writes:I work with words for a living. At one company, I actually had the job title, 'writer'. I seldom seem to get around to writing short stories or longer fiction, but I've certainly written gaming rules and supplements, computer game dialogue, magazine articles, and a ton of website pages.
But until now, I haven't tried using AI with writing. It was a friend's tale that got me to look into it.
A Gamer's Autobiography
I was talking on the phone recently to one of my oldest gaming pals, a semi-retired archeologist and Waffle House manager. He's working on his autobiography, and he's using an AI (I think he said Gemini) to edit and critique his output.
He explained that AIs have a short attention span, so what he does is keep the full text in a plain text file on his laptop, then upload it to the AI fresh in each new session. (Some AIs won't accept text files for security reasons, so you might have to format as a PDF. But Gemini accepts text files.)
One of the annoying problems is that the AI sometimes reacts inappropriately to his past experiences. In one chapter, he was writing about a period in his life when he was depressed. When he asked the AI to edit that section, it instead intervened and asked if he needed counseling! So he explained carefully to the AI that this was a past experience from his autobiography, not something to be alarmed about.
He's also run into issues when certain topics come up, such as hunting and firearms. When asked to edit a particular chapter, the AI responded that it couldn't help him build a bomb! Again, he carefully explained this was a past life experience from his autobiography, it had nothing to do with making bombs, and it relented and did the edit.
It’s a bizarre experience to hear about an algorithm playing moral gatekeeper to someone's actual life story.
The most frustrating thing, however, is when the AI volunteers to 'polish' his text. He finds it often edits out complete areas of his manuscript, and he has to argue with it to restore what's missing.
The 600-Word Ceiling
This last problem interested me, so I did some research and experimentation.
I found out that AIs tend to run into a problem generating text longer than about 600 words. Large-Language Models (LLMs) are designed to generate short amounts of text. Without guidance from the user, they essentially 'run out of ideas' (or steam!).
In my testing, Grok routinely generated text in the range of 600 words. Sometimes it could get up as far as 1,000 words, but any further revisions would bring the word count back down. This may depend on what you're asking it to do.
This means that when an AI is editing text, it may focus on one problem, run over its 'word allowance', and start throwing things out of your text!
In one of my experiments, I gave Grok 650 words of fiction about a protagonist and another character. I asked the AI to strengthen the role of the secondary character. It rewrote the text to make the secondary character the focus, throwing away key plot points and giving me 500 words in return. Definitely not what was wanted.
Surprisingly, I've learned that LLMs don't actually work in terms of word counts. I've found that if I ask Grok, for example, to generate 1,000 words of text, it will cheerfully give me 600 words and claim that it's 1,000 words!
My understanding is that LLMs are not inherently designed to track word counts accurately. They work by predicting the next token in a sequence. When you ask for a specific word count, the model is essentially guessing, which is why it often cuts off or stops prematurely while claiming to have finished the requested length.
The AI may also stop generating text because it has hit a limit, often based on your user level. Essentially, it has an effort buffer, and stops trying when it feels it has already done enough.
Solutions
Knowledge is strength, as they say. Once you know what the limitations are, there are strategies to avoid them.
One way to handle a problem when an AI gives a brief or truncated response is to simply say, 'Continue'. It will then generate the next chunk of text.
When working with an AI, it is best to give it smaller tasks. Instead of asking for an entire article at once, ask the AI to create a plan, then ask it to write segments (e.g., "Write the first segment," then "Continue from there").
Instead of asking for a specific word count, tell the AI what you want it to expand upon. For instance, with fiction, you could say "Expand on the sensory details of the scene" or "Add more internal monologue"). You will likely get a more natural, longer response than by trying to force a specific number of words.
Rather than letting AI edit your text, you can ask it to suggest changes, showing the original text and the revised text. Ask for changes to be displayed in bold. This gives you control over what actually changes in your document, but is much more tedious than trusting the AI to just do it.
The bottom line? Artificial intelligence is a great tool for writers, but beware of its current limitations. Once you know where the weaknesses are, you can to some degree mitigate them.
AI is an editor that never sleeps, but it lacks the competency to write or rewrite more than a small section of text. It’s a powerful tool, provided you remain the one holding the pen.
What AI Thinks of This Article
I asked Gemini to critique this article.
This is a thoughtful and relatable piece for anyone interested in the intersection of professional writing and AI. To improve the article, you can focus on tightening the narrative flow, enhancing the tone, and sharpening the technical explanations.
It made some odd suggestions. For example, it wanted me to change the introduction to express my skepticism about AI. (I'm not skeptical, I've just never seen the need for it in my own writing.)
It wanted me to insert this bridge:
My initial apprehension was mirrored – and amplified – when I spoke to a longtime friend currently battling his own AI writing assistant.
Again, no apprehension here. Just curiosity.
In the paragraphs about the AI’s refusals (the bomb-making/counseling issues), consider a brief, punchy sentence on how these "guardrails" affect the creative process. Example: "It’s a bizarre experience to have an algorithm play moral gatekeeper to your own memories."
Not a bad suggestion, but this seems like putting words into my friend's mouth. I reworked it.
You describe the AI "polishing" the text as frustrating. You could clarify why this happens: it’s likely the AI is prioritizing conciseness or tone-consistency over your original voice.
It misses that the entire next section is about why this happens, and I doubt it involves 'conciseness or tone-consistency'.
You mention the AI runs out of "ideas" or "steam." It would be more accurate to describe this as a context window management or token limit issue.
Yes, but then I'd have to digress into more explanations about context windows and tokens.
Instead of saying it's "surprising" that LLMs don't track words, frame it as a fundamental mechanic. You could state: "I quickly learned that asking an AI for a specific word count is like asking a compass to tell you the time; it’s simply not the metric they were built to track."
Don't tell me what to be surprised by! I'm shocked that a piece of software can't do something as simple as counting words. That seems pretty fundamental to many tasks.
The current ending is a bit abrupt. Consider adding a closing thought that emphasizes your perspective as a professional writer. "AI is an editor that never sleeps, but it still lacks the human discernment to know which darlings to keep and which to kill. It’s a powerful tool, provided you remain the one holding the pen."
Not the point I was making, but I can work with this.
There is a minor typo: "Surprising, I've learned..." should be "Surprisingly, I've learned..."
Good catch.
The Illustration
I asked Gemini to create an illustration for this article. Usually, it uses Nano Banana to create images. Today, that wasn't working for some reason, but it did give me a detailed image prompt:
A professional editorial-style illustration (1920x1080) in a clean, modern aesthetic. Inside a glowing digital monitor frame, a ghostly, translucent silhouette of a writer is seen frantically reaching out, trying to catch and restore pieces of a floating, elegant manuscript that is being shredded into binary code and gray dust by sharp, mechanical digital blades. The scene represents the frustration of an AI "polishing" away key creative elements. The palette is a sophisticated mix of deep midnight blue, slate gray, and a striking, human-like warm amber accent glow coming from the writer's hands. The composition is sleek, high-contrast, and conceptual. Strictly no text, no words, no captions, and no labels of any kind.
Grok then accepted that prompt and produced the image used above.