Detailed instructions and human writing samples cannot override how AI models are fundamentally trained to write, according to new research from TextPulse. The study tested whether prompting alone could make four leading AI models (GPT-5.6 Sol, Claude Opus 5, Gemini 3.7 Flash, and DeepSeek V4 Pro) produce text indistinguishable from human academic prose. The answer was no, even when researchers provided explicit stylistic guidance and human examples to imitate.

Can Prompting Override AI Training?

The research team gave each model three different prompts to write about 300 words on academic topics. The first was a simple request. The second included detailed instructions on properties that distinguish human writing from AI writing, such as uneven sentence length, plain vocabulary, and natural hedging. The third added an actual human-written passage as an example to follow. Despite these escalating attempts to guide the models toward human-like prose, the results remained stubbornly machine-like.

When analyzed on a stylometric spectrum designed to measure human versus AI characteristics, the models showed some improvement under detailed instructions, but the gains were modest. Only 14 of the 40 instruction-loaded texts fell within the middle range of human writing samples. More tellingly, when run through GPTZero, a widely used AI detection tool, all 120 AI-generated texts scored as 100 percent likely to be machine-written, regardless of which prompt was used. By contrast, all 10 human passages scored 0 percent.

Why Does Training Matter More Than Instructions?

The core finding challenges a widespread assumption among users: that AI models have a "machine register" as a default setting that instructions can override. The research suggests the opposite is true. The machine-like quality isn't a choice the model makes; it's baked into how the model was built.

Modern AI models go through multiple training stages. They start with pretraining on vast text corpora, which teaches them the underlying patterns of language. Then they undergo supervised fine-tuning on human-written demonstrations, followed by reinforcement learning from human feedback, commonly known as RLHF. This process shapes which parts of the language distribution the model can access. An instruction, the researchers explain, is merely a condition applied to that already-limited distribution, not a replacement for it.

The models in the study actually overshot the properties the prompts named. When asked to write with uneven sentence length, they produced text with even more variation than human academic journals. When instructed to use plain words, they simplified vocabulary more than necessary. Yet they left untouched the one property the prompt didn't mention: vocabulary range. This pattern reveals that prompting works within narrow bounds set by training, not beyond them.

What Do These Findings Mean for AI Detection?

If prompting cannot undo preference tuning, then AI detection tools that rely on surface-level stylometric features face a fundamental advantage. These tools detect patterns rooted in the underlying training process and distribution, not just surface-level writing quirks. This explains why GPTZero flagged every single AI text as machine-written, even those that passed stylometric tests.

The research team also tested a purpose-built humanizer tool on the same texts. A single pass through the TextPulse humanizer moved texts an average of 47 points toward the human end of the spectrum while preserving 97 percent of the meaning. Importantly, the rewritten texts passed GPTZero detection in cases where the original AI texts did not. This suggests that post-hoc rewriting, rather than prompting, may be necessary to fool detection systems.

How to Understand AI Model Limitations

• Training Determines Output: Supervised fine-tuning and RLHF heavily constrain which language patterns a model can produce, making instructions work only within those preset boundaries.
• Prompting Has Limits: Even detailed stylistic instructions and human examples cannot override the fundamental distribution a model was trained to sample from, meaning prompts cannot make AI writing truly indistinguishable from human writing.
• Detection Relies on Process: AI detectors that identify machine text based on underlying probability distributions are more robust than those relying solely on surface-level stylometric features, which prompting can partially manipulate.
• Rewriting May Be Necessary: Post-processing tools designed to rewrite AI text can achieve what prompting cannot, moving generated text closer to human characteristics while maintaining meaning.

The implications extend beyond academic writing. As AI models become more integrated into professional and creative work, understanding what prompting can and cannot do becomes crucial. Users who assume they can simply instruct a model to "write like a human" may be operating under a misconception. The machine register is not a default the model chooses; it is a property the model carries from training.

All data, code, and analysis from the study are publicly available, allowing other researchers to verify the findings and build on them. The work adds to a growing body of evidence that surface-level stylistic differences between human and AI text reflect deeper, training-level distinctions that cannot be overcome through prompting alone.