Living with Artificial Intelligence
The use of Artificial Intelligence (AI) has obviously become more rampant and more ubiquitous in the last couple of years. There has been lots of valid concerns coming from it such as jobs getting replaced by AI, AI being sentient, and how much water AI consumes to fulfill your request. You might also have seen those ridiculous AI-Generated videos like an interview or an unusual event having absurd background and oddly unnatural voice and movements of the entities involved. Usually seen on short-form content platforms such as Reels and Tiktok. This use of AI is of course, controversial—the AI slop.
But what I would like to tackle about in this post is the use of AI in academic setting. More specifically, the students’ use of ChatGPT, Gemini, Grok, Claude and among others to do their academic work for them.
Just a bit of disclaimer.
- I am broadly not anti-AI myself.
- I do not condone academic dishonesty.
- I am not an expert in AI and designing an academic curriculum.
I just think that there is a gap on how AI generated text is viewed, used, and detected.
AI, LLM, Machine Learning?
To avoid confusion, I think it is important to clarify what these terms technically mean. Not trying to sound like a PhD here—I just think it is important to make sure we’re all in the same boat. AI is an umbrella term that encompasses technologies (like machines, typically computers) that aim to mimic human knowledge or behavior to attain tasks that typically require problem solving skills, reasoning, and creativity. Machine Learning is one of the methods for developing AI Systems. Machine Learning relies on big data for training their AI models—such as the Neural Network model. Machine Learning is built upon deep learning which is driven by neural networks. The jargons go on and on. You can do your own research to demystify these concepts.
For now, you can think of Machine Learning as a method of teaching computers on how to think for themselves given with a large volume of data. It’s like giving the computer the ability to think like a human, so they can create their own inferences and predictions. Giving them human brain-like power. Machine Learning is what powers these Large Language Models (LLMs). The models used by AI Chatbots like ChatGPT.
Just note that I will be using the terms “LLM” and “AI” interchangeably. Within the context of this post, they will generally refer to LLM Chatbots such as but not limited to ChatGPT.
LLM Detectors, on the other hand, are services who claim that they can determine the ratio on how much of a written text is generated by AI versus how much it was written by a human. Ironically, they typically also come with an AI Humanizer service.
LLM use in Academic Setting
One of the main concerns of the use of LLMs in academic setting is its possible use for academic dishonesty. That a student might just use LLMs to write an essay which the teacher gave a week ago and the deadline’s tomorrow. Maybe use an LLM to solve complex mathematical equations instead of building the neurons in the brain to recognize the mathematical patterns. The concern is valid. Expediting a work that was supposed to make your brain think by using an LLM because the deadline is in the corner defeats the purpose of learning. Or maybe the student is just simply—lazy. But even if the student used AI, can an AI Detector really detect it? Considering that even OpenAI killed their own AI detector due to its low accuracy. How much can we trust AI Detectors if the same detector offers an AI humanizer service to counter it? It’s like the same vendor sells both the swords and the shields to the bandits and villagers. The vendor does not really care who wins. The vendor only cares about gaining profit with the villagers’ problems (Selling swords to the bandits). At the same time, selling the solution for it (Selling shields to the villagers).
This is just my analogy. I do not really know if ancient bandits did buy swords or ancient villagers did buy shields.
“Should we just impose a total ban on AI in schools and universities? But wait a minute, even teachers and professors can use AI too. Does that make it a tie? Maybe the use of AI in pedagogy is fine, but not for academic dishonesty.
Well, maybe I am asking the wrong questions.
AI (LLM) Detectors
How do they work?
TLDR: They make assumptions by comparing the writing styles of humans to Artificial Intelligence.
LLMs generate texts by performing computations that mathematically predict what the next word will be. Making AI-Generated texts more consistent and predictable. Detectors analyze text for patterns if it looks like something an algorithm would generate. Below are what I found the most common ways on how AI Detectors work.
I. The Core Metrics Overview
- Perplexity: Measures how “surprising” the text would be to an AI model. Since AI-Generated texts are often more predictable while human-written texts vary more and can make unexpected writing choice.
- Burstiness: Refers to variation of length and structure of sentences throughout the text. AI tends to write with more consistency and little to no variation in sentence structure and length—low burstiness. Texts with more variation are scored high burstiness.
AI-Generated texts tend to be low in perplexity and burstiness.
II. Classifier Model
Using LLMs to catch an LLM. Feed an LLM model with texts labeled “Human” (Reddit posts, forums, etc.) and “AI” (ChatGPT Outputs) to teach it what a human-written text is supposed to look like vs AI. Having an LLM act as Machine Learning Classifier.
III. Watermarking
Some AI providers might embed a subtle statistical signature within the text generation which can then be later reviewed. Checkout Google’s SynthID
The Problem with AI Detectors
Principal
The main problem that I see with AI Detectors is it is just not that reliable, yet it somehow becomes a de facto solution to refrain students from using AI. This pseudo-solutioning to address academic dishonesty with AI is, in my opinion, lazy. AI Detectors just cannot prove it absolutely. Insights from them might be helpful for analyzing writing tone, but they should not be the sole and main basis of academic integrity. If we fully rely with these tools, we risk students getting falsely accused of cheating. I know personally some of my peers had to rewrite their own work because it sounded robotic, having low perplexity and burstiness. Which is usually the case in formal academic writing (see the 1776 U.S. Declaration of Independence gets flagged as AI case) and for non-native English speakers (a study from Stanford).
Is using AI cheating?
The answer for this one is obvious and honestly just common sense. It depends. If you solely rely on AI to do all the heavy lifting for you, then yes it is cheating. But if you use them to gather data, do research, brainstorm ideas, and enhance your writing style, then no. The rule of thumb for me is if you know and truly grasp what AI gives you, and you make your own original work with the information you acquired from it, then it is not cheating. Do your own work. Even AI Companies admit their LLMs can make mistakes. At the end of the day, it is our responsibility to double check the information and it is our accountability on what we do with it.
We have been using paraphrasing tools (also powered by LLMs) to enhance our writing skills mostly for formal and academic registers such as writing an email or writing a research paper—even before ChatGPT existed. Was it viewed as cheating? Generally, no. Because using paraphrasing tools at that time only implied that you came up with the intellectual idea. Provided that you tried to use your own words, paraphrasing tools only improved the articulation for it. Which was very helpful especially for non-native English speakers. We almost had no problems with it from an academic integrity standpoint.
AI Humanizers
AI Humanizers are designed to rewrite a supposed AI-Generated text to make it sound “more human” by increasing perplexity and burstiness. However, they can just make your work worse. They will add poor choices of words, awkward phrasing, and nonsensical sentences. Potentially losing the original meaning from your original work. Basically making your work “dumber” to bypass AI detection. The writing tone then may not be quite suitable for academic writing such as a thesis or essay. A poor-worded essay or research paper will not be graded well. Despite that, it is not even a full guarantee that it will bypass AI detection. AI Detectors also get updated, some even have labels like “AI-Generated and then Humanized”. And even if in theory, it does work perfectly all the time, we cannot still tell apart whether a “humanized” text originally came from an LLM or was genuinely written by a human but was forced to be fed to the humanizer to evade false “AI-Generated” accusation.
What about Plagiarism Detectors?
Pre-ChatGPT era, we were broadly only worried with plagiarism. Plagiarism Detectors make sense because they rely on a database of evidence while AI Detectors just make assumptions. When a plagiarism Detector says your work is plagiarized, it can provide you with a link to the source that almost exactly matches your texts. It literally scans a large database of documents to find near-exact text matches with your work. Plagiarism Detectors also reinforce the ethics of not taking the work of others and make it appear like yours by citing your sources properly.
Meanwhile, AI Detectors just say “I think your work is 80% AI because it resembles writing patterns of a computer. I think this part of your work is likely AI. I am not sure though. I can’t tell where it came from. I might make some mistakes. You might want to use an AI Humanizer also found within this website to make your text sound more human. Apparently, humans are not that literate”. Paradoxically, AI is trained on millions of academic papers written by humans to sound “smart”. So if you sound smart, you sound AI.
The Pedagogical Pivot
Learning with AI
Learning process differs to everyone. Instead of figuring out whether a student used AI, use the time to see their learning progress as a proof of work. Below are my two cents on what I think are better alternatives than having an AI score.
- Require Process, not just the Product: For writing assessments, have drafts and outlines required or showed for the final output. Drafting shows the thinking process. It serves as an evidence of the work’s progress from being an idea to a refined product. For engineering works, have Git commits shown to see how the code changes over time, schematic diagram drafts to show iterative design process, make scratch papers notes part of the documentation, etc.
- Integrate, not ban: Utilize AI tools to save the boring or repetitive parts while keeping the learning progress intact. Encourage learning techniques such as mind mapping, active recall, and Feynman technique while having AI assist with the process.
- Reflection within local and personal context: Encourage students to write their own opinions about something context-specific topics such as campus stories, volunteering experience, and personal experience. AI will not easily grasp the context because personal experience and campus gossips are not a shared or common context.
For the first one, it might be less convenient to require the working process documentation than having a printed AI score passed along the submission. But if we’re prioritizing addressing short-circuiting the learning process, I believe it is just reasonable. For the second and third, even though in opinion, they might have been widely practiced already, it is good to include them in case a total ban on AI is imminent.
The Caveat
No matter what solution is employed to prevent academic dishonesty, slackers will always find their ways to bypass systems that aim to impose disciplined learning. Whether the world we live in today is AI-driven or not, it is always the students’ choice whether they will half-ass their work or not.
Academic Reform
Encourage the students to genuinely learn. Not just consume learning materials. Particularly in the Philippines, a PISA result in 2022 shows that there has been shortcomings with regards to basic education (K-12) in the country. It was mentioned in the report that there are factors that affect the students’ academic performance including students’ well-being and socio-economic status. Another factor is the curriculum being overloaded, leaving students with less time to master fundamental skills. There has been a reform to address the country’s gaps with regards to basic education. The matatag curriculum aims to address the poor performance and learning gaps. This is beyond the scope of the topic we’re discussing right now, and I actually think this deserves its own post. But this is worth mentioning as AI could become a shortcut solution for affected students to circumvent their current limitations.
AI Effects on Student Life | My take
Despite the rise of AI, the essence of being a student is still intact. I believe that AI barely influences the subjective experience of being a student. The vast majority of a student’s life will be spent on building relationships. In school, a student gets to learn how to collaborate, communicate effectively, and resolve conflicts. There are also times where they have to work solo. Which I believe builds understanding of oneself and independence. Skills that are invaluable in building relationships personally and professionally. Learning how to establish interpersonal and intrapersonal understanding, for me, is just as crucial as absorbing the body of knowledge.
Students often have this “love-hate” relationship with their academic institution. This “love” is often rooted in the relationships within friends and peers. Sharing the same struggles, missions, priorities, and passions. Making them feel like they’re not alone in the battlefield. Having a supportive and trustworthy circle in school or in life is valuable—just be wary of whom you spend your time with.
The “hate” comes from the burden of “you have to”. In my country particularly, there are cases where students are often desensitized to the external factors and are expected to prioritize studying even during a typhoon—the Filipino resiliency culture (another wholly interesting topic). In college, life starts to become the preview of adulting—the soft launch. Academic institutions are often bureaucratic and more often than not, unforgiving. Stringent enrollment processes, crappy systems, terror professors, deadlines, grueling schedule, financial stress, long commute, just to name a few. It dictates a student’s life. It teaches students that joy is not the only thing in life. That they have responsibilities to fulfill and the world demands results. The joys and struggles. The experience that AI offers no alternative, is what I believe, makes a well-rounded person.
AI hardly touches these aspects of life as a student.