am not. I want to handle it the best way I can, so I don’t get swallowed up.”
The stance of educational institutions
Consider these two antagonistic cases: as soon as the use of ChatGPT became highly popular – it gained more than 100 million users in the first two months – The New York Department of Education banned its use in the city’s schools. At the same time, technology professor Ronaldo Lemos, who teaches at Schwarzman University, in China, began to force the use of ChatGPT in the tests of his discipline.
What explains the difference between the two approaches? Would it be the bold personality of the Schwarzman University educator versus the slower, more cautious culture of the government? Maybe. However, the question probably boils down to an age factor: the students of the higher education institution are adults; those of the New York schools are children and adolescents.
Leonardo Foletto, editor of the BaixaCultura website and free-culture activist, agrees that ChatGPT should be used cautiously in elementary and high school. According to him, adopting a liberal stance is more difficult when it comes to teaching children at the literacy stage.
Foletto cites three different positions of educational institutions regarding the use of artificial intelligence.
- To ban: first educational institutions try to understand AI before deciding if they are going to allow its use. To Foletto, this stance is problematic, but plausible and understandable in elementary and high school.
- To observe and learn: instead of entirely curtailing the use, rules for using it are agreed upon. Sciences Po, France’s University of political sciences, has banned the use of the tool in the production of written papers or presentations. On the other hand, the institution has made moves to rethink the forms of evaluation alongside teachers, to include artificial intelligence in the proposed exercises and carry out training courses for teachers.
- To use and expand: educational institutions are enthusiasts of ChatGPT as a teaching and learning tool and encourage its use in the classroom.
In the spectrum above, Foletto falls between “to observe and learn” and “to use and expand.”
“It is necessary to prepare teachers to understand ChatGPT and, from there, see how to use the tool in their classes”
Leonardo Foletto
We need to talk about the “P” word
In higher education, the great reticence regarding the use of ChatGPT by students is the issue of plagiarism. With the advent of this new technology, previously existing anti-plagiarism control tools have become obsolete. Such tools work by crossing the content of students’ work with the content of existing pages on the internet, looking for similarities in writing that indicate copying.
This method is ineffective with ChatGPT. The texts generated by the system are not openly available on the internet, and therefore, are not indexed in the Google database. Since every ChatGPT production is unpublished and, in general, very well written, the only person who knows about the non-human origin of the text is the one who requested it from the program.
Another problem with plagiarism is the lack of references. Even if ChatGPT does not specify where it got the information from, it came from somewhere – the tool learned it from millions of texts from its training. So, it is constructing its answers on the intellectual productions of uncredited third parties.
But could it be considered plagiarism? If so, who is being plagiarized? According to Leonardo Foletto, there is no easy answer. The current legal system does not embrace these new issues.
As an enthusiast of free access to culture and information, Foletto advocates the openness of knowledge. ChatGPT provides a paradigm shift, in which access to knowledge is no longer restricted and difficult and becomes abundant, which would be positive. But he knows that this abundance of content can make room for misinformation.
“How can we work with abundance? How can we constitute and take care of it? Because it is not enough to just open [access to knowledge]: it is necessary to take care of the access, since there may be a proliferation of false information”
Leonardo Foletto
OpenAI itself (the company that developed ChatGPT), seeks to be attentive to the discussion of plagiarism with the launch of a scan tool which ranks the probability that the authorship of the text is human or generated by artificial intelligence. The page of the tool, however, warns that it is still limited: the verified texts have a minimum size (thousand characters) and must be in English. Even if these requirements are fulfilled, the result of the analysis is not always correct. If the machine-generated text gets later edited by humans, verification is pretty much useless.
A change in assessment
All interviewees for this report – mostly teachers – were unanimous: plagiarism happens in the classroom with or without ChatGPT. What technology has done is reveal a pre-existing problem: the need to rethink student assessment methods.
According to Leonardo Foletto, it makes no sense for a professor to demand that university students focus “only on form” while writing a text. What really matters is to create relations, to connect contexts with contemporary events.
A popular form of assessment after the emergence of ChatGPT is the one based on question formulation. The logic of a common test is reversed: the student is evaluated for their ability to generate good questions for the ChatGPT prompt. A complementary task would be for the student to evaluate the program’s answer: is it correct? If not, where is the error?
And is it that bad to have access to a machine that can provide textual work? According to Professor Crediné Silva de Menezes, from Faculdade de Educação at UFRGS, artificial intelligence does not steal the intellectual work of humans, but rather optimizes it.
“The machine cannot process beyond syntax. It does not tackle a new problem. Artificial intelligence developers are creating machines that help process information, which allows us to make an intelligence leap”
Crediné Silva de Menezes
According to Crediné, artificial intelligence enables humans to focus on innovation and meaningfulness of experiences while using the machine to expand its natural information processing capacity. With this, education advances.
Those who believe that only students take advantage of artificial intelligence should pay attention to this: teachers could also benefit from a digital helping hand. Professor Crediné can’t wait to use ChatGPT to organize the hundreds of exercises that he has developed over twenty years of teaching. That way, the program can help him select challenging exercises for each class.
Tips for using ChatGPT in Academic contexts
- Ask what you know: Crediné points out that the use of ChatGPT is safer for the user who masters the subject to be researched. He compares using the program to “talking to someone who has read a lot of books.” The user with knowledge will be able to chat with ChatGPT and abstract what is useful from what is doubtful.
- Provide the main text: artificial intelligence works better if you provide the source of information with which it should work. Paste in the prompt a scientific article and give instructions based on the text provided (summarize, create exercises) and the answers are likely to be more accurate.
- Formatting references: it is a great tool to generate and convert references of works in ABNT, APA and others.
- Goodbye, reports: ChatGPT can handle the systematization of research and extension activity reports just like any other document that contains a rigid structure. It is up to the researcher to provide the input.
- Learning to structure papers: the student can ask ChatGPT to write a paper and, instead of handing it to the teacher, observe how the program structured it to then write their own text. “It has a pedagogical function,” explains communication teacher Luiza Santos. “Unless the student is really lazy, it is joint work. We are also learning from this relationship.”
What’s next
The free version of ChatGPT, available on the OpenAI website, operates using GPT-3 technology. The GPT-4 version is already available to ChatGPT Plus subscribers and allows the user to send images to the program prompt while also providing better responses to user inputs.
An OpenAI report revealed that GPT-4 performed well on several academic and professional tests in the United States – including college admission tests and those to practice a profession, such as the OAB (Order of Attorneys of Brazil) exam here in Brazil. In some of them, the program’s score placed among the top ten.
As ChatGPT advances, society works to put barriers in its operation. In addition to an open letter signed by developers and representatives of the technology industry demanding a pause in artificial intelligence research, OpenAI suffered another recent blow: on March 31, the Italian Data Protection Agency and the government of Italy banned the use of ChatGPT in the country until it complies with European Data Protection law.
According to the agency, there is a bug in ChatGPT Plus that makes information such as users’ names and their credit card numbers vulnerable to leaks. OpenAI immediately removed the service from Italy and took a stance in favor of the integrity of the personal information of its users, as well as being in favor of the regulation of artificial intelligence.
The Italian case sheds light on another question: what are ChatGPT and OpenAI doing with all the data generated by users in their interactions with the program? Could ChatGPT use them for other purposes, such as targeted advertising?
It is worth mentioning that, despite its name and the fact that they were born as a non-profit organization, OpenAI became a for-profit company in 2019, open to investments. This had a direct impact on the development of their products: the details about the GTP-4 training are scarce compared to the GTP-3 documentation, precisely in order not to give any advantage to the competition.
And there is plenty: in the path of ChatGPT’s success, Google launched its own artificial intelligence, Bard, in 2023. However, it was launched prematurely: the rush to show results and a disjointed communication about the product delivered a tool which is prone to errors. So far, ChatGPT is still the absolute leader in its segment.
The newcomer with the most potential for growth is Microsoft’s Bing. Once an unpopular alternative to Google search, it now features the Bing Chat, a search tab that uses GPT-4 to communicate with the user using human natural language.
Bing Chat generates responses by combining the GPT-4 engine with web search results. In other words: finally artificial intelligence is connected to the constantly updated database of internet pages. This solves two limitations of traditional ChatGPT: the lack of references and updated information, since now the resulting text includes links to the sites from which the information was taken.
The launch of Bing Chat in such a short amount of time is no coincidence: Microsoft has made an investment of more than 1 billion dollars in OpenAI, and there are ongoing plans for investments over the next few years totaling 10 billion dollars.
According to professor Jesué Graciliano da Silva, this billionaire investment is full proof that OpenAI is no longer free. After all, those who provide its funding hold the power.
“Thanks to this [business] model, it is possible that ChatGPT continues to be free [for the final audience]. But there’s no free lunch. Perhaps in the medium term, we will receive embedded advertising in some form. So, it can also be a way to manipulate people.”
Jesué Graciliano da Silva
ChatGPT: the interview
During the preparation of this report, I had several dialogues with ChatGPT. It’s only fair that the program also gives its side of the story. See below a transcript of the conversation in which a reporter and a computer debate the source of the program’s information, the reason why it did not reveal its references and the difference between “coherence” and “truth”.
Is it impossible to point out the exact source of the information used by ChatGPT to generate its responses?
Yes, it is impossible to point to an exact source of information used by ChatGPT to generate its responses. That occurs because ChatGPT is an artificial intelligence language model trained on a large amount of textual data from various sources, and its response is generated based on a general understanding of natural language rather than a specific source of data. During the training process, the model learns to identify and learn patterns in natural language data, including grammatical structures, word meanings, and relationships between words, and becomes able to apply such knowledge to generate relevant answers to questions.
So, the objective of ChatGPT is to generate a response containing natural language, not a response with the correct information.
Not exactly. The primary goal of ChatGPT is to generate a response that is relevant and appropriate to the question or input provided, based on its general understanding of natural language. The language model is trained to learn the structure of natural language and the relationship between words so it can generate a coherent and meaningful response.
But a coherent and meaningful answer is not synonymous with a true answer. A statement can be coherent and have a meaning, but that meaning could be false.
You are right. It is possible for a response generated by ChatGPT to be coherent and meaningful, but still be false or inaccurate in relation to the facts. The language model has no specific knowledge about any topic or area of study and its ability to provide accurate information depends on the quality and accuracy of the input provided. Therefore, it is important to critically evaluate the responses generated by ChatGPT and verify the information in other sources, especially in cases where accuracy is essential. While ChatGPT can be a useful tool for providing insights and general information about a topic, it is important to remember that it is an automated tool and as such, it cannot replace critical judgment and careful evaluation of the information provided.
How does ChatGPT determine which information from its database will be used to compose a response? I suppose that the database is made up of conflicting information about the same topic.
The selection of specific information to include in the answer depends on several factors, such as the context of the question, the relevance of the information to the question, and the frequency and reliability of the information in the database. In cases where the database contains conflicting information on the same topic, the model may consider multiple sources of information to generate a response or simply indicate that there are different opinions on the subject. However, it is important to remember that ChatGPT has no specific knowledge about any topic or area of study and may not have the ability to evaluate the veracity or accuracy of conflicting or misleading information.
In the previous answer, you mentioned that the selection of specific information to include in the response depends on the reliability of the information in the database. How does ChatGPT measure the reliability of the information in its database?
ChatGPT does not directly evaluate the reliability of the information in its database. Instead, it uses a data-driven approach to learn natural language and identify patterns and relationships between words and phrases. During training, the model is exposed to a large number of texts and learns to recognize which information is most frequent or relevant to a particular topic or context. This means that, in theory, ChatGPT can be exposed to conflicting or misleading information in its database, which can lead to inaccurate or misguided responses.
Translated into English by Lilianne Dias Souza, undergraduate student enrolled in the course “Supervised Translation Training I (English)” of the Undergraduate Program in Language and Literature, under the supervision and translation revision of Professors Elizamari R. Becker (PhD) and Márcia Montenegro Velho (PhD) – IL/UFRGS.
:: Read in portuguese
ChatGPT impacta rotinas na pesquisa e na educação e levanta questionamentos sobre veracidade e metodologias de avaliação