A Python tool that analyzes user frustration levels in ChatGPT conversation exports over time, using ML-based detection methods.
This tool processes your ChatGPT conversation export (conversations.json) and measures frustration indicators in your messages:
- ALL CAPS Detection - Counts words written in ALL CAPS (excluding common acronyms like API, URL, JSON)
- ML Profanity Detection - Uses machine learning to detect English profanity (
alt-profanity-check) - French Profanity - Word-based detection for French swear words
- Sentiment Analysis - Measures negative sentiment using TextBlob
The script aggregates data by week and produces visualizations showing frustration trends over time, with annotations for major GPT model releases.
python frustration_analyzer.py [conversations.json] [output.png]Both arguments are optional and default to conversations.json and frustration_plot.png.
The script generates two plots:
Four subplots showing each frustration indicator separately:
- Shouting (ALL CAPS words)
- ML-detected profanity (English)
- French swear words
- Negative sentiment
A single plot showing the overall frustration index (weighted average of all sources) with 4-week smoothing.
The analysis covers conversation data from late 2022 through early 2026, with vertical lines marking major model releases (GPT-4, GPT-4 Turbo, GPT-4o, o1, GPT-4.5, GPT-5).
Key observations:
- Peak frustration occurred around July 2023, shortly after GPT-4's initial release
- ML profanity scores show the highest spikes, reaching ~60% during the July 2023 peak
- Sentiment analysis reveals elevated negativity in the early GPT-4 era
- Overall trend shows frustration stabilizing at lower levels (~3-5%) in 2024-2025
- Minor upticks visible around new model releases
The analyzer filters out code and technical content before analysis:
- Removes markdown code blocks and inline code
- Filters out URLs
- Excludes lines with code indicators (imports, function definitions, etc.)
- Skips lines with high density of special characters
MIT