AI washing or credible disclosure?
- Industry
- Finance
- Type of work
- Research & NLP
- Period
- 2026
Architecture
- 1.5M+ paragraphs 10-K, 10-Q, 8-K, proxies and earnings calls from the S&P 500, segmented without LLMs
- AI candidates logistic prefilter on vocabulary and embeddings
- Semantic frames LLM judges extract claims and concrete activities
- Firm-year panel washing index: narrative against operational grounding
- Econometrics DiD, postures and bootstrap stability
Goal
Measure whether what S&P 500 companies say about AI in their filings (2021–2026) is backed by concrete implementations, or is just narrative.
What I did
- Built a layered Parquet data lake (raw, bronze, silver, gold) with Polars, reproducible stage by stage.
- Segmented 10-K, 10-Q, 8-K, proxy filings and earnings call transcripts by section, without LLMs.
- Trained a logistic prefilter that narrows the corpus to candidates, then used LLMs to extract semantic frames and activities.
- Computed a firm-level "washing" index, clustered postures and estimated econometric models on the panel.