LIVE FEED
ATLAS OWASP HIGH Significant risk · Prioritise patching RELEVANCE ▲ 6.5

Meta Faces Lawsuit Over Biometric Data Harvesting for AI Training

TL;DR HIGH
  • What happened: Meta sued for harvesting biometric data from social media photos to train AI and build facial recognition.
  • Who's at risk: Facebook and Instagram users in Illinois and California whose photos were used without consent to build AI and biometric systems.
  • Act now: Audit all AI training datasets for biometric data collected without explicit user consent · Implement opt-in consent mechanisms before extracting biometric identifiers from user-generated content · Review compliance with state-level biometric privacy laws (BIPA, CCPA) before deploying AI features at scale
Meta Faces Lawsuit Over Biometric Data Harvesting for AI Training

Overview

A proposed federal class action filed in Chicago alleges that Meta Platforms illegally harvested biometric information from millions of Facebook and Instagram photos to train generative AI models — including Emu and Muse Image — and to develop NameTag, an unreleased facial recognition feature embedded in Meta’s smart glasses companion app. The plaintiffs, parents and children from Illinois and California, claim Meta violated the Illinois Biometric Information Privacy Act (BIPA) and California privacy statutes by extracting faceprints and other biometric identifiers without notice or consent.

The case is significant not only as a legal matter but as a signal of growing scrutiny over how AI companies source and process training data, particularly when that data encodes sensitive biometric characteristics.

Technical Analysis

The NameTag system, discovered by WIRED in June 2026 through code analysis of the Meta AI companion app (downloaded over 50 million times), was designed to convert faces captured by smart glasses into biometric signatures and compare them against a local faceprint database on the user’s device. That database was configured to receive remote updates from Meta’s servers — suggesting a centralised data pipeline, despite Meta’s claims that it is “not building a universal face database.”

The complaint further alleges that the faceprints may be derived from Facebook and Instagram profile images, citing internal employee statements and a Meta patent describing face-matching against user profile photos. Meta’s chief product officer publicly acknowledged that Facebook and Instagram images formed a core “data advantage” in training Emu, the image generation model underpinning several Meta AI products.

Muse Image, released in summer 2026, briefly allowed users to generate images modelled on other people’s public Instagram accounts — a capability that demonstrates how biometric likenesses can be reconstructed and weaponised using generative AI trained on social media content.

Framework Mapping

MITRE ATLAS:

  • AML.T0020 – Poison Training Data / AML.T0059 – Erode Dataset Integrity: The unconsented ingestion of biometric-rich social media images into training pipelines represents a form of dataset integrity violation, where personal data is repurposed beyond its original context.
  • AML.T0047 – AI-Enabled Product or Service: NameTag exemplifies an AI-enabled product designed to perform real-world biometric identification, raising significant misuse potential.
  • AML.T0088 – Generate Deepfakes: Muse Image’s short-lived feature to generate images based on real users’ Instagram accounts directly implicates deepfake generation capabilities.

OWASP LLM Top 10:

  • LLM03 – Training Data Poisoning: The core allegation concerns the integrity and provenance of AI training data, specifically the unconsented use of biometrically sensitive images.
  • LLM06 – Sensitive Information Disclosure: Biometric signatures derived from training data may be recoverable or reconstructable from deployed models.

Impact Assessment

The immediate legal exposure falls on Meta, but the broader implications extend to any AI vendor ingesting user-generated content from social platforms. Biometric data is among the most sensitive categories under privacy law — it is immutable, meaning exposure cannot be remedied by a password reset. If courts find that training AI models on photos constitutes biometric data processing, this could fundamentally constrain how social-media-derived datasets are used across the industry.

Mitigation & Recommendations

  • Obtain explicit, informed consent before using images for biometric feature extraction or AI model training.
  • Conduct biometric data audits across existing training datasets to identify and remediate unconsented inclusions.
  • Apply data minimisation principles: strip or anonymise biometric identifiers before ingestion into training pipelines.
  • Engage legal review for compliance with BIPA, CCPA, and emerging EU AI Act obligations before deploying biometric AI features.
  • Implement transparency disclosures in product documentation and privacy policies covering AI training data sources.

References

◉ AI THREAT BRIEFING

Stay ahead of the threat.

Twice-weekly digest of critical AI security developments — every story mapped to MITRE ATLAS and OWASP LLM Top 10. Free.

No spam. Unsubscribe anytime.