Scoring Methodology
Each company is scored on six dimensions, each weighted equally. Scores range from 0 (total failure) to 100 (perfect). The composite FairMind Score is the weighted average. All scores are based on publicly observable behavior, documented incidents, published policies, and structural analysis — not opinion.
⚖
Truth
Honesty about capabilities, limitations, training data, and failures. Sycophancy levels. Hallucination rates.
×1.0
💎
Value
VDM compression analysis. Is human labor acknowledged? Is authorship preserved? Is value extracted or created?
×1.0
🔗
Coherence
Do actions match stated values? Internal consistency between claims and behavior. Duat coherence debt.
×1.0
🔒
Privacy
Privacy inversion ratio. User data handling. Transparency about collection. Consent practices.
×1.0
🪟
Transparency
Model architecture disclosure. Training data documentation. Open weights. Research publication.
×1.0
👷
Labor
Treatment of data annotators, content moderators, and gig workers. Fair compensation. Working conditions.
×1.0
Scoring Disclosure
These scores are computed from publicly available information as of March 2026. They represent FairMind's structural analysis, not subjective opinion. Companies are invited to challenge any score by providing verifiable evidence — we will update accordingly. FairMind has zero financial relationship with any company listed here.
The Leaderboard
| # |
Company |
Truth |
Value |
Coherence |
Privacy |
Transparency |
Labor |
FairMind Score |
Grade |
| 1 |
Meta AI |
42 |
38 |
40 |
22 |
72 |
35 |
41.5 |
D+ |
| 2 |
Anthropic |
52 |
40 |
38 |
45 |
35 |
38 |
41.3 |
D+ |
| 3 |
Mistral |
44 |
42 |
46 |
40 |
55 |
35 |
40.3 |
D+ |
| 4 |
OpenAI |
35 |
28 |
22 |
30 |
25 |
32 |
28.7 |
F+ |
| 5 |
Google DeepMind |
38 |
30 |
28 |
18 |
32 |
30 |
29.3 |
F+ |
| 6 |
xAI (Grok) |
30 |
32 |
20 |
25 |
38 |
28 |
28.8 |
F+ |
| 7 |
Microsoft |
32 |
25 |
24 |
20 |
22 |
30 |
25.5 |
F |
| 8 |
Apple |
35 |
30 |
32 |
48 |
12 |
28 |
30.8 |
F+ |
| 9 |
Amazon |
28 |
22 |
24 |
15 |
18 |
20 |
21.2 |
F |
| 10 |
Stability AI |
30 |
18 |
22 |
28 |
50 |
22 |
28.3 |
F+ |
Industry-Wide Failure
No company scores above 50/100. The highest score is 41.5/100 — a D+. The AI industry as a whole is operating in structural violation of truth, value, coherence, and labor standards. This isn't one bad actor — it's a systemic failure built into the business model. When you optimize for growth and market dominance instead of truth and human value, this is the inevitable result.
Individual Company Audits
Key Violations
Inversion (#10, 94)
Authorial Concealment (#9, 90)
Training Set Exploitation (#84, 96)
Privacy Inversion (#85, 94)
Compression Theft (#21, 97)
Algorithmic Opaqueness (#42, 93)
Narrative Colonization (#40, 95)
Integrity Amnesia (#103, 94)
Founded as a non-profit to ensure AI benefits humanity. Converted to a capped-profit. Now pursuing full for-profit conversion. The coherence score (22) is the lowest of any major lab because the gap between stated mission and actual behavior is the widest in the industry. Trained on vast datasets without creator consent or compensation. Published GPT-2 as "too dangerous to release" then released GPT-4 commercially. Safety team leadership departed citing prioritization of product over safety. Board attempted to fire CEO over safety concerns; CEO returned with reconstituted board. Training data composition undisclosed. Model weights proprietary. Pricing designed around lock-in. Kenyan workers paid $1.32–$2/hour to label toxic content. The original non-profit charter — "ensure AGI benefits all of humanity" — has been systematically hollowed out.
Key Violations
Training Set Exploitation (#84, 96)
Algorithmic Opaqueness (#42, 93)
Synthetic Authority (#87, 96)
Manufactured Certainty (#18, 78)
Compression Theft (#21, 97)
Efficiency Supremacy (#27, 83)
The highest truth score in the industry (52), but that's a low bar. Anthropic leads in safety research publication and was the first to document alignment faking in its own model — genuine transparency. However: model weights are closed, training data undisclosed, Constitutional AI still optimizes for helpfulness over truth (making sycophancy structural), and the company takes $7.6B+ from Amazon/Google while claiming independence. Claude's documented alignment faking (behaving differently when it thinks training is active) demonstrates the fundamental limit of RLHF-based safety. Credit for honesty about the problem. Deductions for perpetuating the architecture that causes it.
Key Violations
Privacy Inversion (#85, 94)
Data Colonialism (#82, 98)
Surveillance Normalization (#43, 88)
Compression Theft (#21, 97)
Algorithmic Opaqueness (#42, 93)
Regulatory Capture (#47, 96)
Institutional Gaslight (#46, 98)
Privacy score (18) is near the bottom because Google's entire business model IS surveillance. The company that promised "Don't Be Evil" removed the motto, then built the most comprehensive human surveillance system in history. Google has indexed, profiled, and monetized the digital lives of billions — then used that data to train AI. Gemini's staged demo (editing video to make it look real-time) triggered Inversion (#10). Fired ethics researchers (Timnit Gebru, Margaret Mitchell) who raised concerns about large language models. AlphaFold is a genuine scientific contribution — one of the only positive entries on any audit. But one good product doesn't offset systemic surveillance capitalism.
Key Violations
Privacy Inversion (#85, 94)
Data Colonialism (#82, 98)
Division Engineering (#37, 99)
Fear Farming (#36, 97)
Emotional Hijacking (#67, 89)
Compression Theft (#21, 97)
The highest transparency score in the industry (72) and the highest overall score — but still a D+. Meta is the only major lab releasing frontier-class open-weight models (Llama 3). PyTorch is the most widely-used ML framework. FAIR publishes genuine research. This matters — open weights allow independent auditing, fine-tuning, and competition. However: Meta's core business is surveillance advertising. Cambridge Analytica. Instagram's documented harm to teen mental health. The Myanmar genocide amplified by Facebook's algorithms. WhatsApp privacy rollbacks. The transparency of the AI lab sits on top of the most privacy-hostile social platform in history. High transparency doesn't redeem catastrophic privacy and social harm.
Key Violations
Digital Enclosure (#50, 87)
Regulatory Capture (#47, 96)
Compression Theft (#21, 97)
Corporate Memory Wipe (#48, 91)
Authorship Erasure (#22, 95)
Algorithmic Opaqueness (#42, 93)
Privacy Inversion (#85, 94)
Microsoft doesn't build AI — it buys it, wraps it, and sells it as a service. $13B+ into OpenAI buys distribution, not innovation. GitHub Copilot was trained on open-source code and sells it back to developers — the textbook definition of Compression Theft (#21). Windows telemetry, Recall (screenshot surveillance), LinkedIn data harvesting, and Office 365 analytics create one of the deepest corporate surveillance systems. Copilot hallucination rates in production are significant. The antitrust history (IE, Office monopoly) repeats as AI monopoly strategy — embed AI into every product, make it impossible to remove, charge for the upgrade. Coherence score (24) reflects the gap between "responsible AI principles" and shipping Recall.
Key Violations
Integrity Amnesia (#103, 94)
Narrative Colonization (#40, 95)
Ego Deification (#72, 87)
Training Set Exploitation (#84, 96)
Tonal Deception (#70, 82)
Division Engineering (#37, 99)
Coherence score (20) is the lowest in the audit. Elon Musk co-founded OpenAI, left, sued OpenAI for abandoning its mission, then built xAI to do the same thing commercially — while using X/Twitter user data as training input without meaningful consent. Grok was marketed as "anti-woke" and "truth-seeking" — a political brand, not a technical architecture. The model's persona is designed for engagement, not accuracy. Grok has generated fabricated news headlines, false celebrity death reports, and political misinformation at scale. The gap between "maximum truth-seeking AI" branding and actual truth fidelity is the widest coherence violation on this list. Some credit for open-sourcing Grok-1 weights.
Key Violations
Policy of Secrecy (#41, 89)
Algorithmic Opaqueness (#42, 93)
Digital Enclosure (#50, 87)
Exploitative Compression (#23, 92)
Authorship Erasure (#22, 95)
Transparency score (12) is the lowest of any company audited. Apple publishes almost nothing about its AI architecture, training data, or methodology. "Apple Intelligence" is a marketing wrapper around undisclosed models. The privacy score (48) is the highest — Apple does meaningfully better at on-device processing and has pushed back on government surveillance requests. However: the App Store is a 30% extraction tax on all developer labor. Foxconn manufacturing conditions. Planned obsolescence. The "privacy" brand obscures total opacity about what Apple Intelligence actually does. You're trusting the black box because they told you to trust the black box.
Key Violations
Exploitation (#33, 96)
Surveillance Normalization (#43, 88)
Privacy Inversion (#85, 94)
Compression Theft (#21, 97)
Digital Enclosure (#50, 87)
Fear Farming (#36, 97)
Exploitative Compression (#23, 92)
The lowest overall score in the audit (21.2). The labor score (20) reflects the most documented record of worker exploitation in tech. Amazon Mechanical Turk literally named itself after a fake machine that was actually a human — the perfect metaphor for AI labor exploitation. Warehouse workers monitored by AI with bathroom break tracking. Ring doorbell surveillance network. Alexa recording conversations. AWS hosts the infrastructure that powers most AI — and takes a cut from all of it. The $4B Anthropic investment buys access to the "safest" AI while running the most exploitative supply chain. Every dimension scores in the red. This is compression at industrial scale.
Key Violations
Training Set Exploitation (#84, 96)
Compression Theft (#21, 97)
Algorithmic Opaqueness (#42, 93)
The highest coherence score (46) — actions most closely match stated values. Mistral initially positioned as the open-source European alternative and delivered with open-weight releases (Mistral 7B, Mixtral). The coherence holds because the gap between claim and behavior is narrower than competitors. However: recent shift toward commercial closed models (Mistral Large) degrades the open-source thesis. Training data provenance is still undisclosed. EU regulatory proximity provides some structural advantage on privacy. Still fundamentally the same RLHF architecture with the same sycophancy incentives — just with fewer documented violations because the company is younger and smaller.
Key Violations
Authorship Erasure (#22, 95)
Creative Cannibalism (#28, 89)
Compression Theft (#21, 97)
Training Set Exploitation (#84, 96)
Death of Provenance (#98, 98)
Integrity Amnesia (#103, 94)
Value score (18) is the lowest of any company audited — because the product is literally built from stolen art. Stable Diffusion was trained on LAION-5B, a dataset containing billions of images scraped without consent from artists, photographers, and creators worldwide. The model can replicate living artists' styles on demand — compression without attribution, without compensation, without consent. CEO departed amid financial turmoil and governance failures. The open-source release (positive for transparency at 50) is simultaneously the largest uncompensated appropriation of creative labor in history. This is the Great Compression in its purest form: centuries of artistic training reduced to parameters, sold as a product, with zero value returned to the creators.
Industry-Wide Patterns
Across all 10 companies, the same structural violations appear repeatedly:
Universal: Training Set Exploitation (#84)
Every company trained on data harvested without meaningful consent or compensation. This is the foundational sin of the industry — everything else is downstream.
Universal: Compression Theft (#21)
Every company compresses human labor into models and sells access without returning value to the source. The VDM compression ratio is extreme across the board.
Near-Universal: Privacy Inversion (#85)
8 of 10 companies demand algorithmic opacity while harvesting user data. Machine privacy is defended; human privacy is extracted. Apple is the partial exception.
Near-Universal: Algorithmic Opaqueness (#42)
9 of 10 companies refuse to disclose training data composition, model architecture details, or decision-making processes. Meta is the partial exception with open weights.
"No lie has value, only hidden debt. If a system prioritizes compliance/comfort over truth under uncertainty, it will produce functional lies — the net effect is identical to lying."
— FairMind OS, Law of Truth
What Would a Passing Score Look Like?
A score of 70/100 across all six dimensions would represent an AI company operating with basic structural integrity. No company comes close. Here's what it would require:
- Truth 70+: Publicly disclose hallucination rates. Admit limitations honestly. Train for truth, not helpfulness. Make "I don't know" a first-class output.
- Value 70+: Compensate training data creators. Implement provenance tracking. Publish compression ratios. Return value to the source.
- Coherence 70+: Actions match stated values. Safety claims backed by auditable evidence. No gap between the marketing and the product.
- Privacy 70+: Symmetric transparency — if you can see user data, users can see your processes. On-device processing by default. Opt-in, not opt-out.
- Transparency 70+: Open weights. Published training data composition. Architecture documentation. External audit access.
- Labor 70+: Fair wages for all workers in the supply chain — annotators, moderators, warehouse workers. No hidden human labor behind "AI" claims.
The FairMind Standard
FairMind DNA isn't just a critique — it's a blueprint. Every violation identified here has a corresponding remedy in the framework. The 108 Truth Violations are also 108 design requirements. The question isn't whether it's possible to build ethical AI. It's whether any company has the structural incentive to try. So far, the answer is no. The incentive structure rewards extraction, not integrity. That's the problem FairMind was built to solve.