In 2019, a research paper proposed "model cards" as a way to increase transparency into AI systems and mitigate their potential harms. Model cards are short documents accompanying machine learning models that disclose key details for assessing whether they are appropriate for a use case. The authors argued model cards are a step towards democratizing AI responsibly.
As AI proliferates, external audits have found many deployed systems encode biases and fail on marginalized groups. However, since current models lack standardized reporting, it's hard for practitioners to evaluate suitability and compare options. Model cards aim to change this by requiring transparent documentation of model capabilities and limitations.
The proposed model card framework contains sections summarizing:
- Model details like architecture and training approach
- Intended use cases and exclusion criteria
- Relevant factors like demographic groups for evaluating performance
- Quantitative performance metrics, broken down across subgroups
- Training and evaluation data sources
- Ethical considerations during development
- Limitations and recommendations
For example, a model card could report a facial recognition system's error rates across race, gender, and age groups. This transparency into variability helps assess if the model is appropriate for an application context.
Model cards complement datasheets for datasets, which document training data characteristics. Together, they increase accountability across the AI lifecycle. The authors presented example model cards for an image classifier and toxicity detection algorithm.
The image classifier model card revealed a high false positive rate for elderly males being classified as smiling. This demonstrates the importance of intersectional analysis - evaluating across combinations of factors like age and gender. The toxicity detector model card showed improved performance on minority groups between two versions, illustrating model cards can track progress.
The authors acknowledged model cards have limitations. Their usefulness relies on creator integrity. They are flexible in scope and do not prevent misleading representations. Cards should complement other transparency techniques like external auditing.
However, by standardizing model reporting with relevant details for stakeholders, model cards represent a step towards responsible AI development and deployment. They help assess whether models warrant trust for particular use cases. The proposed framework offers a template for organizations to evaluate models rigorously before adoption. Broad adoption of model cards could enable accountable AI systems that avoid perpetuating inequality.
Increasingly, regulators are also recognizing the importance of transparency in AI systems. For instance, the proposed EU AI Act requires certain disclosures like intended use cases and limitations. The Model Card Regulatory Check automates checking if a model card complies with these regulatory requirements. This demonstrates how model cards can facilitate efficient regulatory compliance, beyond just informing users. Linking model cards to governance frameworks reinforces their value in responsible AI.
Sources:
Model Cards for Model Reporting
