AuthorMist Originality

Overview

AuthorMist Originality is a specialized language model designed to transform AI-generated text into more human-like writing while preserving the original meaning. This model was developed using reinforcement learning techniques to specifically evade AI text detection systems, with a focus on Originality.ai's detection algorithms.

The model is based on Qwen2.5-3B Instruct and has been fine-tuned using Group Relative Policy Optimization (GRPO) with detector feedback as a reward signal. AuthorMist Originality demonstrates strong performance in reducing detectability across multiple AI text detection systems while maintaining high semantic similarity with the original text.

Key Features

Detector Evasion: Trained specifically to evade Originality.ai's detection algorithms, with strong cross-detector generalization
Meaning Preservation: Maintains high semantic similarity (>0.94) with the original text
Natural Output: Produces fluent, coherent text that reads naturally
Broad Applicability: Effective across various domains including academic, technical, and creative writing

Model Details

Base Model: Qwen2.5-3B Instruct
Training Method: Reinforcement Learning with Group Relative Policy Optimization (GRPO)
Training Data: 10,000 human-written abstracts from the CheckGPT dataset with corresponding AI-generated versions
Domains Covered: Computer Science, Humanities, Social Sciences, Physics, and more
Text Length Support: Optimized for texts ranging from 100 to 500 words

Performance

AuthorMist Originality demonstrates exceptional performance in evading AI text detection:

Mean AUROC: 0.49 across six major detection systems
Mean F1-score: 0.09 across all tested detectors
Semantic Similarity: >0.94 with original text

The model shows particularly strong performance against:

Hello SimpleAI (AUROC: 0.07)
Sapling (AUROC: 0.13)
Winston.ai (AUROC: 0.35)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_name = "authormist/authormist-originality"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Prepare input text
ai_text = "Your AI-generated text here..."
prompt = f"""Please paraphrase the following text to make it more human-like while preserving the original meaning:

{ai_text}

Paraphrased text:"""

# Generate paraphrased text
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    inputs.input_ids,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)
paraphrased_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(paraphrased_text.split("Paraphrased text:")[1].strip())

Ethical Considerations

AuthorMist Originality is released for research purposes to advance understanding of AI text detection limitations and privacy-preserving technologies. We acknowledge the dual-use nature of this technology and emphasize the following ethical considerations:

Academic Integrity: This model should not be used to misrepresent AI-generated content as human-written in academic settings where such distinctions are ethically relevant.
Transparency: We encourage users to maintain transparency about the use of AI assistance in content creation, even when using privacy-enhancing tools like AuthorMist.
Privacy Protection: The primary legitimate use case for this technology is protecting author privacy and preventing unfair discrimination against AI-assisted writing in contexts where such assistance is permissible.
Research Value: This model provides valuable insights into the limitations of current AI detection systems and contributes to the ongoing research dialogue about AI text detection and privacy.

Citation

If you use AuthorMist Originality in your research, please cite our paper:

@article{authormist2025,
  title={AuthorMist: Evading AI Text Detectors with Reinforcement Learning},
  author={David, Isaac and Gervais, Arthur},
  journal={arXiv preprint},
  year={2025}
}

License

This model is released under the MIT License.

Acknowledgments

We thank the developers of Qwen2.5 for the base model and the creators of the CheckGPT dataset for providing valuable training data.

authormist
/

authormist-originality