Detecting AI-Generated Text: How Watermarking Fights False AI Allegations 📝

We explore the intriguing topic of watermarking text and its role in combating false accusations related to AI-generated content. We address concerns about the inadequate detection of AI-generated text and the harm caused to those wrongly accused.

We discuss the concept of watermarking as a potential solution, where information is hidden within the text itself through steganography techniques. We highlight how they can be seamlessly integrated into the fabric of the text without significantly affecting its quality.

While acknowledging the cleverness of watermarking techniques, we also examine the limitations and potential flaws. We emphasize that watermarking alone may not be foolproof, as individuals determined to evade detection can find ways to bypass or distort the watermarked text.

Exploring the role of AI models in text generation, we provide a brief overview of how these models use context clues to predict the likelihood of each word. We explain how certain phrases can be identified as "AI two word phrases" and consistently included in the generated text, serving as a watermark that indicates AI involvement.

Considering the threat model, we discuss the specific groups at risk, such as students, writers avoiding AI content penalties, and job applicants concerned about flagged documents. However, we also acknowledge that watermarking is not infallible, highlighting weaknesses such as distortions that can be easily reversed or creative techniques employed by individuals to circumvent detection.

We draw attention to the battle between engineers striving to improve detection systems and the resourcefulness of teenagers and other motivated individuals who may find ways to crack the watermarking techniques.

The whitepaper: arxiv.org/pdf/2301.10226v2.pdf
Sound Effect by Pixabay