Intellectual cyborgs and the future of science

Abstract

Unlike classical artificial intelligence (AI), generative AI has the potential to transform scientists into intellectual cyborgs. Leveraging embodied cognition and extended mind theories can help us understand this scientific revolution. Despite ethical concerns, generative AI can enhance research efficiency and accessibility. However, this requires unprecedented proactive regulation and responsible development.

The dawn of intellectual cyborgs

Deep learning, genetics algorithms, reinforced learning, Bayesian networks, Fuzzy logics, and support vector/gradient machines are integral components of everyday scientific pursuits. Recent artificial intelligence (AI) frameworks discussed in this journal include mapping biological intelligence to that of machines [1], and combining deep learning, explainable AI, and transfer learning to improve mental state decoding from neuroimaging [2]. My team has utilized these tools to predict violence, characterize brain diseases with entirely unprocessed neuroimaging, and understand the combination of physiopathological and cognitive pathways in specific neurological conditions. Although these advancements have been successful, such applications may only represent limited approaches compared to other types of AI.

Generative AI could spark a groundbreaking revolution in scientific practices. It focuses on creating new and original content using combined deep learning techniques, such as neural networks and generative adversarial networks, creating content that resembles the input data on which the system has been trained. While classical AI primarily concentrates on solving specific problems using logical rules and expert knowledge, generative AI focuses on creating more flexible, open-ended, original content (but see below and Box 1).

Box 1: Current limitations of LLMs in scientific arenas.

Although the future of LLMs may become integral to scientific exploration and research, current limitations are substantial and include different subjects:

1. Potential lack of genuine novelty: Currently, LLMs are very limited in generating genuinely original ideas or identifying novel questions in a specific field. LLMs generate content based on pattern matching of previous data identified in the training data. The models cannot generate insights not previously suggested in their input data. They cannot infer information that is not explicitly written or understand complex relationships beyond their training data. The content they generate doesn’t come from a unique perspective or personal experience, nor does it involve the kind of creative synthesis that humans engage in when generating ideas.

2. Reliance on high-quality data: The quality of output from an LLM is directly tied to the quality of the input data. Poor or misleading data will result in incorrect or low-quality responses. Similarly, if biases exist in the training data, the LLM will propagate them in its outputs, which can lead to biased scientific insights and conclusions.

3. Lack of comprehension: LLMs do not understand the content they generate. Opposed to researchers, they don’t explicitly use consciousness (or awareness) to modify or generate new insights, theories, or interpretations in a scientific context. LLMs often struggle with understanding and maintaining context over longer content and can lose track. This lack of contextual understanding can lead to the generation of misleading or incorrect information.

4. Inability to handle independent verification and validation: Unlike human researchers, LLMs cannot independently verify the validity of the data or the accuracy of the information they generate. They cannot conduct additional experiments or interrogate data as a human researcher would. However, the recent coupling of LLMs with application programming interfaces (APIs) and other tools may soon partially allow for verification and validation.

5. Lack of intrinsic ethical restrictions: LLMs cannot understand ethical considerations or make value judgments about the appropriateness or potential impacts of the content they generate. This raises concerns about their use in critical areas like scientific research, where responsible conduct and ethical considerations are paramount.

The most talked-about example of generative AI currently is ChatGPT [3], a large language model (LLM) used in various fields. LLMs have countless applications in creative writing, journalism, and scientific research. They have been used in natural language processing [4] involving machine translation, text summarization, and sentiment analysis. In drug discovery, LLMs have been used to generate novel molecules and predict their properties, or new compounds with potential therapeutic properties, and to predict the toxicity and efficacy of existing drugs [5]. In data analysis, these tools are currently used to analyze and summarize large datasets [6], such as social media posts and customer reviews. Multiple scientific fields, such as physics, biology, and astronomy, use this type of AI to generate hypotheses and predictions that can lead to discoveries and breakthroughs.

Although LLMs are already making an impact, they now have the potential to rapidly and exponentially transform scientific practices. In an online survey [7], an emerging number of researchers use LLMs to help with writing, producing presentations, conducting literature reviews, creating graphics, or writing code. Non-native English researchers can also benefit from these tools to improve the coherence of their texts. I have used an LLM in this text to detect typos, find additional examples, and ask for criticisms, all helpful practices for improving science.

As new versions of LLMs emerge, they may go beyond their current use in guided user interfaces and become directly connected with other software for programming, data analysis, illustration generation, and search engines. With these tools, a researcher could ask the software to identify the main gaps in their field, generate novel hypotheses based on patterns and trends in scientific literature, conduct data mining in a dataset, identify relevant relationships between variables, create reports and graphics, and even write the paper.

LLMs could also contact editors, find adequate reviewers to suggest and submit manuscripts through the journal’s online system, requiring minimal input from the researcher. In turn, AI assistance could help editors make informed decisions about manuscript suitability by analyzing content, novelty, relevance, as well as recommend potential reviewers based on expertise, publication history, and research interests. Reviewers would also rely on AI-generated summaries, highlights of strengths and weaknesses, and plagiarism detection to accelerate their evaluation.

Thus, LLMs could automate human-based scientific initiatives or at least hybridize them. This may seem like a futuristic scenario because this technology is not yet free from risk and mistakes (Box 1), and the ability of generative AI to interact on our behalf with other software and people has not yet been fully developed. However, the scenario presented above could become a concrete reality, bringing unprecedented opportunities and concerns to the scientific community.

Generative AI can revolutionize science but requires urgent regulation [3, 8–10]. Their potential negative impact is evident and raises ethical concerns about privacy, accountability, and transparency violations, which require careful consideration [3, 8–10]. For example, the journal Nature reported that researchers use chatbots as research assistants, helping to organize thinking, feedback, writing code, or summarizing research [11]. Later on, the same journal proposed two rules regarding LLMs [9]: These tools will not be accepted as a credited author on a research paper, and researchers using LLM tools should document this use in the methods or acknowledgments sections.

Both challenges and potential negative impacts of LLMs in science are manifold. LLMs can lead to unpredictability regarding what the AI will generate, as it might create a scientifically plausible but misleading or false hypothesis, potentially leading researchers astray. As LLMs generate information based on patterns from their training data, they may unintentionally propagate scientific misinformation, flawed theories, or outdated information without comprehending or verifying its accuracy. The capabilities of LLMs might lead to an overreliance on these tools, potentially undermining critical thinking, creativity, and skepticism - all fundamental to the scientific process. LLMs can unintentionally propagate and even amplify biases present in their training data. This could boost skewed results, biased interpretations, and a lack of diversity in thought and approach. The benefits of LLMs might not be distributed equitably. High costs, technical requirements, or data access issues in the future might prevent researchers from low-income countries from leveraging these tools. LLMs, devoid of the ethical understanding, may inadvertently breach confidentiality, privacy, or intellectual property rights, mainly when dealing with sensitive scientific data. Questions about authorship and accountability arise if an LLM contributes significantly to a research paper. Current norms do not provide clear guidelines about the role of AI in scientific authorship. Finally, who is accountable if an AI generates incorrect or harmful information? The users of the AI, the developers, or the AI itself? These are not just philosophical questions but ones with real-world implications for regulation and legal frameworks. While LLMs hold considerable promise for the future of science, their potential negative impacts warrant careful regulation, monitoring, and proactive ethical considerations to ensure they are used responsibly and to benefit the scientific community.

Regardless of the potential benefits and drawbacks, scientists may become intellectual cyborgs. Despite concerns raised by academics such as Noam Chomsky about the false promises of chatbots [12], the multidimensional improvement of generative AI may exponentially grow. Becoming intellectual cyborgs would bring numerous benefits, including making scientific advancements more efficient, developing revolutionary tools for researchers, especially those in underserved regions, and improving the educational skills of massive learners, among others. According to philosopher Andy Clark [13], the human mind is extended and distributed beyond the body and brain through the environment and technology. Our cognitive processes are embodied, extended, and augmented by ecological settings [14, 15], whether natural or artificial. In terms of embodied cognition, external tools such as pencils, loupes, calculators, smartphones, and the internet have helped us navigate our scientific and everyday worlds more efficiently for centuries. Generative AI will accelerate this process.

A critical unanswered question underlying the current debate is whether this new technology will make us passive observers of leading AI or whether we will direct and embody such technology, allowing us to extend our scientific minds to spatiotemporal dimensions never before reached. As soft intellectual cyborgs, providing answers to this complex and fascinating question now is critical. By anticipating and regulating the future of generative AI in science, we can use and focus on our ancient human creativity, ethics, and responsibility skills to guide its development.

On a personal note, I find myself oscillating between being fascinated by how these tools will expand our scientific mind, and feeling deeply worried due to multiple concerns. However, I do not doubt that the future of generative AI will revolutionize the entire landscape of scientific practices.