Product Update: Launch of Wordvice AI Detector 2.0

Nine months after the release of AI Detector 1.0 in December 2025, we are rolling out a major model update.

Over the past nine months, artificial intelligence has advanced at an extraordinary pace, with new technologies and models emerging almost daily. The ways in which humans and AI collaborate in the writing process have also evolved rapidly. At the same time, discussions have intensified around how AI-generated text should be detected, how detection results should be reviewed by humans, and what ethical or practical measures should follow. Yet despite these ongoing discussions, there is still no clear consensus, while AI technology continues to advance at remarkable speed.

Against this rapidly changing backdrop, Wordvice has continued to improve the performance of its existing models while also considering more fundamental changes to both the model itself and the overall service. In particular, our research paper, "Style as a Confound: False Positives in AI Detection of Non-Native Academic Writing," which is scheduled to be presented at EMNLP 2026, helped deepen our technical understanding of how human-written and professionally edited text interacts with AI detection systems. Read more in our blog post.

And today, we are launching AI Detector 2.0.

This update is not simply about detecting more AI-generated text. It focuses on more reliably identifying AI-rewritten text and difficult borderline cases that appear in real-world Wordvice usage. We have also enhanced sentence-level information to help users better understand and review detection results.

The three main improvements in AI Detector 2.0 are:

• An AI-generated text detection model trained on data from a wide range of sources
• Hard Negative Mining (HNM), which focuses additional training on difficult cases that are frequently misclassified
• A Boundary Detector that highlights parts of a document that show relatively stronger AI-generated patterns at the sentence level

In our internal evaluation, the final model achieved an F1 score of 0.9773 and an AUROC of 0.9979 on the in-domain test set. In particular, for AI rewrite detection, where AI is used to rewrite existing text, the model achieved an F1 score of 0.9819 and a recall of 0.9917.

However, strong performance metrics do not mean that AI detection results can be treated as definitive evidence of authorship. An AI detector provides a signal indicating how closely a text resembles the AI-generated patterns learned by the model. Its results should therefore be interpreted together with the writing process, context, revision history, and human review.

We are also updating our usage policy. To allow paid users to use the AI Detector with virtually no practical limitation, the monthly usage allowance is being increased to 1 million characters. This reflects our view that an AI detector should not be used only once to check a finished document. Instead, it can be used throughout the writing and revision process to identify and reduce AI-specific patterns or "AI slop," helping users ultimately produce higher-quality writing.

This September, Wordvice also launched an AI detector performance improvement project with official support from the Publication Industry Promotion Agency of Korea. Building on this work, we are planning a second major update by the end of this year. The next update will go beyond simply identifying AI-like writing patterns and will focus on practical issues that can arise from careless or inappropriate use of AI, including source verification and logical consistency checks.

We look forward to sharing more as the project develops. Below, we take a closer look at the model architecture, training methodology, evaluation data, and detailed performance of AI Detector 2.0.

What's Technically Different in Detector 2.0?

1. Trained on a Broader Range of Real-World Writing Patterns

Detector 2.0 was trained on approximately 600,000 human-written and AI-generated texts. The training data includes academic and general writing, as well as a wide range of AI-generated and AI-rewritten texts, allowing the model to learn patterns that more closely reflect real-world use.

However, simply increasing the amount of training data was not enough.

Our analysis of the initial model showed that misclassifications tended to cluster around difficult boundary cases. In some cases, the model identified human-written text as AI-generated. In others, AI-generated text closely resembled human writing and went undetected.

To address this, Detector 2.0 identifies these difficult cases separately and gives them greater emphasis during additional training.

2. Hard Negative Mining Focuses Training on Difficult Cases

After the initial training phase, we selected examples that the model found relatively difficult to classify and incorporated them into an additional round of training. Internally, we refer to this process as Hard Negative Mining (HNM).

Specifically, we selected human-written texts that received relatively high AI-generation scores, as well as AI-generated texts that received relatively low AI-generation scores. In other words, the process does not simply collect examples that the model classified incorrectly. It focuses more broadly on cases in which the distinction between human and AI writing is difficult for the model to make, regardless of whether its original prediction was technically correct.

During the second training phase, these difficult examples were given greater weight so that they appeared more frequently in the training process. This allowed the model to focus not only on examples that were already easy to distinguish, but also on texts that present genuine challenges in real-world detection.

3. Added More Detailed Sentence- and Segment-Level Analysis

Detector 2.0 goes beyond providing a single result for an entire document. It also provides more detailed information about where AI-generated patterns appear more strongly within the text.

The Boundary Detector is a token-level model that classifies human- or AI-generated patterns for individual tokens. These token-level scores are then aggregated at the word and sentence levels, allowing users to identify sections of a document where AI-generated patterns are relatively more prominent.

This Boundary analysis is separate from Detector 2.0's sentence-level analysis path. While the Boundary Detector aggregates token-level predictions to highlight detailed patterns within a document, the sentence score is generated by a separate model that evaluates each sentence together with its surrounding context.

The final detection result considers both document-level detection and sentence-level analysis, while these individual analysis paths also provide additional explanatory information. The Boundary highlights and sentence scores are therefore generated in different ways and should be interpreted according to their respective roles.

How Has AI Rewrite Detection Improved?

The impact of HNM was particularly clear in our evaluation of AI-rewritten text, where AI is used to rewrite existing content.

Evaluation Metric	Initial Model	Detector 2.0	Change
AI Rewrite F1	Approx. 0.955	0.9819	+0.027
AI Rewrite Recall	Approx. 0.933	0.9917	+0.059

On a separate in-domain test set, the final model also achieved an F1 score of 0.9773 and an AUROC of 0.9979.

In particular, recall on the AI rewrite evaluation increased from approximately 0.933 to 0.9917. This indicates that, within this evaluation dataset, fewer AI-rewritten texts were incorrectly classified as human-written.

The goal of Detector 2.0 was not simply to improve overall performance metrics. A key objective was to improve detection of AI-rewritten text and difficult boundary cases that existing detectors may struggle with in real-world use.

How Can Sentence-Level Information Be Used?

When a detector provides only a document-level result such as "Similar to AI-generated patterns," it can be difficult for users to understand what contributed to that result.

Detector 2.0 provides more detailed information about where AI-generated patterns appear relatively strongly within the text, helping users identify which parts may warrant closer review.

For example:

Sentence	AI Similarity
"This study examines…"	0.82
"We propose a novel…"	0.31

A score of 0.82 does not mean that there is an 82% probability that the sentence was written by AI.

These scores are relative indicators designed to make the patterns identified by the model easier for users to interpret. A high score for a particular sentence should therefore not be treated as proof that the sentence was AI-generated.

The purpose of sentence-level information is to help users identify which parts of a document may require closer review.

How Was the Boundary Detector Evaluated?

We evaluated the Boundary Detector using a public AI text detection benchmark designed to identify transition points between human-written and AI-generated text.

The primary evaluation metric was MAE (Mean Absolute Error), which measures the average distance between the predicted human/AI transition point and the actual transition point. A lower MAE indicates more accurate boundary detection.

For reference, previously reported results on the same SemEval test set are shown below.

System	Overall MAE ↓
SemEval official baseline	21.54
TM-TREK	15.68
DeepPavlov	13.38
Wordvice Boundary Detector	7.16

However, this table should not be interpreted as a direct ranking of the systems.

Each system was trained using different datasets and training conditions, so results measured on the same test set alone are not sufficient to establish absolute performance superiority. Rather, these results demonstrate how the Wordvice Boundary Detector performed on the same public benchmark.

Stronger on Familiar Academic Texts, but Greater Caution Is Needed in New Domains

Overall performance metrics tell only part of the story. It is equally important to understand where a model performs well and where it may encounter greater difficulty.

When we break down the Boundary Detector results by data environment, a clear difference emerges.

Evaluation Environment	MAE ↓	Token F1
Academic texts similar to the training distribution	2.72	0.987
Unseen external domains	15.96	0.956
Overall	7.16	0.973

The Boundary Detector performed strongly on academic texts similar to its training distribution, while the error rate increased on external domains that were not represented in training.

This is an important limitation of AI detection systems. Strong overall performance on one dataset does not guarantee the same level of accuracy across every language, domain, or type of writing.

Detector 2.0 reflects this limitation in the product experience as well. When there is insufficient input information, or when a language or domain has not yet been sufficiently validated, users are provided with additional guidance encouraging more cautious interpretation of the results.

At present, sentence-level analysis has been most extensively validated on English academic text. Sentence-level results for non-English text or writing that differs substantially from the training distribution should therefore be interpreted with greater caution.

What Changes in Real-World Use?

The practical improvements in Detector 2.0 can be summarized as follows.

Usage Scenario	Challenge with Existing Detection	Detector 2.0 Improvement
AI generates the entire text	Difficult to capture the full range of AI-generation patterns	Training on a broader range of human- and AI-generated writing patterns
Human-written text is rewritten by AI	Rewriting can make the text appear more natural and harder to detect	Greater training emphasis on difficult AI rewrite cases through HNM
A human writes in an AI-like style	Human-written text may be incorrectly identified as AI-generated	Difficult human-written cases are incorporated into additional training
AI patterns appear strongly in only part of a document	A document-level result alone does not show which sections require review	Sentence- and segment-level information through the Boundary Detector and sentence analysis
Text is short or belongs to a less-validated domain	Limited information may lead users to place too much confidence in the result	Additional guidance encourages more cautious interpretation