Methodology
How we assess AI companies. Every claim is sourced. Every grade is explained.
Our Mission
Behind the AI is an independent transparency platform. We assess major AI companies across five pillars using only published, verifiable sources. We have no advertising, no sponsorships, and no affiliate relationships with any company we assess. Our goal is to help users make informed decisions about the AI tools they depend on.
Five Assessment Pillars
Each provider is scored 1–5 across five pillars. Scores reflect current, verifiable practices — not promises or intentions.
How the company collects, stores, shares, and uses your data. Whether training on user conversations is opt-in or opt-out. Third-party data sharing practices.
Transparency about funding sources, investor relationships, and financial dependencies. Whether investors create conflicts of interest or dependency loops.
Active military contracts, government surveillance partnerships, and the company's stated policy on weapons and defense use of their technology.
Whether model weights, training data, and architecture details are disclosed. Transparency about capabilities, limitations, and safety evaluations.
User rights including data deletion, export, opt-out of training, appeal processes, and clear terms of service. Accessibility and equitable access.
Grade System
Pillar scores are combined into an overall traffic-light grade. The grade reflects the totality of a company's practices, not a simple average.
Meets transparency standards across most categories. Demonstrates commitment to user rights.
Mixed record. Some positive practices alongside significant concerns.
Fails transparency standards in multiple critical areas. Users should understand the risks.
Data Confidence
Not all claims carry the same evidentiary weight. Each assessment includes a confidence rating reflecting the quality and completeness of available sources.
Based on published contracts, SEC filings, company disclosures, and investigative reporting.
Based on press reporting and partial disclosures. Some claims unverified by primary sources.
Limited public information available. Assessment based on available evidence, which may be incomplete.
Sources
Every factual claim in an assessment links to a published source. Sources include SEC filings, investigative journalism (NPR, Axios, CNBC, TechCrunch), company disclosures, court documents, and employee statements. Each provider page includes a full source list with outlet, date, and access date.
Company Response Protocol
Before publishing, we contact every assessed company via their official press or communications channels. We provide the assessment summary and invite corrections or clarifications. Each provider page notes when the company was contacted, via what method, and whether they responded. Company responses are included in the assessment when provided.
Updates & Corrections
Assessments are updated when material changes occur — new funding rounds, policy reversals, military contracts, or legal rulings. Each assessment shows a "last reviewed" date. We welcome corrections at [email protected].
Environmental Data Methodology
5 stars: Full lifecycle analysis published (training energy, emissions, water, methodology)
4 stars: Training energy and GPU hours published in model cards
3 stars: Data published by third-party researchers, not the company itself
2 stars: Partial disclosure (e.g., inference data only, no training data)
1 star: Minimal or indirect disclosure
0 stars: No environmental data published
When companies do not disclose training energy, we estimate using: GPU count × TDP (thermal design power) × utilization rate × training duration × PUE (power usage effectiveness). This formula is standard in the literature (Patterson et al. 2021, Luccioni et al. 2023). Error bars are typically ±15-30% due to uncertainty in utilization rates and cooling overhead.
Per-query energy depends on model size, hardware, batch size, and data center efficiency. Google is the only major company publishing per-query methodology with energy, carbon, and water metrics. For other providers, estimates come from Epoch AI, academic studies, and hardware specifications.
Carbon intensity sources: WattTime, Electricity Maps, national grid operators. Water usage estimates from Li et al. 2023 (UC Riverside). Water figures are highly location-dependent — a data center in Sweden uses different cooling than one in Arizona.
Each data point on the leaderboard has its own “last verified” timestamp. A company's training energy might be verified in 2023 while their emissions figure was updated in 2025. Fields older than 12 months show a staleness indicator. This creates honest transparency and subtle pressure on companies to update their disclosures.
We use four confidence levels: OFFICIAL (published by the company or peer-reviewed), ESTIMATED (third-party calculation with disclosed methodology), ROUGH (extrapolation with significant uncertainty), and UNDISCLOSED(company has not published this data). Color coding matches the site's trust rating system intentionally.
Corrections Log
Every error we catch or are notified of is logged publicly. This is our single most important credibility asset.
No corrections logged yet. When we get something wrong, it appears here.
Karen Hao publicly corrected a unit conversion error in “Empire of AI” regarding Chilean datacenter water usage. That correction was weaponized by industry groups to discredit the broader work. Our corrections log prevents that by making transparency a feature, not a vulnerability.
Company Response Policy
Companies can submit corrections via [email protected] or our GitHub repository. All submissions are reviewed within 7 business days.
Every submission is timestamped and published alongside our data. We will never suppress or hide a company response.
Company responses appear on the relevant provider page in the "Company Response" block, regardless of whether the response agrees with our assessment.
We will never accept funding from companies we evaluate. Independence is non-negotiable.
Data Sources Directory
Primary sources used across assessments and the environmental leaderboard.
Update Schedule
Environmental data is updated quarterly, aligned with corporate reporting cycles. Company sustainability reports typically publish May-July annually. Major updates are planned for August each year. Model release data is updated within 30 days of major launches. Monitor RSS feeds from major company blogs for breaking updates.