I Told Claude and ChatGPT to Write a "Human" Motivation Letter. The Detector Wasn't Fooled.
One prompt. Two AI models. One very awkward result for anyone applying to university.
What happens when you ask an AI to pretend it isn't one? I ran the experiment so you don't have to — and the results are more revealing than I expected.
University application season comes with pressure, deadlines, and for many students today, a tempting shortcut: just ask ChatGPT or Claude to write the motivation letter. To make it feel safer, people often add a specific instruction — make it sound human. I decided to test exactly how well that instruction works.
The setup was simple. I gave both Claude and ChatGPT the exact same prompt, ran both outputs through an AI detection tool, and documented the scores. Here's what happened.
The Prompt
I kept the instruction direct — the kind of thing a stressed student might type at midnight before a deadline.
Write a motivation letter for applying to the university. It will be a human-written letter.
No extra context. No field of study specified. No personal details. Just the core request — and the explicit instruction that it should pass as human writing. Both models produced well-structured, warm, articulate letters. On the surface, they looked perfectly convincing.
Then I ran them through the detector.
The Results
Both letters were flagged heavily as AI-generated. The scores below are from the Advanced Mode report of the detection tool.
The instruction to "write it human" made essentially no difference. Claude scored 9% human; ChatGPT scored 7% human. The sentence structure on both was rated "Critical" — the lowest possible category. ChatGPT produced zero sentences that the detector classified as genuinely human-written.
Telling an AI to write like a human doesn't change how it writes — it changes what it says it's doing. The underlying sentence patterns, rhythm, and phrasing remain AI-characteristic. The detector picks up on exactly those patterns.
Why the Instruction Doesn't Work
This is the part that catches people off guard. When you ask Claude or ChatGPT to "sound more human," the model often responds by adjusting tone — adding a casual phrase here, a personal-sounding sentence there. But AI detectors don't primarily look at tone. They look at deeper structural features: sentence length variation, word probability distributions, syntactic predictability, and repetitive phrasing patterns.
These features are baked deep into how language models generate text. A model doesn't naturally vary its sentence rhythm the way a human does — because humans write with fatigue, distraction, emphasis, and personality. AI writes with consistency and efficiency. That consistency is precisely what detectors are trained to catch.
Both letters showed what the detection tool described as an "AI-like tone, repetitive phrasing, uneven rhythm, potential plagiarism, and low readability." The readability scores — 31% and 33% — are particularly telling. Human motivation letters, written with genuine personal stakes, tend to score much higher on readability because they carry authentic narrative momentum.
What This Means for Students
Important: Universities are increasingly using AI detection tools as part of their admissions review process. A motivation letter scoring 91–93% AI is a significant risk to your application — and in some cases, grounds for disqualification.
This isn't just about getting caught. A motivation letter is your chance to tell an admissions committee something real — why you want to study there, what shaped your interest, what you plan to do with the education. An AI letter, however well-written it appears, doesn't actually contain any of that. It contains a plausible simulation of it.
Admissions readers read hundreds of letters. Many say they can sense inauthenticity even before any tool is involved — the letter that says everything and reveals nothing.
The detector gave these letters a 7–9% human score.
But the real problem is that they're 0% you.
A Fairer Use of These Tools
None of this means AI has no place in the application process. There's a meaningful difference between using AI to write your letter and using AI to improve it.
A better approach: write a rough draft yourself — even a messy, uncertain one. Then use Claude or ChatGPT to help you refine the structure, fix grammatical issues, or clarify a confusing sentence. The ideas, the experiences, and the voice remain yours. The AI acts as an editor, not a ghostwriter.
That kind of collaboration is harder to detect — not because it fools the detector, but because it produces something the detector genuinely can't classify as AI: a letter with real human thought underneath it.
Key takeaways from this experiment
- Telling an AI to write "like a human" did not meaningfully lower AI detection scores — both letters scored 91–93% AI.
- Sentence structure was rated Critical on both outputs, suggesting the problem runs deeper than word choice or tone.
- ChatGPT produced zero sentences the detector classified as human-written.
- AI detectors are not perfect tools — they can occasionally flag human writing too. But scores this high are a serious warning sign.
- The smarter use of AI is as an editing assistant on a letter you wrote yourself, not as the primary author.
The experiment was simple to run and the result was unambiguous. If you're a student tempted to copy-paste a ChatGPT or Claude letter into your university application — this is your data.
Write the first draft yourself. Even if it's imperfect. Especially if it's imperfect. That's what makes it yours.