GPT-3 Wrote an Entire Paper on Itself. Should Publishers be Concerned?

Research indicates that businesses have been relying more on artificial intelligence (AI) over the past few years. The average number of AI capabilities that organizations have embedded within at least one function or business unit doubled from 1.9 in 2018 to 3.8 in 2022.

And it is not just businesses; even the general public is taking notice. AI models like GPT-3, DALL-E, ChatGPT, and Alphacode have been the talk of the town on social media. So, it’s no wonder that advancements in generative AI are now having an effect on science and academia as well. A researcher got GPT-3 to write an entire paper with simple prompts. The paper was initially rejected by one journal after review, but it was subsequently submitted to and accepted by another — with ChatGPT being listed as an author — a trend that’s becoming more common these days. 

GPT-3, or Generative Pre-trained Transformer 3, is a Large Language Model that generates output in response to your prompt using pre-trained data. It has been trained on almost 570 gigabytes of text, mostly made up of internet content from various sources, including web pages, news articles, books, and even Wikipedia pages up until 2021. 

So when you enter a prompt in natural language, it uses the training data to spot patterns and then gives you the most appropriate response. You could use it to complete sentences, draft compelling essays, do basic math, or even write computer code.

In this article, we will discuss the impact of GPT-3 and related models on research, the potential question marks, and the steps that scholarly publishers can take to protect their interests.

person using a laptop with a visual overlay of ChatGPT

The impact of GPT-3 on academic research

The model has been around since 2020 and has already been used to develop a range of new applications, such as chatbots, translation tools, and search engines, among others. 

Perhaps the most talked about feature has been its ability to draft human-like essays. You could generate an original piece with a basic prompt like “draft an academic essay in about 800 words on how AI is impacting academia.”

GPT-3’s deep learning algorithm allows it to write from scratch, auto-complete your sentences, and build upon what you have already written. Microsoft plans to integrate the model into its productivity suite, which includes MS Word, Powerpoint, and Outlook. So far, they have added a version of the model to their Edge browser, Bing search, and Teams, their collaboration tool. Other tech giants, such as Google and Amazon, are also pushing forward in the AI space. Google is set to launch Bard, a conversational AI service, while Amazon Web Services is partnering with AI startup Hugging Face to create cost-effective AI applications for customers.

Now, this is only a tiny part of the GPT-3 story. Since the release of its open API, research labs and companies worldwide have been building new applications powered by the model. ChatGPT, a sibling model of GPT-3, is completely changing how we interact with machines. Its dialogue-based approach has got the public hooked, as it enables them to get things done with prompts and questions in plain language — a change from menus, buttons, and predefined commands. 

These developments are sure to change the writing workflow. Identifying who wrote what — human or AI — will become challenging. Turning ideas and thoughts into fully fleshed-out points on a document won’t be the same as before with AI working alongside us.

Bringing this technology to academic circles raises complex questions. Can GPT-3 be listed as an author? How does copyright play a role in this? What about the ethics of such usage?

On the plus side, non-native English speakers will have an easier time overcoming the language barrier. They will be able to produce high-quality research papers without worrying about grammar or syntax issues. Moreover, AI-assisted writing can help researchers save time, allowing them to focus on refining their ideas, framing their arguments better, and conducting more in-depth analyses.

Additionally, you can even command the model to format the output in a certain way. And manuscript formatting is an activity that typically takes up to 14 hours per paper

In short, these capabilities allow researchers to complete their manuscripts much faster and share breakthroughs with the world more quickly.

This has led to the creation of a whole new range of applications: from developing spreadsheet formulas and creating Python code to writing SQL — all from simple text prompts. That’s not all, you also have tools to help you with your literature search and reading process.

At this point, I’d like to disclose that I run SciSpace. We recently added an AI assistant to our research paper repository. It helps break down, summarize, and translate research papers, as well as explain math, tables, and text. And we’re not alone. There are also other tools out there that can help extract more information from research papers.

Beyond that, there are also models like DeepMind’s AlphaFold that can make protein structure predictions and OpenAI’s Codex that can solve complex university-level math problems or provide coding assistance

The reasons to be wary of GPT-3

The critics of GPT-3 have raised numerous questions about the output generated by the model, from plagiarism and bias to a lack of reliability. And rightfully so.

A 2021 investigation into articles published in Microprocessors and Microsystems revealed that the journal published nearly 500 questionable articles. The study showed that they contained broken citations, scientifically inaccurate statements, and nonsensical content, rendering the papers non-reproducible. The investigators believe that authors may have used GPT and reverse-translation software to hide plagiarism and to enlarge their manuscripts.

Another concern is the potential for bias in GPT-3 generated output. The model is trained on unstructured web data. So it can easily borrow from existing stereotypes and beliefs about various subgroups, like races, political ideologies, religions, or genders. Past investigations have revealed instances of severe bias, leading to offensive output. So using these models for research purposes has the potential to pollute science with discriminatory language and unwarranted homogenization.

The model was trained on data from 2021, so unless you give all the right pieces of information in prompts, it might provide you with outdated output. Also, GPT-3 tends to hallucinate, meaning it produces an output that doesn’t make sense or isn’t true. For instance, when you ask questions about a particular theory and why it was derived, the model may respond with something completely unrelated or nonsensical. 

Why does this happen? It comes down to the fact that the internet contains our thoughts, data, and facts but not the reasoning, logic, or context to truly make sense of them. So, GPT-3 has no way of knowing what is true or correct or why something is the way it is. And the model ends up producing probabilistic output without understanding the context around the question.

One way to avoid this issue is to use the chain of thought prompting technique, which involves providing the model with examples and instructions that help decompose the problem into smaller steps, eventually leading to the correct answer.

There are also other ethical and moral concerns. Is it right to use AI to write papers when publishing papers are used as a barometer of researcher competency, tenure, and promotion? Also, if an author uses an AI tool to write papers, does it mean the tool should be credited for the work and not the writer?

What can scholarly publishers do?

First, it is crucial to recognize that:

  • Large parts of academia run on publish-or-perish mode
  • Paper mills and predatory journals are not going away
  • English language dominates the academic and scientific discourse

GPT-3 and other AI models are evolving and hold tremendous potential for academia. However, writing-related AI technologies aren’t new — Google Docs, MS Word, and mobile keyboards have provided word and phrase suggestions and spell checkers, and grammar corrections for a while now. GPT-3-powered writing tools are now taking it further: rather than offering a list of words to choose from, they enable AI to anticipate and finish entire sentences and paragraphs probabilistically.

But at the same time, scholarly publishers need to protect the integrity of their journals from manipulation, disinformation, plagiarism, and bias.

Here are some steps that publishers can take to ensure their continued success in the face of the changes brought about by GPT-3:

  1. Use AI tools for quality control: Integrate AI tools into your internal screening workflow as the first line of quality control. Use them to determine if the paper meets the journal’s scope, detect text overlap and plagiarism, detect formatting and grammar errors, and assess the appropriateness of experimental design. It should help editors and peer reviewers deal with the deluge of submissions, trim their workload, and focus on the most relevant papers.
  2. Establish a clear framework: Formulate policies around the usage of AI tools. It should outline the acceptable research methods, the ethical standards that authors must adhere to, and the consequences of non-compliance. Moreover, if a publisher plans to use AI tools in their workflow, say to locate relevant peer reviewers, then they must clearly outline how to reduce the risk of bias or prejudice in the process. 
  3. Monitor existing papers: Take the help of research integrity experts, AI sleuths, and AI image detection tools to ensure that published articles are free of image duplication fraud, nonsensical content, or tortured phrases. Retract the papers that fail to meet the standards of your journal.
  4. Educate authors: Research paper writing and submission are tedious activities. Often researchers may need help with what to do. Create a blog or a YouTube channel and use that to address these knowledge gaps and ambiguities. Also, use that to build awareness of paper mills, predatory journals, and the ethical and moral implications of using AI tools. Tap into existing resources created by organizations like the COPE and CSE, who share practical advice and assistance around publication ethics, to ensure submissions align with accepted standards.
  5. Offer additional services: Since most papers are published in English, non-English speakers are forced to write English for academic success. Many see this as a burden, making communicating new ideas and insights difficult. Publishers can turn to AI-enabled translation tools, like DeepL, to capture the subtlest nuances of language and retain those nuances in the translation. This will enable them to receive more submissions, get publications ready faster, and ensure that non-English papers remain true to the original intent.
  6. Encourage Open Access: Urging authors to archive their pre-print in a repository like ArXiv or share their datasets in Zenodo will help promote transparency and openness. The greater visibility will lead to more call-outs and expose any suspicious behavior. For paywalled papers, publishers should have a dedicated internal team verify the raw data, seek reader feedback and monitor the web for commentary to ascertain accuracy and credibility.
  7. Check the integrity of submissions: Make sure all the papers in the backlog run through a GPT Detector. It should help identify authors who use AI to shape the core theories of their manuscripts. Also, use databases like Dimensions, Scopus, and Web of Science to detect fake or made-up citations — common occurrences in GPT-3 generated papers. AI often cites papers that do not exist or are unrelated to the topic.

By following these steps, publishers will be better equipped to identify potential issues and establish policies that ensure the integrity of their publications.

Final thoughts

Given the pace of development, the role of AI tools in scientific research and communication will only increase in scope. The jury is still out on how good or bad the impact would be. 

On the one hand, it could democratize research and knowledge. And on the other, it could worsen information overload and enable more people to take advantage of shortcomings in our educational systems, which often reward quantitative achievements.

Scholarly publishers and other stakeholders will need to carefully evaluate the impact of AI tools and take the necessary steps to ensure that its usage does not lead to fraudulent activities or unethical research practices.

Editor’s Note: This article by Saikiran Chandha, CEO and founder of SciSpace — the only integrated research platform to discover, read, write, and publish research papers, is republished with permission of the author. First publication was on Scholarly Kitchen.

Posted in: AI, Education, KM