Behind the Scenes: Multilingual Detection Update
Back in April, we released a major update (Model 3.1m) that outperformed competitors across the board. Since then, we’ve trained on more challenging human and AI data, and have achieved a less than 1% false positive rate on all supported languages (Model 3.2m).
AI-generated writing isn’t just an English-level issue, which is why, as language models grow increasingly sophisticated across different languages, detection tools have to keep up. This is why we’ve updated our multilingual AI detection model again – this time, it’s even more accurate.
Back in April, we released a major update (Model 3.1m) that outperformed competitors across the board. We added support for German and Portuguese, and retrained on more recent French and Spanish data. Since then, we’ve trained on more challenging human and AI data, and have achieved a less than 1% false positive rate on all supported languages (Model 3.2m).
How did it perform?
We evaluated our model on the same benchmarks as before, with data from each of the languages we support. This includes benchmarks split ~50/50 between human and AI-generated text, containing multiple datasets for each language.
Table 1 contains the improvement in FPR from our previous model. Our FPR increased slightly on French texts, but is now consistently below 1%.
Table 1: FPR Scores Per Language on Internal Benchmark
Table 2 contains the overall accuracy, recall, and false positive rate for the same benchmarks. This model outperforms our previous one on accuracy and recall for all languages.
Table 2: More Performance Metrics Per Language on Internal Benchmark
Why this matters
Our commitment to having the industry-wide lowest false positive rate is driven by our mission to ensure responsible use of AI detection. Fewer false positives means fewer incorrect accusations of plagiarism, all while providing reviewers and educators transparency into content origin. The latest multilingual model release demonstrates the effectiveness of our approach to AI detection. We are capable of rapidly improving our detector to meet user needs and stay true to our commitment to having the best AI detector.
If you have requests for more languages you would like to be supported, please fill out this form here and help us continue improving your experience.
