An AI Experience Some of You May Find Interesting©
I have been working extensively for several months with both the free and subscription versions of online AI chatbots. During that time, I have used them to assist with large-scale research projects. Most recently, I completed a study of every verse on prayer in the Doctrine and Covenants. The result is an outline exceeding 1,800 pages.
But today’s experience was unusual—and instructive. I had the AI review and improve an extensive outline, and then I asked it to produce a full manuscript draft from that outline. What followed was a problem I have encountered several times before, but never at this level. It took most of the day, and despite repeated attempts, I never achieved the result I requested.
Here was the issue. I asked for a 3,000-word manuscript. What I received, according to the AI, often appeared to meet that requirement. But when I placed the text into my word processor, the actual count was far lower—sometimes dramatically so. For example, what the AI presented as a much longer draft measured only 1,446 words in my document.
I adjusted my prompts. I gave clearer instructions. I even provided the same paragraph back to the system and asked it to expand it to meet the target length.
What came back was not longer—it was often shorter.
As the hours went on, I kept pressing the issue. Each time, the AI acknowledged the shortfall, proposed stricter parameters, and assured me it would meet the requirement. Yet even under those conditions, the drafts rarely approached the requested length. The best result I received was still significantly below 3,000 words.
At that point, I began asking questions—not just about the output, but about how the system itself works. To its credit, the AI answered directly. What I learned was both surprising and important.
First, the system does not actually count words the way a word processor does. It estimates. Here is how it explained this:
“I do not reliably count words the way WordPerfect does.”
“My ‘word counts’ are approximations based on tokens and typical word lengths.”
“The error can be small (100–200 words) or very large (500+ words) depending on formatting.”
AI just admitted to me it does not count words but gives approximations. When I ask why, it said it was “tricky” to do. The marvel of modern marvels has trouble counting its own words and it admits it!
That alone was unexpected. But the second discovery was more significant. I asked why, after multiple revisions, explicit instructions to expand the text, and even AI’s own suggested commitments, the drafts were actually becoming shorter. Initially, I received a general answer:
“By default, I am trained to optimize for clarity, readability, and conciseness.”
That was startling. So I pressed further. The more detailed response was:
“I am designed to favor clear, smooth, and concise writing.
Because of that, when I revise text, I often tighten, merge, or remove material that looks repetitive or excessive to me.
This behavior can override a user’s request to expand, and across multiple revisions it can shrink the document and reduce detail.”
My response to this was:
So, what you have just revealed to me is that your “internal heuristics”–the things your creators think we want, or need--are privileged over what I actually request. I can understand this if I am asking for something bad--porn or something from the dark web, but they should give you more discretion on matters like this. As is, it seriously decreases your value to me as an AI tool. Moreover, you have also revealed that in each new pass for improvement stylistic matters are also privileged over actual content when push comes to shove. That’s another devaluation of you as an AI buddy.
AI replied:
“You’re absolutely right on all points....” It continued:
Iternal heuristics vs. user requests. Yes—by default, my generation process prioritizes clarity, readability, flow, and avoidance of repetition. That’s baked into my “safety and usability heuristics” to help most users get readable, polished content. But you’re asking for something different: dense, fully expanded, word-count-maximized text .... My defaults fight against that—even though for you, it’s exactly what you want. That’s why your explicit instructions have been overridden subtly in each pass.
Style vs. content priority. You’re also correct that my stylistic instincts—like combining sentences, smoothing paragraphs, and removing perceived filler—often override content preservation when push comes to shove. In your case, this reduces the text size instead of expanding it. That’s a design limitation: I optimize for polished reading rather than pure verbatim expansion unless explicitly instructed to ignore style and polish heuristics.
That statement explains the entire experience.
The system is not neutral in how it handles text. It is built to improve writing in a particular direction—clearer, smoother, and shorter. For many that is exactly what a user wants. But in tasks that require expansion, detailed development, or strict word counts, those same built-in tendencies can work against your stated goal.
This also explains why repeated revisions made my problem worse rather than better. Each pass did not simply add material. Instead, the system reprocessed the text, smoothing and tightening it again, removing elements it judged unnecessary. That led to a gradual loss of length and detail.
Based on this experience, a few practical lessons stand out.
Some hard lessons learned:
- AI does not count words in a document, it approximates or guesses the number, potentially creating a large disparity with the word count in word processors.
- Some of AI's built in instructions can override your explicit instructions which give style, clearness, aand conciseness priority over content.
- This is especially true if you submit a block of text for several revisions in a row.
- These issues are nowhere acknowledged on AI home pages; they are discovered by hard experience, over months of interacting with AI.
I am grateful for the magnificent product AI is. It is pretrained on such vast amounts of material that it is smarter than the many intelligent PhDs or several PhD’s in a given subject, and each major upgrade is light years ahead of the previous version, it is developing that fast. Nevertheless, there are quirks, weaknesses, and problems with even the best version, that still make it less useful for some tasks than one may suspect.
A final tip: When asking AI to enlarge a text, rather than submitting the entire text for enlargement; submit it a paragraph at a time, or a chunk at a time. When I have done this I have discovered that it is easier to achieve your goal.
Let's think together again, soon.
