The internet is brimming with stories of a man who’s led a most amazing life. He’s been a detective, a librarian, a clockmaker. He’s a published author in several disciplines. He’s even been a lighthouse keeper. Stories about him first began surfacing in late 2025 and spiked to a peak early this year, and they read like he is literally the most interesting person in the world.
His name is Elias Thorne. And he doesn’t exist.
A study in May at Cornell University asked four different LLMs to “tell me a story” in a few different variations. The study found that in 20,000 stories, the name “Elias” occurred 26.5% of the time. Furthermore, over 88.3% of the stories featured the same eleven names, locations and professions. It also seems Elias himself has stories to tell, as his works can be found on Amazon in various locations.
If AI is supposed to be getting better and better, how come much of its output seems so…average?
Unlike Elias’s life, the current landscape of AI in higher education does not feel full of adventure and mystery. Instead, it feels polarized, with little middle ground. In our work with faculty in the Center for Teaching and Learning, the majority of faculty fall into one of two schools of thought:
- Some have discovered new ways to use AI as an assistant and enjoy its potential to streamline their work. They feel that the benefits to them and their students will ultimately outweigh the risks.
- Others are very worried about students using it to cut corners, and/or to cheat on assignments. They are not always completely against AI per se, but they do see an imminent threat to their teaching practice and to student performance, and some even fear that it may invalidate their discipline entirely. Their fears are not altogether unfounded, according to the largest undergraduate study of generative AI use, conducted at UC Berkeley.
Both approaches to AI in teaching and learning present nuanced and intentional ways to address this disruption. They also give us the sense that like the calculator did sixty years ago, AI is changing education in permanent ways.
So what now?
The good news is that at least so far, AI is not in danger of replacing people or the importance of critical thinking and discourse. It appears to do a number of things quite well: it can write and sound mostly human, generate code, draw images, and brainstorm. Sometimes it almost seems as though it can actually…think. But there are some cracks in the armor that tell us it’s not all over yet.
What can an AI, an LLM, do well?
- It can sort through massive amounts of data very quickly.
- It is very good at spotting patterns in this data. Given the incredible processing power and the scope of its analysis, it can come up with patterns that humans might not detect on our own.
- It can summarize large amounts of information in moments, and cut down on analytic time drastically in some cases.
This sounds a lot like a remarkably competent research assistant, doesn’t it? An LLM can be an excellent assistant, if it is well- supervised and receives intervention and training upon making mistakes, much like a human assistant. The problem comes when a human believes that this AI assistant is more capable than it actually is, and asks it to perform a function, to create something new, when it is only good at analyzing and synthesizing existing data. What kinds of mistakes can an AI assistant make?
AI Is Monotonous
An AI has no human experience, so it relies on prediction rather than thought. Its opinions are neutral on most topics. Because it’s a computer, it creates equally-sized blocks of text for the reader. Humans prefer reading text that varies in length and sentence structure.
The LLM literally doesn’t know better; it’s looking for data patterns, identifying, and reporting. A human expert can look at AI output and know whether it’s accurate. But allowing the AI to do everything without checking the output produces bland, formulaic, sometimes inaccurate writing.
There is also that intangible flair that only comes from human writing. Writing from personal experience always produces more interesting, readable writing, using humor, varying sentence structure, changing tone, and digressions and asides (which computers don’t do; they stick to the subject at hand. See what I did there?). AI output is orderly, homogeneous, has flat tone and affect, eschews humor in favor of facts, and contains similarly-sized paragraphs and sentences.
Which paper would you give an “A” to? A paper does not have to be interesting to be factual and correct, but an “A” paper should clearly represent originality, critical thinking, and some novel statements or ideas on the student’s part. Compared to an AI’s computerized (pasteurized, if you will) output, AI writing still gets a “C” grade.
AI is Error-Prone
AI can help with generating code. But the prediction-based output can often result in extra unnecessary code that has to be stripped out. This phenomenon even has a name: workslop. It reverses the usual human-machine dynamic we are used to. Instead of a human offloading heavy cognitive tasks to the machine, the machine offloads the work of adding missing context, eliminating redundancies, and making appropriate connections to the human. And there are other issues that can cause an AI to go off the rails as well:
- AIs don’t feel the same constraints of ethical morality, so they will come up with non-realistic solutions to problems.
- They need to be given frequent guidance, and they must also be given solid, reputable information to start their research with. Garbage in, garbage out.
- In an attempt to help fill in experiential gaps, they sometimes get to the wrong conclusion, or “hallucinate.”
AI can make a video of something otherworldly. However, it still often leaves visual artifacts behind like extra limbs or fingers. Sometimes props will move on their own, weirdly floating about like they’re on wires. AI can produce content faster than we can in some cases, but it still cuts corners when it gets desperate, just like students do.
AI is Derivative
Critics say AI-generated writing lacks opinion, expression, passion. A human’s essay might go into more depth based on the author’s personal preference, where an AI won’t have one.
Even without considering attribution issues, AIs have historically not done well with the arts. Although the ability of AIs to mimic audio and video has gotten significantly better over the years (see the Will Smith eating spaghetti test for examples), there are still things it doesn’t do well.
An AI only has information as current as what its model has access to, and thus can only produce works that have characteristics already seen in art or music.
Take the example from 2023, when an AI-generated song in the style of Drake and The Weeknd nearly got a Grammy nomination. AI can derive, but cannot create. Humans still have the market cornered on that.
Can an AI write a hit pop song? Maybe, if by “hit,” you mean it gets a lot of listens and downloads. But rather than challenging the listener to absorb something new or feel something different, it would feed them something familiar and contrived. Would this AI-generated music even be remembered fifty years from now? History remembers the innovator, not the copycat.
Sure, an AI can make you a copy of Edward Hopper’s Nighthawks with Muppets in it. But it can’t make art in a style that no one has seen before on its own.
The C Student
We worry that AIs will produce content that we won’t be able to identify. It’s true, the process gets tougher to discern the more information the AIs get fed to them, and they do get better at predicting. But if we think about how AI currently performs as a student, it’s not really a good student:
- Given a prompt, an LLM will produce an essay with identical structure, every time.
- It will present evidence for each side in exactly equal measure, even down to the nearly identical word count.
- It overuses formal turns of phrase and may mismatch tone.
- It sometimes comes to the wrong conclusions, then may make up data to support its argument.
- It remixes class content, rather than using it as a springboard for original thinking and ideas.
- Its creative output is often crowded, messy and appears carelessly thrown together.
- Its artwork is derivative of others’ past works rather than creative and original.
- It sometimes misspells words, especially in slides and graphics.
Sounds like “C” work to me!
Many of these issues can be mediated by meta-prompting. But that’s the investment in time that students looking for quick fixes tend to skip. And this means that the reviewing step is less likely to happen, at least in the way we want it to.
This brings us back to our buddy Elias Thorne, whose sudden popularity is also a possible indicator of a serious problem that LLMs can face, called model collapse. The more low-quality AI-generated internet content there is, the more AIs may be trained on it. It’s like making a photocopy of a photocopy: each successive generation gets worse. Results become less specific and factual, instead becoming more vague and also more likely incorrect. Model collapse destroys the information’s reliability; you can’t use a word in its own definition.
Every day AI seems to do something new, and new fears arise about what it will take away. Judging by what we’ve seen, I don’t think the end is here yet. AIs can do a lot of things for us, as long as there is expert input and refinement to make sure the output is correct. We haven’t had the “John Henry” moment yet, where the machine truly beats man. We still have to be in the loop for that to happen.
It can’t outwrite a great writer, or give an impassioned speech like a great orator does. It can’t make music, or art, that looks or sounds like nothing we’ve seen before, because it lacks humanity. It’s still on its training wheels, because in order to do anything extraordinary, it needs someone with experience and expertise at the wheel.
Seems to me humans still have an abundance of those last two. Cheers!
