A new detection system can identify sentences written by artificial intelligence mixed into documents co-authored by humans and machines, achieving significantly higher accuracy than existing methods by analyzing how sentences relate to one another. The breakthrough, called SenFlow, addresses a growing real-world problem: students polishing essays with AI help, researchers inserting machine-generated summaries, and legal analysts drafting clauses with language models. Unlike older detection tools that examine each sentence in isolation, SenFlow treats entire documents as interconnected networks, recognizing that AI-generated passages tend to cluster together and that stylistic shifts happen gradually across sentence boundaries.

Why Can't Older Detection Methods Keep Up?

The challenge of spotting AI-written sentences in hybrid documents has grown more urgent as large language models (LLMs), which are AI systems trained on vast amounts of text to generate human-like writing, become more sophisticated. Existing sentence-level detection methods suffer from a critical blind spot: they classify each sentence independently, ignoring the structural patterns that reveal authorship transitions. When researchers tested the previous best method on sentence-level detection, its accuracy dropped below 85% even on documents it had seen before during training.

Beyond methodology, a parallel gap existed in evaluation benchmarks. Current detection systems were built using older AI models like GPT-4o and DeepSeek-V3, but the newest reasoning models such as OpenAI's o1 and o3 were absent from these benchmarks. This meant no one knew whether detection tools trained on earlier AI outputs would work against current-generation systems.

How Does SenFlow Detect AI-Generated Sentences?

• Graph-Based Analysis: SenFlow treats each document as a sentence graph, propagating contextual information across sentences to understand how they relate to one another, rather than evaluating them in isolation.
• Pattern Recognition: The system recognizes that AI-generated passages tend to cluster in contiguous blocks and that authorship transitions evolve gradually across sentence boundaries, not at random points.
• Structural Signals: Even when AI text is polished to match human writing quality, it retains subtle structural patterns like sentence-length gaps that differ between AI generators, which SenFlow exploits for detection.

Researchers also constructed MOSAIC, a new benchmark dataset containing 16,000 hybrid documents generated across biomedical and news domains using two different AI systems: DeepSeek-V3.2, a reasoning model, and Kimi K2, a chat model. The dataset was built with stringent quality controls, including a perplexity-consistency filter absent from prior benchmarks. Perplexity measures how surprised an AI model is by text; by equalizing this metric, researchers could isolate other detectable patterns beyond surface-level naturalness.

What Results Did the New System Achieve?

SenFlow reached state-of-the-art performance across three testing protocols of increasing difficulty. Most impressively, it achieved a 4.15 percentage-point average improvement in accuracy on cross-domain transfer tests, the hardest evaluation scenario where the system must detect AI text in domains it has never encountered before. The system accomplished this while using approximately 20 times fewer computational calls to proxy models than the previous best method, making it far more efficient.

A key empirical finding emerged from the research: even after filtering away obvious perplexity-level cues that make AI text detectable, machine-generated insertions still retain a generator-dependent sentence-length structural gap compared to surrounding human prose. This gap was larger for DeepSeek-V3.2 than for Kimi K2, demonstrating that different AI systems leave distinct fingerprints. The discovery proves that simply making AI text sound more natural does not render it undetectable at the sentence level.

The implications extend beyond academic interest. As hybrid authorship becomes standard practice in education, research, and professional writing, the ability to identify which sentences originated from humans versus machines becomes essential for maintaining academic integrity, protecting intellectual property, and ensuring transparency in professional documents. SenFlow's approach of analyzing inter-sentence relationships rather than isolated sentences represents a meaningful step forward in this ongoing technological arms race between AI generators and detection systems.