Artificial Intelligence (AI) and Language Models: AI Research Tools

As many know by now, one of the biggest issues with using AI language models for academic research is that they make stuff up and present falsehoods as facts. This makes them unsuitable for tasks where accuracy is extremely important, but at the same time, presents opportunities for important discussions about misinformation and bias.

What can tools like ChatGPT, Gemini, and Copilot do? They can help you get ideas for how to begin your research and search for scholarship. It can suggest keywords, search strings, and format citations. You can prompt ChatGPT to be more specific in one or more areas, but note that the same will happen - a mix of good ideas, wrong ideas, and overly broad suggestions. 

There are better alternatives to ChatGPT/Gemini/Copilot for academic research, instead try Perplexity AI or Elicit AI. Do not trust or accept citations that are suggested by any AI program- track down and verify every source that is mentioned.

The number of new AI LLM tools has exploded in the past couple of years. I have tested and experimented with the more well-known tools listed above, but here is a sampling of a few more and very brief info about what they do. You can find them all with a quick search in your browser. Some are available free to users who create an account/login, others require a paid or institutional subscription.

• CORE-GPT (scans millions of Open Access scientific publications)
• DataSeer (scans scientific texts for data descriptions)
• GDELT Project (monitors world news and broadcasts)
• Hum (analyzes how publishing audiences interact with content)
• Iris.ai (specific to chemistry, pharma, medtech searching & filtering)
• LASER AI (tool for systematic reviews)
• Prophy (ranks articles & researchers for research, review, recruitment)
• Scholarcy (article summarizing tool)
• Scite.ai (analyzes citations in scientific papers) 
• SOMA (automates data prep, model training, and evaluation)
• Writefull (writing and proofreading)

Inaccurate Information: Some tools, such as ChatGPT occasionally generate entirely fake information that they will confidently assert is true. Programmers call these "hallucinations." They can be citations or facts that do not actually exist. If you are going to rely on a piece of information found in AI Tools, it's generally a good idea to confirm its existence from another source.

Data Bias: Like any tool dependent on data, most face the same biases as the data they received, much of which is Western in perspective, English in language, and reflect the norms and status quo of the content providers’ population. 

Lack of New Information: Although this is improving with newer and more advanced GPTs, some AI tools have limited abilities to comprehend recent events or new information they were not trained on during their development. Asking questions about news or recent events is likely to be unfruitful. Outdated information may also be provided. 

Data Privacy: Many AI Tools use the information provided them by users to further their own development. Accordingly, it can be risky from a data privacy perspective to provide them with sensitive or private data. This is especially true when you have other people's private data that you are entrusted with safekeeping. 

Ethical Issues: Many AI Tools limit the capabilities of their application to prevent it from being used in ways the company believes is immoral, unethical, or sensitive. ChatGPT for instance will refuse to help if it thinks a user request promotes hateful stereotypes, criminal activity, academic misconduct, suicide advice, sexual content, medical prescription, or stock shorting. While many of these are admirable goals, many ordinary users with good intentions may find they do not agree with the values interpreted by the AI Tool, or find that guardrails are preventing them from doing uncontroversial work as well. 

-Content adapted from Univ. of Fraser Valley, AI Tools

Citing AI is becoming more and more important as we incorporate these tools into our research and writing. What’s tricky is that chats in AI may be “non-retrievable data,” which means other people can’t access the chat to check your source and if they replicate your prompt or question, they may get a different answer. Since AI tools generate content that does not quite fit into traditional sources, we need to adapt the company, tool, and prompt information to fit.

• If you use an AI tool that does cite its sources, you also need to acknowledge secondary sources in your work.
• Some styles require a version number of the AI tool. For example, the format for the version number in ChatGPT references includes the date because that is how OpenAI is labeling the versions. Different large language models or software might use different version numbering; use the version number in the format the author or publisher provides, which may be a numbering system.
• If you choose to use ChatGPT or some other AI technology for writing, be sure you are transparent about your use of it with your instructors. Each citation style has different recommendations for citing generative AI tools like ChatGPT. Some examples are listed below, but check with the individual style guides for more detailed instructions on in text citation and variations. 

APA 7:

OpenAI. (2024). ChatGPT (July 8 version) [Large language model]. https://chat.openai.com/chat

MLA 9:

“Summarize the book Little Women” prompt. ChatGPT, 8 July version, OpenAI, 30 Jun. 2024, chat.openai.com/chat.

Chicago 17 (Numbered Footnote): 

For prompt included in text (URL optional):

1. Text generated by ChatGPT, OpenAI, July 8, 2024, https://chat.openai.com/chat.

For prompt not included in text:

1. ChatGPT, response to “Explain how to make pizza dough from common household ingredients,” OpenAI, July 8, 2024.