← Back to Writing

Why I don't use AI to write

Your mental skills and work product suffer when AI writes for you.

FROM: Ben Lovell DATE: February 2, 2026 RE: The real problem with AI writing.

QUESTION PRESENTED:

Does work suffer from using AI to write for me and is that a big deal?

SHORT ANSWER:

Yes. AI writing leads to converging suggestions and arguments. It puts the killing in “deskilling.”

DISCUSSION

“Walker, there is no path; the path is made by walking.” [Traveler, your footprints] Antonio Machado

The internet abounds with blog posts espousing the power of Large Language Models like ChatGPT as creative collaborators. This is one of them; however, I argue that the power LLMs wield causes harm. What happens when professionals collaborate with AI? Do we get better ideas? Do we converge on the same basic ideas and output? Do we lose our edge?

Artificial Hivemind

LLMs converge on the same words and ideas over time. For example, Liwei Jiang et al., studied what they call the “Artificial Hivemind” effect. The group posed open-ended questions to a large subset of popular LLMs and recorded the results. The responses showed that different models (ChatGPT, Gemini, Claude) tend to converge on nearly identical phrasing, structure, and ideas when answering the same prompts.1

Jiang et al. discussed the implications of the artificial hivemind effect on skilled professions. They argue that this convergence poses a risk of cognitive homogenization. We are all guided to the same narrow set of “safe” answers to our questions and our language and phrasing converges.

Research indicates that offloading too much of your cognition to AI models leads to “de-skilling.” Law is a skilled profession; As an attorney, I am paid because of my skills. Offloading my skilled work to a company robs me of experience. I am doing nothing more than paying for the privilege of training my purported replacement.

Legal arguments may not seem like a time for creative writing, but creativity is required for novel arguments and interpretations. Humans require nuance to persuade. If we attorneys are all converging on the same analyses and argument because we are allowing LLMs to unduly influence our thoughts, what is the point? What, really, is the point of working as an advocate if we are not going to advocate?

AI deskilling is a known problem in a legal field. Jones and Walker LLP wrote about what they cal the “four phases of skill loss:”2

  • Phase 1 (Enhancement): AI assists with routine tasks, improving efficiency while professionals maintain a full understanding of underlying processes.
  • Phase 2 (Integration): AI handles increasingly complex tasks as professionals become comfortable with algorithmic assistance and begin relying on it for standard workflows.
  • Phase 3 (Dependency): Professionals struggle to perform tasks without AI assistance, losing familiarity with manual processes and independent analysis capabilities.
  • Phase 4 (Atrophy): The skills necessary for independent practice deteriorate, rendering professionals unable to effectively verify AI outputs or operate when systems fail.

As a person who tries to understand LLMs and their implications, I use them regularly. The models do a good job with menial tasks. I often feel the pull toward dependency. Claude performs basic searches and outlines legal requirements with ease. It is possible to generate a lot of work product in a lot less time than before. I face myriad tiny research projects throughout the day and the desire to offload those tasks is strong. But I have to draw the line somewhere.

Fortunately, I find AI writing bland and in need of extensive correction. AI writing frequently disappoints and dissuades me from using it. I’m not alone in this opinion. Most people find AI writing bland. We collectively joke about bad writing “sounding like AI.”

Selling Skills to Others

Where do our skills go once offloaded? Corporations. Attorneys practice for years on end to develop skill, experience, and judgment. Legal Tech and AI companies offer attorneys convenience and productivity at the cost of skill and experience. We pay them to take our work and use it to improve their models. Our skills will be repackaged and sold back to us at an ever increasing price.

Reverse Centaurs and Liability

Cory Doctorow writes about a coming group of people he characterizes as “reverse centaurs”3 Doctorow explains the concept:

Start with what a reverse centaur is. In automation theory, a “centaur” is a person who is assisted by a machine. You’re a human head being carried around on a tireless robot body. Driving a car makes you a centaur, and so does using autocomplete.

And obviously, a reverse centaur is machine head on a human body, a person who is serving as a squishy meat appendage for an uncaring machine.

These scenarios create liability issues. A radiologist verifies machine-analyzed scans presented at an inhuman rate and misses a tumor. A patients die from preventable issues. Who is liable? The human doctor who offloaded their skill but not their liability.

Deskilling and Replacement Are Not Inevitable

AI replacing humanity is not inevitable. Media discourse surrounding the AI takeover and AI-invested companies present this replacement as inevitable. The end result is a environment of despair and hopelessness. I attended a CLE recently wherein the presenter opined to the audience that a technological takeover of our profession is inevitable. That we should stop moaning and get on board. CEOs tell us the same. It convinces us that we cannot stop capitalism from running us off a cliff. The opposite is true—the future is uncertain therefore anything is possible.

AI (LLMs) Cannot Forge a New Path (Yet)

LLMs seemingly cannot generate novel ideas in human domains like art. I was a musician before I was a lawyer so I like to use music to make a point. Generative AI remarkably productive, but can it “create”? Given ragtime and delta blues music as a model, can AI create jazz? Given only the pop music of the 70s as a model, can AI create hip-hop? The current state-of-the-art as we understand it cannot. That may change in the future.

Advocacy requires something different than pattern recognition. Current iterations of Large Language Models make their path not by walking, but by retreading all previous paths. Human persuasion requires something more than a synthesis of what has happened—It requires someone to tell a story of the future, most often a story of hope. For now, that is us.

Keep writing.

lulu-climbing-web

Footnotes

  1. Liwei Jiang et al., Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond), ArXiv (Oct. 27, 2025), https://arxiv.org/abs/2510.22954. ↩

  2. From Enhancement to Dependency: What the Epidemic of AI Failures in Law Means for Professionals, Jones Walker: AI Law Blog (Jan. 15, 2025), https://www.joneswalker.com/en/insights/blogs/ai-law-blog/from-enhancement-to-dependency-what-the-epidemic-of-ai-failures-in-law-means-for.html. ↩

  3. https://doctorow.medium.com/https-pluralistic-net-2025-12-05-pop-that-bubble-u-washington-8b6b75abc28e ↩