Is Using AI Plagiarism?

Generative AI plagiarism exists as a possibility due to the nature of generative AI algorithms. 

AI plagiarism, therefore, is a real and growing concern. The US Government Accountability Office lists [1] four focal points it regards as key to ethical AI use:

• Governance

• Data

• Performance

• Monitoring

Is AI plagiarism the predictable outcome of a rising new technology that you can avoid by careful programming? Or, is AI plagiarism less a question regarding the technology itself than the philosophical underpinnings of the concept of plagiarism? 

What constitutes plagiarism?

The traditional definition of plagiarism refers to passing off someone else’s work as your own, either knowingly, with the intent to deceive, or by omission (i.e. by declining to attribute portions of your work to others or by doing so incorrectly). 

But can you ever accuse AI of plagiarism? After all, AI has no intent, no agency, and no ethics and receives neither plaudits nor punishment regardless of the relative originality of its output: Programmers train it on freely available data, and it makes decisions about that data by means of complex statistical computations. 

Understanding AI content generation

Leaving those deeper questions aside for now. It’s important to first understand how AI generates content. 

How AI generates content

Programmers input massive amounts of data into an AI algorithm. By means of large language models (LLMs) and deep learning via neural networks designed to mimic the structure of the human brain—and, therefore, theoretically, the mechanism of human knowledge acquisition—an AI model learns to predict which word would logically follow the previous one in a sentence based on the patterns it detects in the training data. This is something like an autosuggest feature, albeit far more sophisticated. 

AI models don’t think or learn the way humans do: They are sophisticated predictive algorithms that take a recombinant approach to content generation. What they are doing is performing statistical analysis. But if you train them on copyrighted material, they may violate norms—and even laws—regarding intellectual property. 

Attribution and ownership

If AI isn’t thinking or creating in the traditional sense, then to whom do we attribute its output? To the owners of the data programmers trained it on? To the AI prompter? To no one? 

Rights to AI-generated content

This is an open question. If a human-like AI generates human-like content, does it follow that an AI programmer has a human-like right to claim that content as their own? Or, is it the writer’s intellectual property by dint of their being the one who prompted the AI, which, after all, is more like a library than a human being capable of thought and creation? 

No one will issue a patent for work derived wholly from AI; experts consider such work part of the public domain and, therefore, ineligible for a copyright. However, if an AI-based work contains human-created elements, then those elements are eligible for copyright protection. 

All of this implies that in order to copyright your own self-created portion of an otherwise AI-derived text, you will have to be able to prove that you indeed wrote that portion of the text, which may not be easy. 

Academic integrity and AI

Students who plagiarize others' work can receive failing grades. In some cases, authorities can even suspend or expel them from school. 

Plagiarism used to be reasonably well defined if not always well understood by students. Since the advent of generative AI, however, teachers and academics have had to debate the development of new norms and protocols regarding AI plagiarism, not just in terms of student work but also in regard to their academic peers. 

Ethical considerations in academia

AI plagiarism policies differ broadly. Some universities have issued official statements on the matter: 

• San José State University considers the use of AI to write a paper to constitute plagiarism. 

• The University of South Florida considers the use of generative AI to be closer to ghostwriting than out-and-out intellectual theft, but would still be considered as cheating if submitted for assessment. It recommends a redefined plagiarism framework centered on intentional or careless lack of acknowledging sources rather than theft. 

• Arkansas State University is concerned that AI may plagiarize resources, lack attribution, and provide outright erroneous citations—hallucinations are a known bug of generative AI models. 

Comparing AI-generated and human writing

What really demarcates the difference between AI-generated and human writing? 

Originality and creativity

AI doesn’t have creative intent. It has no intent whatsoever, being a machine-learning program allied to a predictive output algorithm. Neither does it have the intent to deceive anyone by passing off someone else’s content as its own. AI is amoral—it cannot act either morally or immorally. 

Furthermore, AI never directly copies someone else’s work: It’s just an algorithm that pulls from a huge number of sources; any resemblance to anyone else’s creative effort is purely coincidental. Think of the old image of an infinite number of monkeys at typewriters, eventually writing a Shakespearean tragedy word for word. It’s a matter of probability which, given enough queries and time, may be possible. What is at issue is your tolerance for uncanny similarities between that monkey-derived tragedy and the original. 

Can you tell the difference between human-written and AI-derived content in the first place? It is possible. AI content tends to be: 

• Intellectually shallow

• Repetitive 

• Monotonous

• Cliché-ridden

Then again, human-written content can be just as bad. That isn’t the issue here, but it means that it’s not necessarily easy to, at a glance, tell that AI has plagiarized someone—especially if that person writes, in the first place, as blandly as AI. 

Detecting AI-generated content

You may be able to uncover plagiarism by using AI-generated content detection tools. 

Tools for identification

Examples of AI plagiarism tools include: 

• GPTZero

• Winston AI

• Quetext

• Smodin

• Scribbr

• Originality

• Copyleaks

• Turnitin

Most AI detectors utilize LLMs similar to those that programmers train generative AI platforms on. They look for the relative presence of two factors: 

• Perplexity: This is the degree to which AI-generated text is unpredictable—that is, how likely it is to perplex a reader. Low perplexity is the aim of AI; however, this content is highly predictable and lacks human flair. Low perplexity may mean greater correctness, but it indicates that a piece was likely AI-derived. 

• Burstiness: This measures the variability of sentence structure throughout a text. When writing at length, people tend to utilize sentences of differing lengths, as well as a variety of grammatical structures. AI generates sentences that are almost uniformly similar in structure. Low burstiness, then, suggests AI-derived content. 

It is important to note that such AI detection tools aren’t foolproof. Even if they do detect that a piece was highly likely to have been AI-generated, that still leaves you with philosophical questions regarding the nature of plagiarism in an AI environment. 

Information retrieval (IR) software already exists to combat plagiarism. By utilizing similar principles, some experts hope to fit generative AI frameworks with logs that track and, so to speak, watermark the content from which they generate their output. 

Ethical use of AI tools

Generative AI sometimes produces “hallucinations”—mistakes based on an algorithm’s misunderstanding of the logical connections between data values. In addition, unethical writers sometimes purposely edit AI-created content in a misleading way, peppering it with misinformation, bias, or libel; someone who implicitly trusts the platform that content is on may then act in ways that harm themselves or others based on erroneous information derived from a supposedly trustworthy source. 

Some people can even purposely create fake content in order to mislead, punish, or exploit people and companies. All of which is to say that the world of AI content creation is a fraught one. 

AI transparency

The lack of AI transparency potentially leads to the large-scale erosion of trust—in governments, in tech, and in one another generally. This would seem to be a natural and predictable consequence of people believing that any media they come across could be an AI-generated illusion. 

To combat this possibility, many people believe that they should easily be able to discover where a piece of content comes from—and determine whether it was AI-generated and, therefore, potentially suffers from plagiarism concerns. People harboring such ethical concerns may decline to work with a business that leverages generative AI. In ways not yet fully quantifiable, AI plagiarism may result in disastrous economic consequences for those it plagiarizes. 

One way of enforcing transparency in AI content is via a provenance scheme. This scheme essentially watermarks the content on which programmers train an AI algorithm. A provenance scheme allows you to trace training data back to its source, make decisions about the reliability of the source data, and cite the human originators of the data. 

Legal implications

Plagiarism—especially copyright infringement—has long had serious legal ramifications. 

Copyright and liability

While plagiarism rarely rises to the level of a legal matter, it is a serious problem. 

Most plagiarism cases that do make it to court are misdemeanors, which can result in fines of up to $50,000 and as much as one year of jail time. Some plagiarism cases even rise to the level of felonies [2]. If a plagiarist makes more than $2,500 from their work, they may face fines of up to $250,000, as well as up to 10 years in jail, depending on the state in which the plagiarism took place [2]. 

Getting started in AI with Coursera

If you’re ready to learn more about AI plagiarism, look on Coursera. 

Consider starting with DeepLearning.AI’s course, Generative AI for Everyone. In it, you’ll learn how to think through the lifecycle of a generative AI project, from conception to launch, including how to build effective prompts.