{"id":568148,"date":"2025-11-18T15:49:27","date_gmt":"2025-11-18T18:49:27","guid":{"rendered":"https:\/\/revistapesquisa.fapesp.br\/?p=568148"},"modified":"2025-11-18T16:28:04","modified_gmt":"2025-11-18T19:28:04","slug":"researchers-use-artificial-intelligence-in-scientific-work","status":"publish","type":"post","link":"https:\/\/revistapesquisa.fapesp.br\/en\/researchers-use-artificial-intelligence-in-scientific-work\/","title":{"rendered":"Researchers use artificial intelligence in scientific work"},"content":{"rendered":"

Generative artificial intelligence (AI) tools are just beginning to find their way into research\u2014for now, their use is largely confined to writing-related tasks. But a new international survey by academic publisher Wiley<\/em> suggests that\u2019s about to change. Of nearly 5,000 researchers surveyed across more than 70 countries (including 143 from Brazil), the majority expect widespread adoption of generative AI in academia within just two years (see infographic below<\/em>). \u201cThere\u2019s a strong consensus that artificial intelligence is poised to reshape the entire research ecosystem,\u201d said Josh Jarrett, vice president for AI at Wiley, in an interview with Nature<\/em>.<\/p>\n

The survey also asked researchers whether AI is already outperforming humans at certain day-to-day research tasks. Over half of respondents said yes\u2014citing AI\u2019s edge in tasks like identifying potential collaborators, summarizing papers into educational content, detecting plagiarism, filling out reference lists, and monitoring field-specific publication activity.<\/p>\n<\/div>

\n \n \n \n <\/picture>Alexandre Affonso\u2009\/\u2009Pesquisa FAPESP<\/span><\/div>
\n

Still, most respondents agreed that humans are irreplaceable when it comes to higher-level tasks: spotting emerging research trends, choosing where to publish, selecting reviewers, managing administrative workflows, and seeking out grant opportunities. Despite rising enthusiasm, 81% of researchers voiced concerns about bias, data privacy, and the opaque training methods behind many AI systems. Nearly two-thirds also pointed to a lack of guidance and hands-on training as major obstacles to putting AI to fuller use in their work.<\/p>\n

\u201cIn Brazil, many researchers still feel uncertain about the ethical paths forward when it comes to using generative AI,\u201d says political scientist Rafael Sampaio of the Federal University of Paran\u00e1 (UFPR). \u201cThat\u2019s why it\u2019s crucial for institutions like the Brazilian Federal Agency for Support and Evaluation of Graduate Education [CAPES] and the Brazilian National Council for Scientific and Technological Development [CNPq] to step in with formal guidance.\u201d Sampaio coauthored a practical guide to the ethical and responsible use of AI in research, released in December in collaboration with business expert Ricardo Limongi, from the Federal University of Goi\u00e1s (UFG), and digital education scholar Marcelo Sabbatini, from the Federal University of Pernambuco (UFPE).<\/p>\n

Sampaio uses AI tools in his own research. One go-to tool is Google\u2019s NotebookLM<\/a>, a platform that lets users ask research questions and get synthesized summaries based on curated materials. The app can \u201ccrawl\u201d as many as 50 documents at once, including PDFs, audio, and video files\u2014essentially serving as a multisource research assistant. For now, it\u2019s free to use.<\/p>\n

\u201cI use it mostly for large-scale scans\u2014for an initial triage or to help decide what\u2019s worth a deeper read,\u201d he explains. \u201cIt\u2019s also great when I\u2019m trying to track down a paper I\u2019ve read before but can\u2019t remember the title or author.\u201d One neat feature is NotebookLM\u2019s ability to generate podcast-style summaries with two virtual voices chatting about the documents\u2014entirely generated by AI.<\/p>\n

The guide Sampaio helped write catalogs several tools\u2014including NotebookLM\u2014that could be useful to researchers, but cautions that human oversight remains essential. AI should serve as an assistant, not an author. Limongi, one of the handbook\u2019s coauthors, regularly uses LitMaps<\/a> and Scite<\/a>\u2014two platforms designed to assist with literature reviews in selected fields. Users can upload a paper and get an interactive map of how that study connects to others, complete with clickable citations and reference links to build stronger arguments.<\/p>\n

But Limongi notes that researchers need training to use these tools wisely. Uncritical use of AI, he cautions, can lead to deskilling\u2014especially in areas like critical reading and the ability to articulate a coherent argument. \u201cAI can assist,\u201d Limongi says, \u201cbut it\u2019s the researcher who designs and drives the study. We can\u2019t reduce researchers\u2019 role to just pushing buttons.\u201d Science journals prohibit AI tools from being credited as authors\u2014human researchers are always ultimately responsible for the scientific content\u2014and for any text written with help from AI.<\/p>\n

Claude.ai, a conversational AI chatbot and rival to ChatGPT, is gaining attention for its usefulness in drafting outlines and structuring scientific writing. At the Federal University of Pernambuco (UFPE), Marcelo Sabbatini uses Claude regularly to sketch out initial drafts and refine both academic papers and public-facing science content. \u201cI\u2019ll ask for ways to frame a topic and then how best to develop the argument,\u201d he says. \u201cIt gives me a structure to build on\u2014fact-checking, digging deeper, and adding my own knowledge and perspective.\u201d<\/p>\n

\"\"Created by the San Francisco\u2013based startup Anthropic, Claude topped the rankings in a May 2024 study published in Royal Society Open Science<\/em>, which tested how accurately ten generative AI models\u2014including ChatGPT and DeepSeek\u2014could summarize scientific papers. But the overall picture wasn\u2019t rosy. The study found that across 5,000 papers, the chatbots produced flawed or overstated conclusions in up to 73% of the summaries.<\/p>\n

Materials scientist Edgar Dutra Zanotto, of the Federal University of S\u00e3o Carlos (UFSCar), takes a multitool approach. He uses a suite of chatbots\u2014ChatGPT, DeepSeek, Claude, Perplexity, and Gemini\u2014for a wide range of research-related tasks. These tools help him polish English-language texts, run calculations (like figuring out the molar composition of glass based on molecular weights), and even suggest peer reviewers for the Journal of Non-Crystalline Solids<\/em>, where he\u2019s an editor. \u201cI look up the most influential researchers in the field I\u2019m working on,\u201d he explains.<\/p>\n

Zanotto had prior experience with AI before the chatbot boom. At the Center for Research, Education, and Innovation in Vitreous Materials (CeRTEV)\u2014a FAPESP-funded Research, Innovation, and Dissemination Center (RIDC)\u2014he and his team trained a machine-learning algorithm using a massive dataset of 55,000 glass compositions. It\u2019s one of the largest databases ever assembled in the field, capable of predicting novel glass structures that don\u2019t yet exist. Before the advent of AI tools, designing a new glass formulation was a long, trial-and-error process. \u201cNow, we can create a new glass in just a month,\u201d Zanotto says. \u201cIt used to take years\u2014and every new formulation came with a dissertation.\u201d<\/p>\n

AI is also proving useful for solving complex mathematical problems through tools like MathGPT<\/a>. And increasingly, AI models are being used as coding companions. \u201cClaude and Gemini are both popular right now for writing and debugging code,\u201d Limongi notes.<\/p>\n

Because these models are trained on massive datasets scraped from the internet, they often replicate the same biases found in those sources. And sometimes, they simply hallucinate\u2014making up terms, names, and even references that don\u2019t exist. To minimize those risks, Zanotto routinely tests the same prompt across multiple AI platforms and cross-checks the results to find the most reliable answer.<\/p>\n

One clear strength of tools like ChatGPT is brainstorming\u2014their ability to scan and combine information from vast data sets makes them powerful partners for early-stage ideation. Researchers can \u201cchat\u201d with the tool to get suggestions on how to approach a research topic or which paths to explore. But since the models are trained on preexisting content, their suggestions might not be entirely original\u2014though they are also able to suggest novel variations of existing ideas.<\/p>\n

Industrial engineer Roberto Antonio Martins, of the Federal University of S\u00e3o Carlos (UFSCar), has been testing out advanced AI tools like ChatGPT, Copilot, and DeepSeek, especially a recently added feature called \u201cdeep research.\u201d This feature digs through both the web and academic databases, returning a synthesis of the topic along with a list of cited sources, Martins explains.<\/p>\n

The process starts with the tool asking the user scoping questions\u2014helping it generate a more precise and well-structured prompt before launching the search. Martins always starts by framing his research topic as a question. \u201cThis isn\u2019t your typical keyword search,\u201d he says. \u201cWhen you ask a full question, you provide more context\u2014and that usually means more relevant output.\u201d Martins now lectures on the academic use of generative AI.<\/p>\n

Researchers need training to understand the limits of AI<\/p><\/blockquote>\n

Perplexity, another chatbot gaining traction among researchers, is praised for its ability to fetch fresh, web-based content and provide clickable footnotes linking directly to sources. \u201cI use these tools early on\u2014when I\u2019m gathering background information and forming initial hypotheses,\u201d says Brazilian-born Cyntia Calixto, a professor of international business at Leeds University Business School in the UK. She\u2019s been using AI tools in her research on CEO activism\u2014how company leaders take public stances on controversial topics online. \u201cIt\u2019s like having a research assistant to bounce ideas off as I dig deeper into the topic,\u201d she explains.<\/p>\n

When she needs to search for scientific papers, Calixto turns to SciSpace\u2014a tool that not only finds research articles but also provides concise summaries of each one. Aydamari Faria-Jr., a biomedical scientist at Fluminense Federal University (UFF), recommends two other platforms: Answer This<\/a> and Elicit<\/a>. Both allow researchers to search across multiple scholarly databases like PubMed and Scopus. \u201cThey include filters that streamline the search process,\u201d he explains, \u201cand allow you to focus on specific study types\u2014like randomized clinical trials, systematic reviews with or without meta-analysis, or cross-sectional studies.\u201d Faria-Jr. still pairs AI-powered searches with manual queries through academic databases, ensuring a comprehensive review of the topic he\u2019s currently researching.<\/p>\n

AI support in academic writing is also finding a role in upholding scientific integrity\u2014particularly in assessing whether citations in a manuscript are sound and accurate. One tool, Scite, allows users to see the original context of a cited passage. It uses a deep learning model to assess whether the reference genuinely supports the citing statement\u2014or if it actually contradicts it. Attorney and law professor Cristina Godoy, from the Ribeir\u00e3o Preto School of Law (FDRP-USP), is a regular Scite user.<\/p>\n

She recalls finding a Nature<\/em> paper through Scite that added depth to one of her own research articles on AI. \u201cScite suggests articles that engage with your topic\u2014and it links directly to the papers,\u201d she explains. \u201cYou can even limit the search to specific journals.\u201d Godoy also uses other AI tools for analyzing legal data. She is careful when it comes to generative AI. She never inputs excerpts from unpublished manuscripts\u2014aware that these models can retain training data, and there\u2019s no guarantee those ideas won\u2019t surface in someone else\u2019s output later on.<\/p>\n

The story above was published with the title “Your new lab assistant?<\/strong>” in issue 352 of April\/2025.<\/p>\n","protected":false},"excerpt":{"rendered":"Researchers use artificial intelligence in scientific work","protected":false},"author":684,"featured_media":553394,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_exactmetrics_skip_tracking":false,"_exactmetrics_sitenote_active":false,"_exactmetrics_sitenote_note":"","_exactmetrics_sitenote_category":0,"footnotes":""},"categories":[166],"tags":[220,219,2413],"coauthors":[2721],"class_list":["post-568148","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-policies-st-en","tag-communication","tag-computation","tag-technology","position_at_home-sumario"],"acf":[],"_links":{"self":[{"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/posts\/568148","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/users\/684"}],"replies":[{"embeddable":true,"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/comments?post=568148"}],"version-history":[{"count":2,"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/posts\/568148\/revisions"}],"predecessor-version":[{"id":569728,"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/posts\/568148\/revisions\/569728"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/media\/553394"}],"wp:attachment":[{"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/media?parent=568148"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/categories?post=568148"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/tags?post=568148"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/revistapesquisa.fapesp.br\/en\/wp-json\/wp\/v2\/coauthors?post=568148"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}