ClimateBERT — Greenwashing Signal Detector

Paste a paragraph from a sustainability report, marketing copy, or corporate disclosure. This demo runs six specialized ClimateBERT classifiers in parallel to surface cheap-talk signals relevant to the upcoming EU regulations:

  • ECGT Directive — applies 27 September 2026, bans vague green claims and "climate neutral via offsetting" statements.
  • Green Claims Directive (proposed) — pre-verification of environmental claims.
  • CSRD / ESRS — the source of text that will be scrutinized.

Models (all from climatebert on Hugging Face, Apache-2.0): distilroberta-base-climate-detector, environmental-claims, distilroberta-base-climate-specificity, distilroberta-base-climate-commitment, distilroberta-base-climate-sentiment, netzero-reduction.

Caveats. Models are trained on paragraphs (not single sentences) and on English only. Outputs are proxy signals, not a legal verdict. Ground-truth greenwashing labels do not exist in any public dataset — every detector operationalizes proxies (specificity, commitment gap, cheap talk).

Try an example

Built on ClimateBERT by Webersinke, Kraus, Bingler & Leippold. Scoring heuristic inspired by Bingler et al., Cheap talk and cherry-picking: What ClimateBERT has to say on corporate climate risk disclosures, Finance Research Letters (2022).