Mind Readings: AI Content Detectors Deep Dive Part 1

Okay. We need to have some conversations about artificial intelligence, particularly generative artificial intelligence, so generative AI, tools like chat GPT and academia and the workplace and plagiarism and AI generation and copyright infringement and kicking people out of school for using AI to write papers or not hiring them for jobs. Because there's a lot of people and a lot of companies doing a lot of things really, really wrong. So this is a five-part series. We're going to talk about the implications of all this stuff, set some definitions, do some examples of different uh infringement detection tools, or actually I should say AI detection tools, because we're we want to separate all these things out.

Um, and show that the act of trying to detect AI is ultimately pointless and harmful. So let's get started with part one definitions. When we talk about the use of AI, particularly in an academic context or in a hiring context, we are talking about uh what people will mention terms that they conflate that should not be conflated, like plagiarism, like uh academic honesty, uh copyright infringement. So let's set some definitions. And to do this, we are uh well first I'm gonna put up the warning banner here.

I am not a lawyer, I cannot give legal advice. If you have a uh if you require legal advice, uh seek a qualified attorney in your jurisdiction for advice specific to your situation. That's really important. I am not a lawyer. Uh and what I say, I mean, I'm just another guy on the internet, so but one that has a lot of experience and expertise in artificial intelligence.

To begin with, let's talk about plagiarism. Plagiarism and AI use are not the same thing. Plagiarism, according to Legal Information Institute from Cornell Law School, is the act of taking a person's original work and presenting it as if it was one's own. Plagiarism is not illegal in the United States in most situations. Instead, it is considered a violation of honor or ethics codes and can result in disciplinary action with schools, person's school or workplace.

It can warrant legal action if it infringes upon the original author's copyright patent or trademark. So that is plagiarizing. Generative AI does not do this. What is inside a generative AI model, like the models that power Chat GPT or Anthropic Claude or Google Gemini are massive, massive piles of statistics. Statistics and data that form statistical relationships among trillions and trillions and trillions of different word and sentence and paragraph combinations.

The amount of data that is used to train artificial intelligence. This is Shakespeare's complete plays, right? This is 800,000 words. To train a generative AI model, today's models, you would need enough of these to go around the equator of the planet twice. That's how much text data they are trained on.

When you use generative AI to spit out term paper or whatever, it is not plagiarizing because it is not pulling A, it is it is not um presenting someone else's original work. You are getting just a pile of statistics. Um you could still make the case that um someone misrepresenting an AI output as their own. I mean, yeah, that if you are saying you wrote this and you did not write it, you could say that that is dishonest. That would be an accurate statement, but it is not plagiarism because what comes out of AI is not the original work, and it is not, and you know what as a result you can't present it as though it was uh someone else's original work.

So that's number one. Number two is copyright infringement. Again, back to Cornell Law School, Legal Information Institute. Infringement of copyright refers to the act of unlawful copying of material under intellectual property law. It is an act that interferes with the right of intellectual property ownership.

A copyright owner has the following rights to reproduce their works, to prepare derivative works based on their original work, to distribute copies of the copyrighted work, to perform certain copyrighted works in public, to demonstrate certain copyrighted works public, to perform the copyrighted work for sound recordings, and to import copies into the United States. So to prove to bring a copyright infringement claim, the plaintiff plaintiff must prove that they hold the copyright interest through creation assignment or license. They must plead that the complaint is of an unlawful copy of the original element of the copyrighted work. To constitute an infringement, the derivative work must be based upon the copyrighted work. Again, this is where generative AI is kind of in a in a new area, and and there is no settled law on this.

Generative AI, again, if you go into the models themselves, if you look what's inside, it is not the original work. There you will not find any original works in a generative AI model. You will find a pile of math. And as you use a generative AI tool, it is invoking copyright. Now you can't do this in your average web-based service.

You can only do this in open models. So I've got the uh cobalt AI. I'm using Mistral Straw Small Instruct, which is a model made by the Mistral Company of France. And what we're gonna do is we're gonna we're gonna give it a simple prompt. Let's give it a prompt like um, how does the Supreme Court of the United States of America impact the economy?

Right? That's a pretty straightforward prompt. And the model is going to start responding. Now, as it responds, if we look at what's happening behind the scenes, let me put this side by side. It is not copying anything.

What instead is happening is it is if you look carefully, it is guessing what the next logical word might be based on the statistical database. Let's scroll back down. And so with each word that it creates, it guesses based on all the previous words what the next likely set of words are going to be. So that's what's going on behind the scenes. This is not copyright infringement if you go by the legal definitions, right?

Uh, because it is not reproducing any original works. And as machines create their output, is it this is the sentence that it all hinges on is what a model creates a derivative of the original work based on the training data. In some places in the world, the answer to this is yes, in the EU in particular. In some places in the world, the answer is no, Japan and China. In the United States, there are a number of lawsuits right now about this.

But um, Dr. Crystal Laser at Cleveland State University, who specializes in uh digital copyright, said we won't have an answer to this question for probably 10 years for all the current cases to work their way through the system and to arrive at settled law. The third thing is academic honesty. And I'm going to use my alma mater, my my bachelor's alma mater, Franklin Marshall College, and look at their academic honesty policy. And the policy is unauthorized aid, making use of prohibitive material study guides or other assistance in academic exercise, for example, obtaining test questions before the exam is given.

Um, that would be uh a violation of academic integrity, plagiarism, reproducing the work or ideas of others and claiming them as your own, right? So claiming authorship of a piece of writing or created by someone else. This is where it is insufficiently clear whether an AI output would be considered plagiarism under this policy. It's not, it doesn't clearly say no AI. It also doesn't say yes AI.

It says nothing about it. And so for institutions that are concerned about the use of AI within academics, you gotta be clear. You gotta be clear how it should and should not be used. Can you use it to brainstorm? Can you use it to write an outline?

Does the final product need to be all the students' fingers on the keyboard? Um there's a few other things here, but this was last updated about a year ago, so well after the invention of Chat GPT, and it's unclear. So to wrap up part one, generative AI is not plagiarism, at least not by the strict legal definition. It might be copyright infringement, the the creation of the models themselves. The works that they prepare, almost certainly not because you can't trace the output of an AI back to any one particular work.

You can't demonstrate like this is this came from here, and the legal liability for the infringement is going to be on the model makers, not the student or the employee. Academic honesty is kind of vague. It's kind of vague. So I think it's important that we set these definitions and that anyone who wants to argue for or against artificial intelligence be clear on these definitions first. And if you're unclear, you need to set policies and then explain the nuances of those policies.

Now, there are a number of institutions that have had declared just blanket no use of AI at all. We'll talk about that in part five of this series. But suffice it to say that the workforce that your students will be going into, or that the companies that your your employees will be working at are using these tools. 77% of knowledge workers, according to Microsoft's uh 2024 work trend index. 77% of employees in knowledge work jobs are using generative AI with or without their company's permission.

So if you are matriculating students who do not have this critical skill, 71% of uh CEOs said in that same report that they would not hire someone who has no AI skills, and they would choose a less experienced candidate with AI skills over a more senior candidate without AI skills. If you matriculate students who do not have AI skills, you are doing them a disservice. So let's let's set that expectation. All right, that's the end of part one. Let's uh let's take a break and we'll come back for part two.

What is the point of writing all these term papers and other things where you would use AI? So talk to you soon. If you enjoyed this video, please hit the like button, subscribe to my channel if you haven't already, and if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.