The "Fairness" of AI - Meta's New FACET Benchmark

Meta's new FACET benchmark promises more rigorous AI fairness evaluations, but the company's checkered past on algorithmic bias leaves crucial questions unanswered.

The "Fairness" of AI - Meta's New FACET Benchmark

Meta recently released FACET, a new benchmark designed to assess fairness in AI systems that classify images and videos. With over 32,000 labeled images, FACET allows for evaluating biases related to gender, race, age, and other attributes. Meta states that FACET enables more thorough bias testing than previous benchmarks. However, Meta's mixed track record on algorithmic fairness raises questions about how effectively FACET will be leveraged.

A More Comprehensive Benchmark

FACET incorporates detailed labeling not found in prior benchmarks. Images are annotated for physical attributes, demographics, and activities. This granularity enables nuanced bias evaluations like determining if models struggle more with people having certain hair types. Meta also provides an explorer tool to simplify benchmarking models with FACET.

Addressing DINOv2 Biases

Meta tested its new DINOv2 computer vision model with FACET, uncovering biases including stereotyping women as nurses. Meta acknowledges these biases likely originated from DINOv2's training data. The company states FACET will help avoid perpetuating such biases going forward.

Lingering Skepticism

However, Meta's track record on fair AI casts doubt on FACET's likely impact. Prior anti-bias tools have proven insufficient, with academics criticizing Meta's algorithms as exacerbating inequalities. The efficacy of any benchmark depends on how rigorously it is applied. Given past insufficiencies, many may remain skeptical of Meta's commitment to algorithmic fairness.

The Need for Careful Oversight

FACET provides a valuable new lens for evaluating AI biases. But benchmarks alone cannot address algorithmic harms. Sincere engagement and oversight are required to ensure fairness findings lead to meaningful changes. Meta's hesitancy to take bold action on past revelations does not engender confidence. For FACET to fulfill its potential, Meta must confront flaws head-on.

Key Features of FACET

  • Extensive Dataset: Comprises 32,000 images with 50,000 people labeled by human annotators.
  • Diverse Categories: Beyond demographics, it labels for occupations and activities such as basketball players, doctors, and disc jockeys.
  • Deep Evaluations: Enables analysis of biases in computer vision AI against various classes.

Why FACET Matters

While benchmarks targeting biases in computer vision algorithms are not a novelty, what sets FACET apart is its claim to a comprehensive evaluation.

  1. Addressing Previous Concerns: Prior algorithms have shown biases against certain demographic groups. FACET's objective is to understand and highlight such disparities.
  2. Previous Controversies: Although Meta's previous endeavors in AI have raised eyebrows due to ethical concerns, FACET aims to set a new precedent.
  3. Granular Analysis: The benchmark can gauge biases in specific scenarios, such as how models perceive gender attributes in the context of certain occupations.

Origin and Preparation of FACET

  • Data Collection: Images for FACET were taken from the 'Segment Anything 1 Billion' dataset, sourced from a photo provider.
  • Annotation: Annotators labeled images for demographic attributes, physical attributes, and classes.
  • Diverse Annotators: The individuals marking up the data hail from regions spanning North America to Southeast Asia. Their compensation varied based on the country's standards.

Potential Issues and Limitations

Every tool has its challenges, and FACET is no exception.

  • Unclear Origins: Concerns arise from not clarifying the consent of people pictured and the recruitment and remuneration processes for annotators.
  • Real-world Relevance: FACET might not represent real-world concepts and demographic changes over time. For instance, changes in the attire of healthcare professionals due to the COVID-19 pandemic may affect the relevance of certain images.

In summary, while FACET moves the needle on AI bias evaluation, Meta must back up this progress with transparent and decisive mitigation efforts. Without earnest commitment to fairness, the new benchmark may amount to a hollow gesture. The efficacy of FACET will be proven not through its release, but through Meta's ongoing actions and accountability.

Read next