Something Is Off, and Everyone Knows It
Ask an AI to write something, read the result, and you will often feel it before you can articulate it: something is off. The words are correct. The sentences are grammatical. The structure is logical. And yet the text feels hollow — like a translation from a language nobody speaks.
This is the problem I kept running into, and I suspect it is familiar to anyone who has tried to integrate AI writing into their actual work.
It took me longer than I expected to understand what was causing it, and longer still to find reliable ways around it. This article is my attempt to save you some of that time.
Take AI-driven development all the way to production
WARP is a hands-on program for teams who want more than headlines. Former enterprise DX and data strategy leads work alongside you until it runs.
The Structural Problem
AI language models are trained to predict the next token in a sequence, given all the preceding context. They are extraordinarily good at this task. What they produce is, in a statistical sense, very plausible text — the kind of text that, on average, appears in the training data in response to similar prompts.
That is also the problem.
"On average" and "plausible" are not the same as "good." Good writing — writing that actually persuades, entertains, informs, or moves the reader — tends to deviate from the average in specific, purposeful ways. It takes risks. It uses rhythm to create emphasis. It includes the precise concrete detail that makes an abstract point land. It knows when to be shorter than expected and when to hold the reader in a longer sentence.
AI-generated text tends to be average, because that is what the training objective produces. It is rarely bad. It is rarely great. It sits in a comfortable, slightly-too-smooth middle ground that reads as professional but not alive.
The Specific Symptoms
Once you know what to look for, the patterns are easy to spot:
Rhythmic uniformity. Sentences in AI-generated text tend to be similar in length and structure. Human writing naturally varies. Some sentences are short. Others extend across multiple clauses, building toward a point, layering context, and releasing the reader at the end. AI text often lacks this variation.
Abstract generality. AI models default to general statements. "This approach has significant advantages for businesses." What businesses? What advantages? A human writer with any actual knowledge of the subject would be more specific. The vagueness is a tell.
Transition phrases. "It is important to note that..." "In conclusion..." "Furthermore..." These transitions appear constantly in AI text because they appear constantly in the training data — in textbooks, reports, and formal writing. They create a bureaucratic register that reads as stiff in most contexts.
Hedging. AI models are trained to be cautious about making strong claims. This manifests as constant qualification: "It is generally considered that..." "Many experts believe..." "While there is some debate..." Some hedging is appropriate. When every sentence hedges, the text has no backbone.
The completeness compulsion. AI models try to cover all the relevant bases, which means they often include information that the reader does not need. Human writers make choices about what to leave out. AI text includes everything that could be relevant, which dilutes the signal.
What Actually Helps
I have tried many approaches over the past year. Here is what has made a consistent difference:
Write with a specific reader, not a general audience. The most common prompt mistake is writing for an abstract audience. "Explain this to a non-technical reader" produces more useful results than "write a general introduction to this topic." Better still: describe a specific person with specific knowledge and specific questions.
Give the AI your actual perspective. If you have an opinion, state it in the prompt. "I think the consensus view on X is wrong because Y — help me write that argument." AI models can argue from a position you give them more effectively than they can generate a position from scratch.
Specify the constraint. Short is harder than long. If you ask for a 150-word explanation, the model is forced to make choices about what matters. The resulting text tends to be sharper than the same request without a length constraint.
Edit for rhythm deliberately. When reviewing AI-generated text, read it aloud. Where you stumble, the rhythm is wrong. Short sentences where there should be long ones, long sentences where there should be short ones — fix these specifically, not just the words.
Add one specific detail. Find one place in every piece of AI-generated text and add a concrete detail that only you would know — a specific number, a particular case, a phrase from an actual conversation. This anchors the text in reality in a way the AI cannot generate.
The Underlying Insight
The AI writing problem is not primarily a technology problem. The technology is good. The problem is that most people use AI writing tools as if they were a faster typist, rather than as a collaborator with a specific set of strengths and weaknesses.
AI is very good at structure, at comprehensiveness, at producing text that meets a specification. It is not good at voice, at restraint, at making the specific choices that distinguish good writing from merely correct writing.
The most effective approach treats AI as a first-draft tool and invests human judgment in the editing process — knowing specifically what to fix, rather than hoping the AI will get it right on the first pass.
That shift in how you use the tool makes a larger difference than almost any change to the tool itself.
