Green Washing Detector

Inspiration

Greenwashing makes it difficult for people to tell whether environmental claims are meaningful or simply marketing language. Companies often use vague phrases like “eco-friendly,” “sustainable,” or “carbon neutral” without providing measurable evidence. We built Green Washing Detector to make sustainability claims easier to investigate and understand.

What it does

Green Washing Detector analyzes corporate websites, sustainability reports, PDFs, marketing copy, and company names to identify potentially misleading environmental claims.

It:

• Extracts sustainability-related evidence from submitted sources. • Searches for official ESG, climate, emissions, and sustainability reports. • Uses Browserbase to fetch pages, navigate JavaScript-heavy websites, and discover relevant reports. • Flags vague, exaggerated, comparative, or unsupported claims. • Detects missing metrics such as baselines, timeframes, scopes, and verification details. • Assigns a Greenwashing Risk Score from 0–100. • Explains each finding with an evidence status and confidence level. • Suggests more transparent alternatives that avoid making unsupported promises.

How we built it

We built the application with Next.js, React, TypeScript, and Tailwind CSS.

The ingestion pipeline supports URL scraping, PDF parsing, pasted text, and company-name discovery. For web research, we use Browserbase’s Search and Fetch APIs first, then fall back to a real Browserbase browser session through Playwright when pages require JavaScript rendering or navigation.

The analysis layer combines an OpenAI-compatible LLM with a deterministic heuristic auditor. The LLM output is validated server-side to ensure that scores, risk levels, evidence statuses, and quoted claims remain consistent with the submitted source. If the LLM is unavailable or produces an invalid response, the heuristic engine provides a reliable fallback.

Challenges we ran into

One of the biggest challenges was distinguishing genuinely unsupported claims from statements that were simply vague or incomplete. We had to account for negations, comparative claims, certification language, missing metrics, and claims spread across multiple sections of a report.

We also had to handle the variety of corporate websites. Some pages are static and easy to scrape, while others rely heavily on JavaScript, client-side routing, or dynamically loaded reports. Designing a safe escalation path from direct scraping to Browserbase Fetch and finally to browser navigation was important for both reliability and cost control.

Another challenge was preventing the analysis model from inventing quotations or treating instructions inside source documents as commands. We addressed this with strict source-grounding and response validation.

Accomplishments that we're proud of

We are proud that Green Washing Detector works even without external AI configuration through its deterministic fallback engine. The application remains useful while still supporting more advanced LLM analysis when available.

We also built a provenance-aware evidence pipeline that preserves source URLs, titles, acquisition methods, and extracted context. Instead of only returning a risk score, the application explains why a claim was flagged and what evidence is missing.

The system can now accept only a company name and automatically search for relevant official sustainability reports, while still safely falling back when Browserbase is unavailable.

What we learned

We learned that greenwashing detection is not just a keyword-matching problem. Context, negation, comparison baselines, measurement boundaries, timeframes, and third-party verification all matter.

We also learned that LLM reliability depends heavily on the surrounding system. Strong prompts are not enough by themselves; model responses need schema validation, quote verification, score normalization, and deterministic fallback behavior.

Finally, we learned that trustworthy AI tools should clearly communicate uncertainty. A claim marked as “insufficient evidence” is more useful and responsible than an unsupported claim of certainty.

What's next for Green Washing Detector

Next, we want to expand the system with:

• More regulatory frameworks and region-specific guidance. • Better extraction from long, complex ESG PDFs. • Historical tracking of company claims and year-over-year progress. • Automated comparison between sustainability claims and reported metrics. • More structured Scope 1, Scope 2, and Scope 3 emissions analysis. • Source quality and third-party verification scoring. • A browser-based report explorer with citations and side-by-side evidence. • Human review workflows for journalists, researchers, investors, and regulators.

Built With

• warp