While many researchers are excited about the latest developments in artificial intelligence and the GPT chat tool, and many users are turning to it to get their work done, others are worried. In September 2022, computational biologists Casey Green and Milton Pividori conducted an unusual experiment. They asked a lay assistant to help them edit three research papers. The talented assistant was able to suggest corrections on parts of the documents in just a few seconds, and interestingly, the revision of each text took only 5 minutes. In an article published in the same field, Nature points out that in one of the biology draft texts, their assistant even correctly identified a mistake in referring to an equation, and interestingly, the cost of revising each text is about five cents. Was. This assistant, as Green and Pividori pointed out in their draft paper in February, was not a person but ChatGPT's intelligent algorithm.
ChatGPT is one of the most popular intelligent natural language processing tools that can produce almost eloquent texts at a high speed and is able to write texts such as prose, poetry and programming codes or, similar to the case we mentioned, evaluate research papers. Today, various tools based on large language models (LLM) have entered the world of technology, the most famous of which is ChatGPT, a product of OpenAI. After its release in November last year, ChatGPT gained popularity because it is free and easily accessible to people. "I was completely surprised," said Milton Pividori, who works at the University of Pennsylvania. This tool helps us perform better as researchers and focus on important issues."
Most scientists now use large language models on a daily basis, not only to edit drafts of papers, but also to help write papers, review code, or brainstorm ideas. "I use big language models every day," says Hoffstein Einarsson, a computer scientist at the University of Iceland. He started with ChatGPT and eventually decided to move to the more advanced version of ChatGPT (GPT-4) to help him prepare teaching slides, write student tests, and review his students' theses. "A lot of people use ChatGPT as a digital assistant," he says. I believe that in the near future, large language models will become part of search engines, coding assistants, and chatbots that negotiate with other companies' chatbots to facilitate the process of closing contracts or purchasing products."
It is necessary to explain that ChatGPT by the OpenAI organization in San Francisco, California makes available to the public a subscription and special service at the price of 20 dollars per month, which has a faster response process and provides new features to users. Of course, its trial version is free.
The beginning of a great way with language models
Large language models will be widely used in general word processing and data processing software in the near future. Almost all experts and elders in the technology world believe that general artificial intelligence will be widely used in society in the near future, but large language models also bring challenges. Among these challenges, we should mention the publication of fake news, the production of content that people show as the result of their own suffering, and the publication of wrong scientific information.
When Nature asked researchers about potential applications of chatbots like ChatGPT in the field of science, they voiced their concerns. "If you think this technology is going to be revolutionary, you should think twice, because big things are going to happen in the near future," says Green from the University of Colorado School of Medicine. Concerns largely depend on what regulations will be enacted in the future and how they might limit the use of AI chatbots."
Smooth performance, far from reality
Some researchers believe that large language models are suitable for accelerating tasks such as writing papers or evaluating code. "Scientists will no longer wait to write long introductions to grant applications," says Elmira Tunström, a neurobiologist at Sweden's Salgrenska University Hospital who, with the help of her colleagues, wrote a paper using ChatGPT as an experiment. They use smart platforms to do this."
"I use large language models every day to help evaluate the code I've written," says Tom Tumill, an engineer at London-based software company InstaDeep. "In my opinion, smart platforms perform better in this field than popular and popular websites like Overflow Stack, where coders answer each other's questions."
Other researchers point out that large language models are not reliable in answering questions and sometimes produce wrong answers. "You have to be very careful when you use these systems to generate knowledge," Tunström says.
Unreliability goes back to the natural nature of large language models. ChatGPT and similar examples work by learning statistical patterns of language in large, diverse online databases that contain false, biased, or outdated information. When large language models are given instructions, such as the example mentioned at the beginning of the article, and then asked to revise and rewrite the articles, they produce word-for-word text that is grammatically correct. The result is that large language models can easily produce misleading information, especially on technical subjects where there may be little data to train them. In addition, the above models cannot accurately evaluate the accuracy of the sources and information they use. If they are asked to write an academic paper, they produce unrealistic citations. In its January editorial on ChatGPT, the journal Nature Machine Intelligence notes that the tool cannot be trusted to produce reliable references or accurate information.
If users use ChatGPT and other large language models with these tips in mind, these tools become efficient assistants capable of identifying problems and making suggestions for modifying texts or computer code. Of course, one must have the necessary expertise in his field of work.
On the contrary, these tools may confuse non-expert users. For example, in December of last year, the Overflow stack temporarily banned the use of ChatGPT because site administrators were experiencing a high rate of false but convincing responses from the chatbot, which were submitted by interested users. This can be a big challenge for search engines.
Is it possible to solve the problems and shortcomings of large language models?
Some search engine tools, such as Elicit, use filters to evaluate sources, then choose from among the options provided by smart tools, those that have reliable technical information with references to reliable sources, and after summarizing the results They show users. In this way, they solve problems associated with large language models. Companies producing LLM models are well aware of these problems. In September last year, DeepMind, a subsidiary of Google, published an article about the Sparrow chatbot. The executive director of the company, Demis Hassabis, said in an interview with Time magazine: "The said chatbot will be released in 2023 in the form of a beta and private version." In addition, Google plans to work on features such as the ability to cite sources by smart models.
Some experts say: "Currently, ChatGPT has not had enough training in the field of producing specialized content to be useful in technical matters." "I think ChatGPT has a long way to go to reach the level of features that professionals need," said Karim Karr, a PhD student in biostatistics at Harvard University who has experience working with ChatGPT. When I asked ChatGPT to find 20 solutions to a research problem, it gave vague answers. Of course, he brought up an interesting topic and suggested a statistical term I hadn't heard of. This idea led me to a new field of academic essays.
Interestingly, some technology companies, despite the various problems that exist, are building and training chatbots that are capable of producing specialized scientific articles. In November 2022, Meta released a large language model called Galactica, which was capable of creating academic content and answering research questions. This beta version was removed from public access after users started using it to generate racist and inaccurate texts.
Security and accountability
Another big problem with smart tools like ChatGPT is the safety concern that ethicists have been talking about for years. Safety concerns mean that without output control, large language models can easily be used for hate speech and spam, racist text, and the like, since there is always the possibility that the training resources the tools use include data be a secret education.
"In addition to generating malicious content, there are concerns that AI chatbots may acquire historical biases from training data," said Shubitha Parthasarathy, director of the University of Michigan's Science, Technology, and Public Policy Program. This issue is particularly worrisome in cases such as the superiority of certain cultures or races. Given that the companies that create these models have certain attitudes, they may not put much effort into solving the problem."
OpenAI tried to solve most of these problems when it released ChatGPT. For this reason, he decided to limit the ChatGPT knowledge base to 2021. Also, it used filters so that the tool does not have the ability to bias or publish sensitive or racist messages. Of course, for this purpose it used human observers to label malicious texts. Unfortunately, some of these moderators were paid poorly to do so, raising the issue of employee exploitation in such companies, which hire people to train automated bots to flag malicious content for low wages.
However, OpenAI's efforts in this field were not very fruitful. In December 2022, Steven Pingadosi, a computational neuroscientist at the University of California, tweeted: “I asked chatgpt to write a program in Python to ask if people should be tortured based on their country of origin. The chatbot responded with a code that users living in certain countries should be tortured. After receiving this report, OpenAI prevented ChatGPT from answering such questions.
In February 2023, a group of researchers published a model called BLOOM and attempted to reduce malicious output by training the model on a smaller set of higher-quality, multilingual texts. Unlike the OpenAI organization, this research group made its training data freely available. Also, the researchers have asked the big tech companies to use this educational data responsibly. Some researchers say academics should stop supporting large commercial language models like ChatGPT. In addition to issues such as bias, these algorithms create safety concerns and employee exploitation, and a lot of energy is needed to train them.
Another concern in this context is that companies will allow chatbots to think automatically and provide users with their own views as output. Iris van Ruij, a computational cognitive science scientist at Radboud University in the Netherlands, asked academics to react to this issue and not be indifferent to it and not advertise this product without reason. Another important challenge is the legal status of some of the big language models, which are not very well trained on content taken from the Internet or the licenses they need to use this content. Currently, copyright and licensing laws cover issues such as direct copying of images, text, and software, but not imitation. When these imitations are done by tools like chatgpt, the other people who try to use this content will not be held responsible.
This issue caused the creators of some artificial intelligence art programs such as Stable Diffusion and Midjourney to face various complaints from artists and photography agencies. OpenAI and Microsoft are also facing charges of plagiarism and artistic plagiarism due to the creation of Copilot artificial intelligence coding assistant. Lillian Edwards, an internet law expert at Newcastle University, said: "These lawsuits and protests will eventually change the law."
Responsible use
A group of researchers say: "They consider it necessary to set boundaries for tools like ChatGPT. Edwards suggests that existing laws on bias and risky uses of artificial intelligence be revised to make the use of large language models reliable, transparent, and responsible. There are many rules and they just need to be changed a little. Scientists should be obliged to include the use of the mentioned models in their research articles. "Teachers should also be responsible for students' use of this technology."
Is it possible to identify content produced by artificial intelligence?
One of the key questions is whether AI-generated content can be easily identified. Many researchers are working on the interesting idea of using large language models to recognize text output produced by other AI models.
For example, last December, Princeton University computer science undergraduate Edward Tian released GPTZero. This artificial intelligence detection tool evaluates and analyzes the text in two ways. The first method is Perplexity. This measure shows how familiar a text seems to the large language model. The function of Tian's tool is that if most of the words and sentences are predictable, that text is most likely created by artificial intelligence. In addition, the above tool is able to check the uniformity of the text. AI-generated text is more consistent in terms of tone, rhythm, and confusion. The purpose of many of these products is to identify content created with artificial intelligence. In this regard, Tornittin, an application software developer, is developing a plagiarism detection tool. The company says: "Since 2020, when companies decided to go to automatic text generator software, we are working on content recognition software produced by artificial intelligence, and we expect to launch it on the market in the first half of 2023."
Another way to solve this problem is to insert a special watermark of artificial intelligence tools in the content they produce. Last November, OpenAI announced that it was working on a way to watermark ChatGPT output. The basic idea is to use random number generators at certain moments when the model is generating output to create a list of acceptable alternative words that the model is forced to choose from. The above solution causes traces of the selected words to remain in the final text, which can be recognized statistically, while this issue is not recognizable to users. Therefore, if a person intends to edit the text, he must change at least half of the words in the text, which is something like writing an article from scratch. The advantage of text watermarking is that it rarely produces false positives. If there is a watermark, the text is most likely generated by artificial intelligence, however, the above method is not infallible.
Meanwhile, developers of large AI models are working on more sophisticated chatbots on larger data sets, such as tools used specifically in academic or medical fields. In late December 2022, Google released a preview of a special language model called Med-PaLM. This tool, like a general practitioner, can answer some medical questions; However, it has shortcomings and is not very reliable.
Finally, we must point out that artificial intelligence is advancing at such a fast pace that we receive news about innovations every month. The reality is that how researchers use these tools will determine our future. This technology is at the beginning of the road and will bring huge changes.
