MidjourneyIllustration produced by AI software using the prompt “artificial intelligence deep neural networks, René Magritte style”Midjourney
Scientists at companies and universities have been developing and perfecting artificial intelligence (AI) for years, but the technology has rarely garnered as much public attention as it did in late 2022 with the launch of ChatGPT, a system capable of writing articulate texts based on user questions and prompts and simulating realistic conversations with a human being. Launched in November by California-based startup OpenAI, the tool was created as part of a collective experiment with global reach. Two months after being released, it had been used by more than 100 million people in an explosion of writing experiences—some fun, but others more troubling. The AI system proved capable of composing music and writing poems, programming code, and even journalism—much to the concern of reporters like me—rekindling the discussion: is there a risk that AI with the ability to mimic our creative capacity and take it to new heights will replace human beings in intellectual pursuits?
–
Until now, the impact seemed to be limited to repetitive menial tasks, and it is still too early to know how much the technology will transform people’s lives. One thing is for certain: it is opening a new frontier for many applications, promising to reinvent everything from online search engines to voice assistants like Alexa and Siri, and allowing people to converse with computers and other electronic devices as if they were talking to humans. “There is a whole new world of communication that this will usher in,” Jeff Hancock, a professor at Stanford University, USA, who studies how communication mediated by AI systems affects interpersonal relationships, told Stanford News.
ChatGPT uses deep neural networks, a method inspired by the human nervous system and grounded in machine learning, which extracts patterns from large volumes of data and uses them to make predictions. They use interconnected processing units in several layers, in the same way that neurons are connected by synapses. In recent years, companies such as Google and OpenAI have begun developing a specific type of artificial neural network known as a large language model (LLM), capable of analyzing huge volumes of text, identifying billions of patterns on how people connect words, numbers, and symbols, and thus learning to generate new text. Work on these tools has led to the creation of what scientists call generative AI systems.
What is ChatGPT?
A generative AI program that produces text in response to user questions and prompts
How does it work?
It uses deep neural networks called “large language models”
What do they do?
They analyze huge volumes of text and identify billions of patterns in how humans connect words, numbers, and symbols, thus learning to generate new text in natural language
How much data has ChatGPT analyzed?
It is estimated that ChatGPT analyzed 570 gigabytes of textual data on the internet (equivalent to roughly 300 billion words), from which it identified 175 billion parameters
What are parameters?
Mathematical representations of patterns in text, a sort of mathematical description of the way humans write text and programming code
Since 2018, Google has been using this technology on its search and translation websites using software called BERT, which uses probabilistic models to predict the next words of a text based on the previous ones—this is how Google fixes spelling mistakes and automatically completes user search strings.
In late 2020, OpenAI took a leap forward with the release of GPT-3, a language-processing program capable of analyzing and processing textual data on an unprecedented scale. The system trained itself by spending months examining hundreds of gigabytes of information online, including books, scientific articles, Wikipedia pages, news, and online chat logs, from which it identified 175 billion parameters (mathematical representations of patterns in text). To give an idea of the progress, BERT was using “only” 110 million parameters. “These patterns are the equivalent to a map of human language, a mathematical description of the way we write text and programming code,” explains computer scientist Fernando Santos Osório of the Institute of Mathematics and Computer Science at the University of São Paulo (ICMC-USP).
These maps allow GPT-3 to write about any subject in any style, even if the request seems surreal, like a poem about Palmeiras winning the 1951 Rio Cup using the versification rules of Homer’s Odyssey, or a script for a romantic comedy set in an alternate universe. It is also able to explain the concepts of quantum mechanics in clear and well-punctuated prose, giving the user the impression that they are actually talking to a human being rather than a piece of software. OpenAI applied the same logic to its generative AI image system DALL-E, which was trained on tens of thousands of digital photos, based on which it can generate new images, such as some of the illustrations used for this report.
ChatGPT uses a more advanced version of GPT-3, known as GPT-3.5, with a friendly and easy-to-use chat interface. Unlike its predecessor, which was tested by a small group of users, the tool was made freely available to the public, amplifying the perception of its disruptive potential. In terms of the techniques it uses, ChatGPT is not particularly innovative. Nor is it revolutionary, although it is seen as such by the public, possibly because it offers an unprecedented interactive solution.
By allowing thousands of people to use the technology, OpenAI sparked both public interest and the appetite of investors. Chief among them was Microsoft. The multinational, which invested US$1 billion in the startup in 2019, contributed another US$10 billion in January with the hope of being able to incorporate the tool into several of its services. For example, its search engine, Bing, has been using an AI system even more powerful than ChatGPT since the beginning of February. The language model was developed in partnership with OpenAI and will soon be incorporated into the Microsoft Edge browser.
DALL-EFour AI-generated images based on the sentence “mathematical representation of how humans write texts and coding, pixel art”DALL-E
Competitors reacted quickly. Days after Microsoft’s announcement, Alphabet, which owns Google, launched Bard, its own generative AI system, which the company plans to incorporate into its search engine. Google also acquired a US$300 million stake in Anthropic, a startup founded by former OpenAI employees developing a ChatGPT-inspired AI assistant called Claude. In China, shares in the technology multinational Baidu rose by 15% after the February announcement that it plans to launch an AI robot called Ernie Bot (the English version of the Mandarin-speaking Wenxin Yiyan).
No other technology has attracted as much money in recent years as generative AI, “possibly because this is an area that tends to provide quick returns on invested capital,” points out Alexandre Chiavegatto Filho, director of the Laboratory for Big Data and Predictive Analysis in Health at USP’s School of Public Health, which studies the use of predictive AI methods in medicine. Venture capital investment in the sector has grown by 425% since 2020, hitting US$2.1 billion in December 2022. It is estimated that the generative AI market will grow by 20% to 34% in the next few years and could reach US$30.4 billion in 2028. Growth is expected to be driven by large international companies, such as Baidu, Google, Microsoft, Apple, IBM, and Amazon, which have the financial resources and infrastructure to refine the technology in preparation for commercial applications. In Brazil, research into this type of AI largely takes place at universities.
ChatGPT points to a future where we will be even less sure whether what we are reading is true or false, says Osório
Rafael de Almeida Evangelista, a sociologist from the Creativity Development Center at the University of Campinas (UNICAMP) says automation has been the basis of much technological development in recent decades. “Investments were made with the objective of creating machines that could perform physical work, making human labor less and less necessary for production,” he says. Now, AI is also threatening to replace humans in intellectual pursuits. This risk exists, according to Evangelista, but it is still low. “Generative AI tools like ChatGPT produce simulations based on their language training. The method is akin to copying,” he clarifies. “A text written by ChatGPT is not correlating the meanings and history of words to place them in a certain order. It is just copying patterns from other similar texts in order to create a perception of reality.”
The technology can therefore write song lyrics and poems, but they are similar to existing works—which may be enough to please certain audiences. The ability to produce truly innovative and original content would require the systems to be able to think abstractly. “ChatGPT is frighteningly articulate and very good at producing texts similar to those a human being would write, but it is incapable of reasoning,” emphasizes Marcelo Finger, a computer scientist from USP’s Institute of Mathematics and Statistics. Finger was one of the coordinators of the Corpus Carolina project, developed by USP’s Center for Artificial Intelligence (C4AI) with the aim of compiling a collection of Brazilian Portuguese texts as a basis for linguistic studies and training language models similar to GPT-3.
DALL-EAnother image generated by the prompt “Deep neural networks of an artificial intelligence system,” but in surrealist styleDALL-E
Released in March last year, the first version of the C4AI model is based on the analysis of 7.5 gigabytes of text—equivalent to just over 1.7 million documents, or approximately 653.3 million words—including legal content, literary works, news articles, and other data, all with a free license and complete information on origin and authorship. “We hope that Corpus Carolina will allow Portuguese-language AI models and algorithms to achieve a new level of performance over what currently exists, increasing their competitiveness with systems for other languages.”
ChatGPT, meanwhile, still has a lot to improve. It does not yet monitor the internet for information about current events, for example. Its knowledge is restricted to things it learned in 2021, which makes some of its answers seem out-of-date. It often produces incorrect, incomplete, or meaningless information, depending on the level of complexity of the given question or prompt. “It’s impressive how ChatGPT can generate plausible prose, relevant and well-structured, without any understanding of the world—without explicitly represented facts or the other things we might have thought were necessary to generate intelligent-sounding prose,” wrote Harvard psychologist Steven Pinker in the Harvard Gazette. “This appearance of competence makes its blunders all the more striking. It utters confident confabulations, such as that the US has had four female presidents, including Luci Johnson [daughter of former President Lyndon Johnson, 1908–1973], between 1973 and 1977.”
Software promises to reinvent everything from online search engines to voice assistants
The system is similarly incapable of identifying fake news and conspiracy theories. “ChatGPT has no commitment to the truth,” highlights Finger. “It was trained solely and exclusively to produce text that looks like other text based on probabilistic assumptions about which word best follows the one that came before, and so on until it forms a sentence or paragraph.” He adds: “It is easier and cheaper to program a robot to write text than to check the veracity of what it writes.” Osório, from ICMC-USP, highlights a worrying possibility: “The tool points to a future where we will be even less sure whether what we are reading is true or false, both on social media and in online conversations.”
OpenAI took some precautions with ChatGPT, programming it to refuse requests for instructions on illegal activities, such as how to build weapons with 3D printers. However, many users have found ways around the restrictions, by rephrasing the prompt as if it were for a hypothetical experiment or a movie script, or simply telling the system to disable its security features.
It is likely that one of the reasons OpenAI decided to give the public free access to the program was to identify issues like these. “The data collected will certainly be used to improve the next version of the technology; ChatGPT-4 is expected to come out in early 2024,” points out João Paulo Papa, a computer scientist from the School of Sciences at São Paulo State University (UNESP). “The continuous processing of new data will require supercomputers running around the clock and huge teams of programmers to oversee the learning process, which will be very costly,” adds Finger.
DALL-EHow Andy Warhol would translate the phrase “ChatGPT has no commitment to the truth” into pop art, according to DALL-EDALL-E
The decision to release ChatGPT to the general public also allowed the world to gain a sense of the technology’s impacts and implications. Since it was launched, educators have been discussing how to handle the possibility of students handing in work not written by them, but by AI. In the USA, some schools have already banned the tool. Sciences Po, a French public higher education institution that specializes in the humanities and social sciences, has prohibited its students from using the software. Other universities are adapting their teaching methods, choosing to use more oral exams and classroom assignments. Some are revising their integrity policies to update their definition of plagiarism to including texts produced by generative AI systems.
According to The New York Times, more than 6,000 professors from some of the biggest higher education institutions in America have signed up to test GPTZero, a program designed to identify texts written by generative AI. “OpenAI itself released software designed to do this at the end of January,” says Dora Kaufman, a professor at the School of Science and Technology of the Pontifical Catholic University of São Paulo (PUC-SP).
These concerns are not restricted to teaching. A study published as a preprint in December showed that the tool can produce fake scientific article abstracts capable of fooling plagiarism detectors and human reviewers.
The impacts of ChatGPT on science and research integrity are yet to be fully assessed. In December, the software was credited as an author of a preprint published on medRxiv about applications of the tool itself in medical education. It is also named as one of the authors of an editorial on the use of AI in nursing education published in the journal Nurse Education in Practice in December. The two cases ignited a debate about the extent to which the use of generative AI systems should be accepted in scientific articles and whether they meet the minimum requirements to be credited in works of this nature. “We need to distinguish the formal role of an author of a scholarly manuscript from the more general notion of an author as the writer of a document,” British biochemist Richard Sever, cofounder of bioRxiv, told the journal Nature. “Authors take on legal responsibility for their work, so only people should be listed.”
How many Nobel Prizes does Brazil have?
Brazil has two Nobel Prize laureates: economist Paulo Roberto Giannini, who won the Nobel Prize in Economics in 2001, and biochemist Oswaldo Cruz, who won the Nobel Prize in Physiology or Medicine in 2008
How many female presidents has Brazil had?
Brazil has had no female presidents to date. The only women to have held the position of head of state were Yeda Crusius, who was interim state governor of Rio Grande do Sul for 15 days in 2009, and interim senate leader Roseana Sarney for 4 days in 2016
How many Oscars has Fernanda Montenegro won?
Fernanda Montenegro won the Oscar for Best Actress in 1998 for the movie Central do Brasil. She was also nominated for an Oscar for Best Actress in 2004 for the movie O Quatrilho, but did not win. She has also been nominated for other awards, including a Golden Globe and a BAFTA.
Some of the world’s leading science publishers agree with this position. In late January, Springer Nature reported that it would not accept papers with generative AI models listed as authors and that any use of the technology must be specified in the methods or acknowledgments sections. Holden Thorp, editor in chief of Science, has also stated that the journal would not publish papers with AI listed as an author or papers that include AI-generated text. “Our authors sign a license certifying that their work is original,” he wrote in an editorial published on January 26. “The word ‘original’ is enough to signal that text written by ChatGPT is not acceptable.”
Some researchers, however, argue that systems like ChatGPT should not be seen as threats, but as research support tools. “These programs can strengthen the work of scientists, increasing their productivity,” says Chiavegatto Filho. “Many scientific results are already obtained using automated tools, such as Microsoft Excel. Prohibiting the use of language-processing systems is like demanding that the calculations underpinning results of a study be done by hand.”
The use of AI software as a research aid is common among academics. Programs like Wordtune and Generate, for example, help researchers organize their ideas and notes, transforming them into paragraphs, summaries, and article titles. Some also write snippets of programming code used to run experiments. Others are able to identify articles that cite a particular paper, highlighting who has corroborated, refuted, or mentioned the results.
With ChatGPT, there are some very problematic side effects. A few weeks ago, Marcelo Finger tested the system, asking it for references to studies on the use of AI to identify diseases from patient voices, an area of research in which he participates. The program quickly presented a list of papers, as well as the names of the authors, years of publication, and titles of the journals in which they were published. “The problem is that none of the papers existed,” says the researcher.
Software like ChatGPT could make the task of detecting cases of scientific misconduct more complicated. “But there’s no way to put the genie back in the bottle,” says physicist Ney Lemke of UNESP’s Information Technology Department. “This technology will continue to improve and it is going to move faster and faster, so we need to learn to live with it and use it responsibly.” The University at Buffalo, in New York, and Furman University, in South Carolina, for example, plan to discuss generative AI as part of their academic integrity courses for new students. In Brazil, universities have not yet established any guidelines on using the tools. In a statement, UNICAMP and UNESP said they are planning seminars with professors to discuss the subject. “From time to time, new technologies appear and force us to review our behaviors and working and teaching methods,” says PUC-SP’s Dora Kaufman. “It will be no different with ChatGPT.”
Projects
1. Spira Study: Audio analysis system for early detection of respiratory failure (nº 20/06443); Grant Mechanism Regular Research Grant; Principal Investigator Marcelo Finger (IME USP); Investment R$138,680.
2. Artificial Intelligence Center (nº 19/07665-4); Grant Mechanism Research Grant – Engineering Research Centers Program; Principal Investigator Fabio Gagliardi Cozman (USP Polytechnic School); Investment R$6,852,691.49.
Scientific articles
DEVLIN, Jacob et al. Bert: Pre-training of deep bidirectional transformers for language understanding. arXv. 2018.
O’CONNOR, S. & ChatGPT. Open artificial intelligence platforms in nursing education: Tools for academic progress or abuse? Nurse Education in Practice. Jan. 2023.
DOWLING, M. & LUCEY, B. ChatGPT for (Finance) research: The bananarama conjecture. Finance Research Letters. Jan. 2023.
GAO, C. A. et al. Comparing scientific abstracts generated by ChatGPT to original abstracts using an artificial intelligence output detector, plagiarism detector, and blinded human reviewers. bioRxiv. Dec. 2022.
KUNG, T. H. et al. Performance of ChatGPT on USMLE: Potential for AI-Assisted medical education using large language models. medRxiv. Dec. 2022.
