We’ve all been bombarded with the latest buzzword everywhere we turn. Many promises have been made. Our eyes and ears are flooded with sensational claims that it’ll save us, end us, or anything in between. We hear:
AI will make us more productive.
AI will make us faster.
AI will enable anyone to create that which they could not before.
I could go on.
“Artificial Intelligence” is everywhere. AI hype may have reached a new high, but AI is far from new. Generative AI can be fun and even useful. There are ways it can help us. However, there are some things that it’s fundamentally incapable of.
A Quick Allegory on AI
I remember a personal story as an allegory for my experience with AI. Equipped with only shoes and a cheap bicycle, I worked three miles from home. Despite the promise of a more efficient commute, I rarely used my bike. Past experience with that bike fostered my distrust. This was many years ago now, yet the memory survives.
I was speaking with the teacher of my dual-credit writing class when a classmate entered the room. Concerned, he mentioned that my bike tire was flat. Upon finishing class, I left for my internship. There was a nearby gas station where I offered to pay for them to turn on the air compressor, which the clerk kindly did for free. However, within yards, my tire was again flat, and I was again on foot, but with the nuisance of a bike I couldn’t ride. I’d struggled with seemingly unfixable brake issues prior. Plans collided with a void that should’ve been filled by reliability. The promise of an efficient commute was effectively broken. My trust in that bike never recovered.
Bicycles have a compelling purpose. Mine failed to fulfill that purpose enough to stay relevant. I learned that time-saving tools must be reliable if they’re to be relied on. With insufficient trust, there’s a need to allot the same time as without the tool. I’d save no time yet accept responsibility for an untrustworthy tool.
With insufficient trust, there’s a need to allot the same time as without the tool.
Working in IT, I’m extensively exposed to AI in both concept and practice. I’ve felt pressured to forget that artificial intelligence is artificial. My work with it has never allowed me to forget that I’m working with a computer. Even the applicability of the word “intelligence” is disputed, but rightly so.
Machines incapable of living a human experience will forever be distinguishable from a human. Computers don’t have serotonin, dopamine, or cortisol. At the time of this writing, no computer has ever felt the joy of a hard-earned achievement or the grief of losing a friend. This is how “artificial” applies to AI. Miles of content can, have, and will be created confronting the ethics and philosophy of AI, but I’ll set that aside and focus on the applicability of a subset: Large Language Models.
How Large Language Models Really Work
LLMs are being tacked on to so many products—often for fear of missing out, obligation to shareholders, or other questionable reasons. You’ve likely used one by now, knowingly or not. ChatGPT, Claude, and Copilot are examples of LLMs. LLMs produce text, typically in response to text they receive. They’re being used in “support” chat, for research, for creating scam content, to generate disposable corporate content, and for many other purposes.
So, how do they work and what can they actually do? They aren’t magic. They are software programs designed to examine text and produce different text using the same patterns. They choose the next word(s) according to their analysis of text that they’ve consumed. Context clues are being added, but fundamentally, LLMs generate text with probability derived from patterns they’ve analyzed before. LLMs are tools for recreating patterns—often to copy information.
LLMs are tools for recreating patterns—often to copy information.
I work with LLMs nearly every day, and I’m confronted with their limits just as often. I use them to supplement my software coding work. Computer code is notoriously best when it’s boring, predictable, and unoriginal. “Clear is better than clever.”
LLMs still disappoint. I’ve repeated the following cycle enough times to be humbled that I still do.
- I ask for a thing (e.g., code, code review, or information).
- I review the response and see something I don’t understand but see potential.
- I ask it to help fix, complete, or explain a broken concept.
- It gives an absurd explanation, adding useless or harmful bloat.
- It apologizes, says it understands what went wrong, and gives the “final” and “correct” version.
- The cycle continues until I eventually realize that it’s incapable of helping with the task and I could’ve completed or made significant progress had I done it myself.
That’s how I’ve seen LLMs handle work that’s more factual than emotional. Communication meant solely for humans is different. It requires empathy developed from a human experience. Visual, textual, and auditory AI products often imitate various forms of communication, but results land hollow.
The True Cost of Heavy AI and LLM Usage
The greatest cost to heavy AI usage is fundamentally human. Multiple studies have shown that we are lazy thinkers: when we can delegate decision-making or other mental work elsewhere, we not only do, but also limit and lose our proficiency in those areas.1, 2 We debit the human experience with no plan for how to repay it. We can change and develop technologies, but we can’t change our nature.
We debit the human experience with no plan for how to repay it.
Studies have shown “cognitive offloading” in several areas. The use of GPS devices has been shown to reduce our navigational abilities.3 The results of a (yet to be peer-reviewed) study by MIT suggest that AI usage can reduce critical thinking and one’s ability to self-evaluate their own quality of work.4 My own observations lead me to expect the same.
Using AI for what should involve creativity, be it writing or image creation—even if the result is indistinguishable from human work—fails to communicate and express through connection to a human experience. To some, this is the purpose of such a work. Human communication contains vastly more information that can be derived from lexical structure alone.
Even people who struggle with social cues interpret countless micro-expressions taken for granted. Computers can’t produce or observe these cues as we can. We struggle to encode these cues for a computer, and computers don’t have the hardware to create or observe them. The best AI can do is echo a message that’s already been spoken. LLMs can be helpful, but when their output is used, it’s often because:
- A human already created it, but with empathy and context of the human experience.
- It probably didn’t need to be made. Examples:
- A summary of a report for a boss who won’t read it.
- An unempathetic, unhelpful automated email to a (now former) dissatisfied customer.
- A chatbot providing partial information from a help article that has more comprehensive information which should’ve been easily searchable.
Where LLMs Shine in Daily Work
If I didn’t find LLMs to be useful, I wouldn’t use them. They sometimes help with formulaic parts of my work, leaving me energy for the more creative parts. I still find them wasting my time. I still find problems they produced days or weeks later and realize that I’d gotten too lazy or hasty. I don’t use them to reduce ambiguity; LLMs confuse the best of plans. I only use them when I’m informed enough to falsify their output and have time to review it fully.
AI cannot know when it is wrong or test its output against the world. AI doesn’t understand meaning as a human would. Current LLMs usually produce output with an air of confidence detached from accuracy. Even a deceptive person is likely to give cues that their claims are inaccurate.
Current LLMs usually produce output with an air of confidence detached from accuracy.
I don’t use LLMs where empathy is required. I don’t use them for things that require a human experience, which happens to be most things. For such efforts, I prefer to maintain sincerity and quality of work through my humanity. I believe that the adoption of generative AI is unstoppable, but we have the power to discern when we’re better off not using it.
When AI is used to produce something, it often appears cheap and detached from humanity. Every time you ask, “Can AI do this?” you owe it to yourself to ask, “Would it be more effective if a human were to do it?” When you use LLMs to help with something you care about, you may find yourself spending more time trying to make it work than you would have on the project without it.
Every time you ask, “Can AI do this?” you owe it to yourself to ask, “Would it be more effective if a human were to do it?”
Promises have been made about AI. Just as with my bicycle, promises have been broken. LLMs require expertise and attention to detail for verification. As such, they cannot be trusted to save time or even not to waste it.
A Viable, Reliable Alternative
Nine Muses Writing Group writes content for humans, by humans. If you need to delegate writing tasks, we provide tailored content informed by empathy that can only come from the human experience. Contact us today for a free consultation.
Sources:
- Skulmowski, A. (2023). “The Cognitive Architecture of Digital Externalization.” Educational Psychology Review, 35, Article 101. https://doi.org/10.1007/s10648-023-09818-1.
- Firth, J., et al. (2019). The online brain: How the Internet may be changing our cognition.” World Psychiatry, 18(2), 119–129. https://doi.org/10.1002/wps.20617. Discusses cognitive offloading, memory externalization, and potential long-term cognitive effects of digital technologies.
- Dahmani, L., Bohbot, V., et al. (2020). “Habitual use of GPS negatively impacts spatial memory during self-guided navigation.” Scientific Reports, 10, 6310. https://doi.org/10.1038/s41598-020-62877-0.
- Kosmyna, N., et al. (2025). “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” arXiv:2506.08872. https://doi.org/10.48550/arXiv.2506.08872. Pre-print paper: not peer-reviewed.
