◎

Behind the AI

Who funds, controls, and profits from the AI tools you use every day? Independent assessments. No corporate sponsors. Plain language.

∞

Meta (Llama)

▲Serious Concern

Reversed its own military ban. China's PLA already using Llama.

Key Facts

Company
Meta (Llama)
Product
Meta AI / Llama
Grade
Serious Concern
Overall Score
2 / 5
Last Updated
2026-03-22
Military Status
Active
Last Reviewed2026-03-22
Company ContactedNot yet contacted
MethodEmail to [email protected]
ResponseNo response received
Data Confidence: HighAssessed by: Behind the AI Research Team
Founded2004 / 2023
HQMenlo Park, USA
StructurePublic ($1.5T+)
CEOMark Zuckerberg
Valuation$1.5T+ (public)
Revenue$160B+/yr (mostly ads)
Core Identity

Open-weight AI to maintain US advantage. Models freely downloadable but unenforceable restrictions. Part of $650B hyperscaler spend.

Military & Government

Reversed its own military ban Nov 2024. Provides Llama to Pentagon, Lockheed Martin, Palantir, Anduril, Booz Allen. China's PLA researchers used Llama for military intelligence. Meta called it 'unauthorized' but cannot enforce restrictions on downloadable models.

Assessment by Pillar

Data Practices

Core business is data collection. Llama trains on PUBLIC Facebook/Instagram posts only. Private messages NOT used. WhatsApp conversations NOT used for Llama training (confirmed Snopes Nov 2025). Training data composition otherwise secret.

Funding Transparency

Public company. Financials fully disclosed. But $5B FTC fine for privacy violations.

Military & Gov't

Reversed its own weapons ban in 3 days. Now actively arms defense contractors. Cannot prevent adversary use of open-weight models.

Model Transparency

Some models open-weight (Apache 2.0). Largest models increasingly proprietary. Training data not disclosed.

User Rights

No meaningful opt-out while using Meta platforms. Ad-free tier exists in EU only (€5.99-7.99/month for Facebook/Instagram, regulatory compliance). Meta AI+ premium announced Jan 2026 but NOT YET LAUNCHED. Self-hosting Llama gives full control but requires technical capacity.

Key Findings

RISK

If you use Facebook or Instagram, your PUBLIC posts may contribute to Llama's training. Private messages and WhatsApp conversations are NOT used (confirmed Snopes Nov 2025). There is no consumer-facing opt-out for public post training while continuing to use the platforms.

RISK

Meta reversed its military ban 3 days after China was caught using Llama for military intelligence. The policy reversal was reactive, not principled.

RISK

Departing Meta AI chief Yann LeCun confirmed in January 2026 that Llama 4 benchmark 'results were fudged.' Zuckerberg reportedly 'sidelined the entire GenAI organization.'

CONTEXT

Self-hosting Llama on your own servers gives full data control and is the strongest privacy option available. But this requires significant technical capacity.

CONTEXT

Beyond the $5B FTC fine, EU fines total €1.9B+ and FTC has ongoing proceedings to modify the 2020 consent order.

What They Say vs. What Happened

What they said

Created "Responsible AI" team. Committed to open-source AI.

What happened

Responsible AI team deliberately steered away from misinformation in ranking algorithms. Fairness testing was optional. Pay incentives tied to engagement metrics. Reversed military use ban. Reports indicate China's PLA used Llama.

Bright spot

Most transparent on training energy data. Publishes MWh, GPU hours, and failure rates for all Llama models.

Source

Investigation into Meta Responsible AI team operations

Karen Hao, "How Facebook Got Addicted to Spreading Misinformation" MIT Tech Review March 2021

The Pattern

↓Ethics commitment announced
↓Ethics team hired
↓Team finds real problems
↓Problems threaten revenue
↓Team defunded, fired, or redirected
↓New commitment announced
↺Repeat

“They always do just enough to be able to put the press release out.”

— Hany Farid, UC Berkeley

The Law Is Catching Up

●
EU AI Act transparency obligations ↗Enforcement begins August 2, 2026
●
73 new US state AI laws in 2025 ↗Rapid state-level legislative action across the US
●
AI Whistleblower Protection ActPrompted by OpenAI NDA revelations threatening departing employees with equity loss
What to look for in an alternative

A provider with clear data training consent mechanisms, consistent military/government policies that haven't been reversed under pressure, and an ad-free product option.

Who Funds Them

Public shareholders Ad-revenue fundedN/A

Recent News

Nov 2024: Reversed military ban, gave Llama to Pentagon
2024: China's PLA used Llama for military chatbot
2025: Working with Lockheed Martin, Palantir, Anduril
Feb 2026: EU Statement of Objections over WhatsApp ban on third-party AI chatbots

Frequently Asked Questions — Meta (Llama)

Does Meta AI train on my Facebook data?

Llama trains on PUBLIC Facebook and Instagram posts only. Private messages are NOT used. WhatsApp conversations are NOT used for Llama training (confirmed Snopes Nov 2025). Training data composition is otherwise not fully disclosed. There is no meaningful opt-out for public post training while using Meta platforms.

Why did Meta reverse its military ban?

Meta reversed its policy banning military use of Llama in November 2024, three days after China's PLA was caught using Llama for military intelligence. Meta now provides Llama to the Pentagon, Lockheed Martin, and Palantir.

Sources

  1. [1]Reversed military ban Nov 2024— TechCrunch, Nov 4, 2024
  2. [2]Provides Llama to Pentagon, Lockheed Martin, Palantir, Anduril, Booz Allen— CIO.com, May 2025
  3. [3]China PLA used Llama for military intelligence; Meta called it 'unauthorized'— IEEE Spectrum, Nov 2024
  4. [4]Policy originally banned 'military, warfare, nuclear industries' use— The Conversation, Nov 2024
  5. [5]$5B FTC privacy fine— Multiple / Public record, 2019
  6. [6]Training data not disclosed— Business Standard, Nov 2024

Related Assessments

Disclaimer

Assessments reflect publicly available information and the published methodology of the Behind the AI Research Team. Grades represent analytical assessments derived from the published scoring framework, not statements of fact about internal company operations. If you believe any claim is inaccurate, contact [email protected] with the specific claim and your evidence.