New technique pinpoints exactly where humans and AI take over in co-written text

Most text detection tools work with a simple yes-or-no answer: is this passage human or machine-generated? But in an increasingly collaborative world where writers work alongside language models, that binary approach falls short. A new research effort tackles the tougher problem of identifying exactly which sentences or paragraphs came from humans and which came from AI.

The Core Problem#

The challenge lies in documents that blend both voices seamlessly. Traditional detectors treat entire passages as single units and classify them wholesale, leaving users unable to distinguish individual contributions when both humans and LLMs have touched the same piece of writing. This limitation becomes critical as hybrid authorship becomes the norm in professional, academic, and creative contexts.

A Time-Series Solution#

The key insight behind the new approach draws inspiration from change point detection, a classical technique used in analyzing sequential data for sudden shifts or transitions. The researchers realized that segmenting human-versus-AI text faces the same conceptual challenge: finding the precise moments where one author's voice gives way to another's.

Using this framework, the team developed multiple algorithms tailored to the problem. Their methods account for varying reliability levels across different detection signals and maintain theoretical guarantees about performance optimality.

What This Means#

Beyond pure detection accuracy, these tools enable fine-grained auditing of collaborative writing. They help verify authenticity at the sentence level, support better understanding of how humans and models interact in the writing process, and strengthen trust in documents where both parties have contributed.

The researchers released Python code alongside their findings, making the approach available to the wider research community.