Attention Chunking Experiment

DOI: 10.5281/zenodo.21523038

Summary

We test whether chunked generation (context-reset between segments) can bypass AI-text detection based on helicity trajectory analysis.

Key Finding

Chunked generation fails to bypass detection. While it destroys the autocorrelation signal (flipping it positive to mimic human values), it exposes a much stronger alternative fingerprint: helicity trajectory decay.

Group	acorr_mean	>0 rate	H_trend	neg rate
Human (n=3)	+0.072	100%	+0.254	67%
Single-pass (n=3)	−0.079	33%	+0.101	33%
Chunked (n=3)	+0.096	67%	−2.363	100%

Conclusion: Generation architecture leaves fingerprints at multiple independent levels — bypassing one detection vector creates another.

Files

• paper.md — Full academic paper
• attention_chunking_experiment.py — Experiment script (text generation + helicity analysis)
• data/attention_chunking_generated.json — Generated AI texts (3 topics × 2 modes)
• data/attention_chunking_results.json — Full analysis results

Reproduction

# Install dependencies
pip install numpy requests

# Run experiment (requires API access)
python attention_chunking_experiment.py

License

CC BY 4.0