I built an open-source AI text detector that ranks #14 on the RAID benchmark and #1 among CPU-only solutions. No GPU required.
https://github.com/xuzhenpeng263/ai-text-detector
Why this matters
Big Tech is approaching AI detection all wrong. They're throwing massive compute at the problem – training billion-parameter models that require expensive GPUs just to detect whether a text was written by AI.
This is fundamentally misguided. Using brute-force compute to fight brute-force compute is a losing game. It's an arms race that only the largest companies can afford, and it's not scalable.
The CPU-first philosophy
I believe the solution isn't MORE compute – it's SMARTER compute. This detector uses only 60 handcrafted statistical features and a lightweight XGBoost model. It runs on any CPU, instantly, with no specialized hardware required.
The features capture the subtle fingerprints of AI-generated text:
- Compression ratio patterns
- Entropy distributions
- Burstiness metrics
- Lexical diversity signals
- And 56 other statistical markers
These signals don't require deep learning to detect. They just require knowing what to look for.
Performance
On the RAID benchmark (the standard academic evaluation for AI text detection):
- Overall ranking: #14
- Among CPU-only methods: #1
This proves you don't need a GPU to compete with the best models out there. You just need the right features.
Usage
from ai_detector import AITextDetector
detector = AITextDetector()
result = detector.detect("Your text here...")
print(f"AI Probability: {result['ai_probability']:.2%}")
print(f"Label: {result['label']}")
Why this approach wins
- Accessible: Anyone can run it, no GPU needed
- Fast: Inference takes milliseconds on any modern CPU
- Transparent: Features are interpretable, not a black box
- Sustainable: No massive energy consumption
- Privacy-first: Everything runs locally
The bigger point
The AI industry is obsessed with scale. Bigger models, more data, more compute. But sometimes the best solution is the elegant one.
AI text detection doesn't require a foundation model. It requires understanding the statistical properties that distinguish AI writing from human writing. That's something a lightweight, well-designed system can do perfectly well.
Open source means everyone can benefit from this approach – not just companies with massive GPU clusters.
Would love to hear your thoughts on the compute vs. algorithm design tradeoff in AI detection.
GitHub: https://github.com/xuzhenpeng263/ai-text-detector
Tech stack: Python, XGBoost, numpy, scipy
Language: English text detection
License: MIT