Introduction
Transparency is a cornerstone of ethical AI, yet it remains a nebulous concept in practice. "Disclosive transparency"—the information a provider shares with users—often suffers from a "confusion dynamic": excessive technical detail can overwhelm laypeople, while insufficient detail leaves the system as a "black box." Researchers at UC Santa Barbara address this by developing automated, neural language model-based metrics to quantify the clarity and completeness of AI system descriptions.
Quantifying Transparency: Replicability and Style
The authors decompose disclosive transparency into two distinct, measurable dimensions: replicability (content) and style-appropriateness (accessibility).
Replicability
Replicability measures whether a description contains enough information to functionally reproduce the system. The study employs two proxies:
- Trigram Recovery: A GPT-2 model is fine-tuned to predict technical details in a full paper based solely on its abstract.
- Sentence Affinity: Using BERTScore, the researchers measure how well sentences in a brief description "cover" the technical content of the full documentation.
Style-Appropriateness
This metric assesses whether the text is written at an appropriate level for the target audience. By calculating a Log-Likelihood Ratio between a technical language model (trained on 100k arXiv papers) and a general-purpose model (trained on web text), the system can determine if a description is overly academic or appropriately accessible.
Experimental Setup
To train and validate these metrics, the researchers curated a custom corpus of 100,000 LaTeX files from arXiv (computer science categories) to define "technical" language. For evaluation, they used the ACL System Demo Corpus, focusing on 55 abstracts of NLP systems intended for lay users. These automated scores were then compared against expert ratings from NLP Ph.D. students and user attitude data from Amazon Mechanical Turk.
Key Findings and Results
The study reveals that "style" and "content" are largely independent; a description can be accessible but shallow, or technical but impenetrable.
Metric Validation
The automated metrics showed significant alignment with human expert judgment. Specifically, the Trigram Recovery metric positively correlated with expert ratings of "Function Transparency" (PCC = 0.309) and "Data Transparency" (PCC = 0.279). The Style-Appropriateness metric correlated with expert ratings of "Task Transparency" (PCC = 0.366).
The "Illusion of Understanding"
A critical psychological insight emerged: when a description clearly explains the task (what the system does), users often feel they understand the function (how it works), even if technical details are absent. This "illusion of understanding" suggests that task clarity can mask a lack of functional transparency.
The Replicability-Ethics Paradox
Interestingly, descriptions with higher replicability (more technical detail) sometimes received lower ethical ratings from users. This suggests that transparency may reveal controversial system aspects that vague descriptions successfully obscure.
Conclusion and Future Directions
This work provides a path toward objective, repeatable measurement of AI transparency. While limited by a small sample size of 55 abstracts, the results demonstrate that automated metrics can serve as reliable proxies for human judgment. Future applications include using these metrics to guide the automatic generation of style-appropriate, transparent summaries for complex AI systems, such as "AI Fact Sheets" or consumer-facing "Terms of Service."
Michael Saxon, Sharon Levy, Xinyi Wang, Alon Albalak, William Yang Wang. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.
Summary updated