Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 27·8 min read
Annotator Positionality as Signal: Psychometric Weighting for Anti-Autistic Ableism Detection
TL;DR
Large language models used in high-stakes decisions can amplify ableist biases, but their ability to detect anti-autistic ableism remains unclear. Researchers introduced a psychometrically-weighted evaluation framework that uses community-informed ground truth labels to measure how well LLMs identify ableist language targeting autistic people.
✦ Why It Matters
Engineers can now audit LLM bias against autistic people using a validated framework before deploying models in healthcare, hiring, or social services.
Key Takeaways
- Adopt psychometric weighting to enhance bias detection in LLMs.
- Incorporate diverse annotator perspectives to improve model accuracy.
- Avoid relying solely on keyword matching for language analysis.
- Evaluate LLM outputs against community-defined standards for fairness.
How It Works
The framework uses psychometric weighting to prioritize perspectives from autistic individuals and those who accept autism, ensuring that their views are adequately represented in the evaluation of language models. This contrasts with traditional methods that may overlook these perspectives, leading to biased outputs.