As someone new to coding, I wanted to see if AI models can actually read between the lines, or if they just look at words literally. To test this, I built a custom evaluation benchmark on Kaggle!
🧪 The Experiment
I set up a benchmark to test how well different AI models understand hidden subtext, human emotions, and patient communication.
I used two fantastic tasks created by the community:
-
infer_patient_subtext— Testing if AI can figure out what a patient actually means beneath their symptoms. -
emotional_subtext_understanding— Testing if AI can detect subtle emotional changes in text.
🤖 The Lineup
I put four massive AI models head-to-head to see how their text interpretation skills differ:
- GPT (OpenAI)
- Gemini (Google)
- DeepSeek
- Qwen (Alibaba)
📊 The Results & Insights
- The Winners: GPT and Gemini tied for the highest scores, navigating the subtext cleanly.
- The Major Surprise: DeepSeek failed completely.
- What Happened: While DeepSeek is famous for its math and complex coding reasoning, it fell into a trap when it came to reading between the lines. It struggled to infer the underlying emotional changes and patient subtext, either taking the texts too literally or missing the implicit cues completely.
- The Lesson: This proves that text interpretation isn't just about general intelligence. A model can be a logic genius but still completely misread subtle human emotions.
🔗 Check Out My Live Leaderboard
You can see the real-time scores, model rankings, and data directly on my live Kaggle page here:
View My Kaggle Benchmark Leaderboard
