AI in Research

Original Editor - Michael Rowe

Top Contributors - Wanda van Niekerk, Jess Bell, Tarina van der Stockt, Kim Jackson and Tolulope Adeniji  

Introduction

Artificial intelligence (AI) is transforming research, offering new possibilities and challenges throughout the research process. It has the potential to enhance researchers' capabilities and efficiency, but its use raises several important ethical questions, as well as concerns about cognitive skills development and the role of human expertise in the academic environment.

Ethical Implications and Responsible Use of AI in Research

As AI becomes integral to research, its adoption demands a focus on ethical considerations and responsible usage. Ensuring transparency in AI-driven processes, addressing biases in datasets, and maintaining accountability for decisions are essential. Researchers must balance leveraging AI’s capabilities with preserving human expertise and critical thinking. Clear guidelines and collaboration across disciplines can help navigate these challenges, fostering a research environment where AI serves as a tool to complement, not replace, human ingenuity and judgment.

Trends in Research

The academic and research landscape is changing. The amount of published research is increasing, but finding research topics or ideas can be more difficult. The rapid increase in published research papers can make it harder for people to follow and keep up with research, and researchers are at risk of being overwhelmed by the volume of research. This could hinder further research, causing existing ideas to remain dominant while novel ideas go unrecognised.

Impact of AI on Research

AI can influence academic research in a number of ways, including:

• writing, peer review and publishing
  • AI has the potential to increase the speed of everything
• how we do research
  • AI can help with text analysis, hypothesis generation and the analysis of massive datasets
  • it can also create simulations of human interaction
• what research means
  • AI can help bridge the gap between academia and society, help academics / researchers explain their work to each other, create funding opportunities for multi or interdisciplinary collaboration, connect researchers with ongoing projects and enable discussions in their field of interest
• what we research
  • integration of AI into professional contexts

Generative AI is an Expert

Generative AI has:

• expertise within, and across, professional domains
• extensive knowledge and the ability to apply it in creative ways
• the ability to understand and navigate complexity through expert communication skills

It is constantly improving at producing accurate responses to complex questions and the trend is moving from information abundance to expertise abundance.

The State of the Art of AI in Research

AI is being used in research in the following areas:

• AI-assisted literature review and synthesis
  • advanced language models are being used to rapidly analyse and summarise large volumes of scientific literature
• generative models for hypothesis generation
  • AI systems are being developed that can propose novel hypotheses and research questions
• AI-augmented experimental design
  • generative AI is being applied to optimise experimental protocols and design more efficient studies
• synthetic data generation
  • in fields where data collection is challenging or expensive, generative models are being used to create realistic synthetic datasets

AI-Supported Research Process

Problem Identification

Generative AI is great at generating ideas. Ask it for ideas, not answers:

• example prompt: "Give me a list of relevant research problems to explore in the field of ...... (add the specific field you are interested in). I am interested in (list areas and be specific)."

AI platforms that can help with problem identification include:

• Elicit
• Perplexity

The video below shows how you can use Elicit to find a topic for your research project:

Idea Generation

Novel Research Ideas

Si et al. investigated if large language models (LLMs) can generate novel research ideas. In their study, natural language processing (NLP) researchers were recruited to create novel ideas and blind review human and LLM ideas. The authors reported:

"By recruiting over 100 NLP researchers to write novel ideas and blind reviews of both LLM and human ideas, we obtain the first statistically significant conclusion on current LLM capabilities for research ideation: we find LLM-generated ideas are judged as more novel (p < 0.05) than human expert ideas while being judged slightly weaker on feasibility.”

In their analysis of the human study, the authors conclude that:

• "human experts may not be giving their best ideas"
• "reviewers of the ideas tend to focus more on novelty and excitement"
• "reviewing ideas is inherently subjective"

Some limitations of LLMs were:

• lacking diversity in idea generation
• unable to reliably evaluate ideas

With the qualitative analysis of the human and AI-generated ideas, the following common failures of AI ideas were:

• vague on implementation details
• datasets misused
• missing or inappropriate baselines
• unrealistic assumptions
• resource-demanding
• not well-motivated
• existing best practices not adequately followed

In contrast, some of the human-generated ideas had unique strengths and weaknesses:

• "Human ideas are generally more grounded in existing research and practical considerations, but may be less innovative"
• "Human ideas tend to be more focused on common problems or datasets in the field"
• "Human ideas sometimes prioritise feasibility and effectiveness rather than novelty and excitement"

If you'd like, you can read the full paper here.

Idea Generator

Claude is an LLM that is not connected to the internet. While it provides different kinds of answers than more research-focused AI platforms, it can still highlight valuable ideas for users to consider.

Research Design

AI can be used to investigate which research design is best suited to answer a specific research question. This is especially useful for researchers who are not familiar with different research designs:

• example prompt: "I need to demonstrate .....(add your specific context) of ......; suggest a few research designs I can consider"
• remember, the more detailed context you provide, the more relatable, localised and specific the suggestions provided by AI will be

Useful AI platforms for research design include:

• Claude
• Perplexity

Literature Review

Some AI research platforms allow you to upload and then ask questions about research papers:

• example prompt: after uploading a paper: "Use the attached papers / articles to give an overview of .... (add in the specific context)
• be specific and detailed

AI platforms that can help to inform the literature review:

• Elicit
• Research Rabbit
• Consensus

Please watch the videos below if you'd like to see how these platforms can be used.

Reading

AI-supported platforms can help summarise research papers and with data extraction. There is, however, a concern that this could hinder the cognitive development that takes place when working through difficult ideas and concepts.

• George et al. highlight that an increasing reliance on technology can negatively impact cognitive skills, such as critical thinking, problem-solving and creativity. Research shows that as people offload more mental tasks to devices and software, less cognitive effort is used and this may lead to weakened neural connections over time. A measurable decline in reasoning, argument evaluation and creative thinking is already seen across the general population. As AI capabilities advance, humans need to make a conscious effort to develop and maintain essential cognitive skills alongside technological advancements.

AI-supported platforms to help you read a research paper include:

• Explainpaper
• Scholarcy
• Scispace
• Sharly
• Semantic Reader (a product of Semantic Scholar)

The video below illustrates how Explainpaper can be used.

Data Collection

AI platforms, such as Claude, can be used to generate ideas on different data collection methods specific to your research:

• example prompt: "How could I go about collecting data for a .... (study design) that aims to answer the following research question? ...Insert question... Give me a few options"

Data Analysis

Qualitative data analysis:

• qualitative data analysis is more subjective and open to interpretation and there is considerable value in the researcher analysing their own data
• when conducting qualitative research, it might be useful to engage with AI on your analysis and see if it comes up with different themes
• check with your institution if this is allowed, and consider using it as a "critical friend"
• MyRA is a new platform that aims to make qualitative data analysis more efficient

Quantitative data analysis:

• frontier language models are capable of writing the code needed for statistical analysis
• results can be presented in a range of different visual formats
• insight is needed to evaluate the quality and correctness of the output

Writer

AI models, like Claude, can also be used to review and provide you with ideas on your writing or the structure of your writing:

• it is important to give AI your written content and ask it for suggestions on the outline or structure, considering the context that you have provided it with
• don't copy and paste something generated by a language model and present it as your work! You remain responsible for the information provided!

Examples of AI-supported writing tools include:

• Lex
• Quillbot
• ProWritingAid
• Jenni AI
• PaperPal

Can AI be a Co-Author?

Recently, some academic papers have been published listing AI as co-authors. This is currently a controversial topic as there are concerns about "authorship, originality, factual inaccuracies and "hallucinations" or confabulations." The New England Journal of Medicine AI has elected the following policy:

"At NEJM AI, we have elected instead to allow the use of LLMs for submissions, as long as authors take complete responsibility for the content and properly acknowledge the use of LLMs. However, this policy does not allow an LLM to be listed as a coauthor."

Other scientific journals and publishers do not accept AI as a co-author (Springer-Nature and Science), some (Taylor and Francis) have indicated that they will review the situation and other publishers (Elsevier) have already included AI as co-authors.

The BMJ group follows an approach in line with the World Association of Medical Editors and the Committee on Publication Ethics and requires authors to be transparent about the use of AI technologies in submitted content, but does not accept AI as an author of any content submitted for publication. "BMJ only recognises humans as being capable of authorship since they must be accountable for the work." Authors and contributors remain responsible for any content produced by AI technology in their work. This includes "responsibility for accuracy, suitable attribution of sources and absence of plagiarism."

Take home message on AI as co-author:

• check with your journal editor, study supervisor, and institutional research director about their policies on the use of AI, but remember: 'Friends don't let friends co-author with ChatGPT."

Peer Reviewer

When using AI as a peer reviewer of your work, consider the following:

• AI platforms, such as Claude and CoPilot, may be useful in providing feedback on your work
• before uploading any of your work, make sure that you understand the platform's terms of service and be careful with what you upload
• provide enough context to the AI platform for it to provide you with constructive and relevant feedback on your work

Liang et al. investigated the effectiveness of using LLMs to provide scientific feedback. They reported the following results:

• substantial overlap between human feedback and LLMs
• positive user perceptions of the usefulness of LLMs feedback
• LLMs sometimes bring up novel points not mentioned by human reviewers

The authors highlight the possible benefits of using AI as a peer reviewer:

• can be a valuable resource for authors looking for constructive feedback
• may be helpful for researchers with limited access to good-quality feedback mechanisms
• useful self-check tool for authors to improve work efficiently

The Potential of AI in the Peer Review Process

Saad et al. explored the potential of ChatGPT as a journal peer reviewer and reported that:

• low levels of agreement were found with human reviewers
• using ChatGPT in its current state as an automated peer reviewer is not recommended and should not be used to substitute reviewers
• it may be used as a supplementary tool to complement and support human expertise and judgement

Dissemination of Academic Information Using AI

AI can be used to disseminate academic information in the following ways:

• visual summaries
  • AI can create flow charts or visual representations of papers and make key concepts more accessible
• audio conversations
  • AI can transform research papers into audio dialogues, using a conversational format to explain the content of the paper
  • example: NotebookLM
• podcast generation
  • possible conversion of reading lists into AI-generated podcasts for on-the-go listening and learning
  • example: ElevenLabs
• public engagement
  • transforming and translating jargon-heavy academic writing to more accessible and understandable formats for the general public

Please note that the idea is not to replace students' and academics' deep engagement with original research papers, but to find complementary ways to interact with the research and disseminate research findings.

The Human Element in the World of AI in Health Research

Human beings can add value in this changing research approach in the following ways:

• problem selection
  • what problems are worth exploring?
• significance
  • what results are significant?
• collaboration
  • we can work in teams to achieve more than any individual can achieve in isolation
• strategic goals
  • what goals matter to us?
• interpretation
  • we put findings into a relevant context
• accountability
  • we are ultimately responsible for the outcomes
• meaning
  • we give meaning to the research enterprise