How much does the prompt change the feel?
The prompt is the single most controllable factor. A 2026 study tested two versions of an AI chatbot—one prompted to be 'empathic' and one 'compassionate'—and found that the empathic version elicited significantly stronger emotions (empathy, compassion, distress) in 122 participants [2]. The only difference was the prompt; the underlying model and training data were identical. This shows that even subtle prompt wording can dramatically shift the perceived personality and emotional tone of the output.
Further evidence comes from text style transfer research. A 2024 study showed that using the same prompt for every input is suboptimal—different inputs respond better to different prompts [4]. They built a system that adaptively selects the best prompt for each input, achieving superior style transfer compared to using a fixed prompt. This confirms that prompt choice is not just a minor tweak; it's a core determinant of output quality and feel.
Does the training data leave a permanent mark?
Yes, and it's detectable. A 2025 study analyzed text from five major AI models (ChatGPT, Claude, Gemini, ERNIE Bot, Grok) and found that each model has a unique 'linguistic style fingerprint' based on features like word choice, sentence structure, and punctuation habits [5]. Using these fingerprints, they could identify which model generated a given piece of text with high accuracy. This means the training data—the corpus each model was trained on—imprints a baseline style that persists regardless of the prompt.
However, the prompt can override some of these fingerprints. The same study noted that their method worked across ten different semantic tasks, suggesting that while the model's 'voice' is consistent, it's not immutable. The prompt can steer the output toward or away from that baseline, but the underlying statistical patterns from training data remain detectable.
What about the model itself?
The model architecture sets the ceiling for quality, but within the same model class, prompt and data matter more. A 2023 study on text style transfer found that a prompt-based editing approach outperformed systems with 20 times more parameters [3]. This suggests that clever prompting can compensate for a smaller or less capable model. The model's architecture determines what's possible (e.g., how well it can follow complex instructions), but the prompt determines what actually happens.
That said, different models do have different inherent feels. The 2025 fingerprinting study [5] showed that models like ChatGPT and Claude produce measurably different text even when given the same prompt. So the model matters, but it's the foundation—the prompt and training data are the paint and brush that create the final picture.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2023 to 2026, 4 from 2024 or later, 1 in Q1–Q2 journals — selected as the most relevant from 5 studies that passed quality screening, drawn from 61 papers retrieved from a database of over 500 million.
Sources used in this answer
Learning to detect AI texts and learning the limits
In a study with 254 Czech speakers, participants who received immediate feedback improved their ability to distinguish AI from human text, correcting initial misconceptions about AI stylistic features and readability [1].
Chatting with an LLM-based AI elicits affective and cognitive processes in education for sustainable development.
In a study with 122 participants, an empathic AI chatbot elicited stronger emotions (empathy, compassion, distress) than a compassionate one, showing that prompt personality directly changes reader experience [2].
Prompt-Based Editing for Text Style Transfer
A prompt-based editing approach for text style transfer outperformed systems with 20 times more parameters, proving that clever prompting can compensate for model size [3].
Adaptive Prompt Routing for Arbitrary Text Style Transfer with Pre-trained Language Models
An adaptive prompt routing framework selected different prompts for different inputs, achieving superior style transfer compared to using a fixed prompt for all inputs [4].
Identifying AI-Generated Text Sources via Linguistic Style Fingerprints
Using linguistic style fingerprints (word choice, syntax, punctuation), a classifier could identify which of five AI models (ChatGPT, Claude, Gemini, ERNIE Bot, Grok) generated a given text [5].