What is ChatGPT, the revolutionary artificial intelligence system that worries some experts

On November 30, as the world was getting into the New Year spirit with its eyes focused on the World Cup in Qatar, a program that was soon considered a step forward in the advancement of artificial intelligence debuted: ChatGPT.

This new system can generate written content that is highly coherent and very similar to that created by humans.

Despite the errors and shortcomings that still exist, the capacity already demonstrated by the program, and its potential to improve in the long term, arouses not only admiration, but also talso fears.

Just look at the number of reviews they associate ChatGPT with the word “threat”. Many feel that the program seems too convincing by mimicking human speech and anticipate trouble.

They are questions about the possibility of a strong rupture in areas such as creativity, learning and education, work, digital security and democracy itself, as stated in a recent article by The New York Times.

According to the authors, what used to be a person expressing an opinion may now be just a robot that artificially generates an argument.

What is the program?

ChatGPT is basically a virtual robot (chatbot) that answers a variety of questions, completes written tasks, converses fluently, and even gives advice on personal problems (although it is noted that it is not intended for this purpose).

Your possibilities of generating content are immense.

He can, for example, teach you how to prepare a stroganoff -and with a different twist on the recipe, if you specify it-, give you advice on getting a job, writing poetry, academic papers and also writing a reconciliation letter to a friend you care about. you have distanced

ChatGPT responds in a matter of seconds to a random request on how to “write a poem about artificial intelligence in the style of Carlos Drummond de Andrade”; the quality of the result, depending on who criticizes it, can be debatable.

According to SEO.ai tests, ChatGPT is available in almost 100 languagesbut model performance varies by language (works best in English).

The system was developed by OpenAI, a company founded in 2015 in the US by Sam Altman (today its main figure) and by the omnipresent Elon Musk (who left it in 2018 considering that there was a conflict of interest with his main company, Tesla).

Five days after launch, ChatGPT reached over 1 million users (who, ironically, have to be human; you have to prove you’re not a robot when logging in). These interactions are being used to train and develop the model.

OpenAI says that use will be free and open to all during this “test and research” stage, leading experts to speculate about future types of monetization for the tool.

The company also warns that, in that period, the software “may occasionally generate incorrect or misleading information” and that its data history is limited to 2021.

Despite being pointed out as a possible threat to Google’s hegemony as a provider of information on the Internet, the system still makes serious mistakes, such as saying that Brazil has won at least five Oscars (in fact, the country has never won a statuette). ).

Why is ChatGPT considered a breakthrough for AI?

Powerful text-based artificial intelligence programs work by storing gigantic amounts of data (with emphasis on words and conversations in this case) and using algorithms to predict the best formulation of a sentence.

These are called big language models (LLM).

Unifesp professor Álvaro Machado Días, a neuroscientist, explains that during the training of the software, to the chatbot you are asked questions like “what is a cylinder?” and the technicians elaborate their own answers.

“If the response of chatbot is not valid, the correct ones are inserted into the system to teach it. This is passed to other situations automatically.

Despite already using a modality that manages to understand the context of the use of words, allowing better concatenation of texts, previous programs did not respond as well to the user or still sounded very artificial.

ChatGPT learned to speak in a way closest to a human.

Machado Días says that what sets this program apart is the use of a technique that understands how language works: learning reinforcement through human feedback (RLHF).

Engineers apply ‘reward’ and ‘punish’ methods to teach the system the most desirable forms of interaction. It is a fine tuning process.

“In practice, engineers rank the responses given by the algorithm according to their relevance and encourage the program to learn the preferences listed in the ranking to increase the relevance of the textual outputs. The result is in the form of texts that seem deeper and more significant than those of the alternatives”, says the Unifesp professor.

ChatGPT has also been trained to admit mistakes, challenge incorrect assumptions, and reject inappropriate requests.

But a professor at the University of California managed to get the system to write programming code to say that only white or Asian men make good scientists.

OpenAI claims that although it has made efforts in this regard, the program “will sometimes respond to problematic instructions or exhibit biased behavior”.

According to the company, the data collected in this testing stage will be used to improve the system.

Is it a threat to learning and creativity?

The threat of changes or alterations is already hanging over work and employment. Fields that rely on text, like journalism, could change a lot, and job openings could be gone forever.

ChatGPT’s competition in code generation is also raising questions in a relatively new sector: programming.

But one of the areas that has been realizing the possible problems of ChatGPT is precisely one of the most affected by the arrival of new technologies: education.

The temptation among students to use the program to find ready answers to their assignments led New York to make a quick decision: just a month after its debut, the system was banned from schools and public devices in the American city.

OpenAI has been working on a kind of flag to identify that the content originated from ChatGPT. Algorithms already exist that calculate with enough precision the probability that a chatbot have made a text.

In addition to ‘copy and paste’, there is the fear of structural impacts on human learning. For example, will the cognitive exercise of writing an essay with a beginning, middle and end, linking ideas coherently, be affected?

“I am very concerned about the algorithmization of thought, which is the alteration of our understanding and relationship with the world due to interaction with AI,” says Machado Días, from Unifesp.

“I think this will be the biggest mindset shift in all of modern history. It is worth noting that the human brain has been slowly reducing in size, as a result of technological development, for more than a thousand years, from a technical-cultural point of view, but also more limited from a neurocognitive point of view.

Martha Gabriel, author of the book Artificial intelligence: from zero to the metaverse and professor at the Pontifical Catholic University of Rio Grande do Sul, affirms that it will be necessary to adapt to the new times.

“What makes the difference in this context are no longer the answers, but the questions. You have to know how to ask. To know how to ask, you have to know how to think critically, ”she says.

For Yuri Lima, a researcher at the Laboratory of the Future of the Federal University of Rio de Janeiro, “teaching itself needs to adapt to stimulate learning that recognizes students as cyborgs increasingly integrated with new technologies”.

“This requires that teachers also know how to use these same technologies and integrate them into their activities. From the moment projects, activities and tasks become more complex and integrated into the current universe in which students live, motivation is no longer affected by this technological issue”.

Another theme revolves around the future of human creativity and the production of content that is not based on artificial intelligence.

Ten days after the arrival of the OpenAI system, a designer from San Francisco (USA) managed to create in just one weekend a children’s book with text and illustrations made with ChatGPT and MidJourney, a program that produces images providing descriptions .

“An important point still to be addressed is the issue of plagiarism. Since the training of models like ChatGPT is based on texts available on the internet such as news, books and blogs, their answers can bring ideas published by certain people without receiving credit for it”, says Lima.

“In creative areas, this recognition of authors is considered important. In addition, not knowing the sources makes it difficult to recognize biases or even lies, as in fake news.

Machado Días points out that “creativity arises from combinations that are both unusual and relevant. Since algorithms are devices for generating combinations, the creative impulse is expected to be reduced.”

“On the other hand, as algorithms perform their tasks, they tend to generate patterns that we don’t even imagine possible, expanding our combinatorial understanding, that is, our creativity.”

Martha Gabriel goes along a similar line: “Technology can be an incredible tool to expand our thinking, since we can test countless hypotheses, formats, solutions in a short time to refine our hypotheses and improve our questions.”

“However, this can also be a very big threat for those who use these systems blindly, without criticizing or questioning morality and ethics. This is not just an individual risk, but for all of humanity.” (YO)